Rewrite long-range actors

[jngrad]

Fixes #4454

Description of changes:

• apply the SOLID principles
  • split electrostatics from magnetostatics
  • factor out code duplication
  • extract P3M/DP3M tuning code
  • hide ELC details in private class methods (the energy/force correction logic used to be exposed in coulomb.cpp)
  • hide Coulomb implementation details by no longer indirectly including coulomb.hpp in 30 unrelated files in the core
• write clear pre-conditions as event hooks
  • these events are triggered upon box size change, periodicity change, node grid change and cell structure change
    • partial fix for Add checks for cell system compatibility #4428
    • setting a node_grid incompatible with a long-range actor is no longer possible (the old node_grid is restored)
  • these preconditions always run once before the integration loop starts
  • write test cases to check for these events
  • bugfix: charge neutrality checks now run once before the integration loop starts, instead of once at actor activation
• encapsulate the state of long-range methods in classes managed by the script interface
  • bugfix: the cached DipolarP3M energy correction is now invalidated when the box length changes
  • remove global variables for CPU methods (partial fix for Remove global variables #2628) and introduce one global variable for each type of long-range actor: dipolar solver, Coulomb solver, Coulomb extension
  • convert Cython files for long-range actors to Python files using the ScriptInterface framework
    • long-range actors and their python object now have the same lifetime, regardless of whether the actor is active or not (partial fix for Core objects should have the life expectancy of the corresponding Python objects #4220)
  • the list of active long-range actors can no longer end up in an invalid state (partial fix for Robust exception mechanism in core setters #4219):
    • updating an active actor with an invalid parameter now automatically restores the original parameter
    • inserting an invalid long-range actor in the list of actors now automatically removes the actor from the list of active actors, even when the exception happens during tuning
• API changes:
  • ELC and DLC are no longer extensions, instead they take another actor as argument
  • P3MGPU is now able to execute the CPU energy and pressure kernels using parameters tuned for the GPU force kernel
    • closes P3M GPU has no energy calculation. #239
    • P3M-based reaction scripts run with 2x speed-up when reaction steps take place every 1000 integration steps
  • P3MGPU is now supported by ELC
  • long-range actor parameters are now immutable, since changing one parameter (e.g. the prefactor) would usually invalidate the tuned parameters
  • the charge neutrality check sensitivity can now be modified via actor.charge_neutrality_tolerance; this is mostly relevant to actors coupled to ICC, whose induced charges can have value spanning several orders of magnitude
• under-the-hood changes:
  • CPU-based long-range actors no longer rely on the MpiCallbacks framework
  • the subtracted Coulomb bonded interaction and pair criteria no longer indirectly include coulomb.hpp

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[reinaual]

where does it come from, that for ScriptInterfaceHelper-objects we need to call _activate() before appending it to the list?

[jngrad]

Here is the intended workflow:

• _activate() calls the C++ actor registration code that adds the actor in the global optional variant
• the now active actor in the core runs the sanity checks and tuning algorithm (if the method supports tuning)
• if checks or tuning fail, a C++ exception is thrown, which is handled in the registration code
  • the actor is safely removed from the core global
  • the exception is re-thrown
  • on MPI rank 0, the C++ exception becomes a Python exception
  • line 54 in actors.py throws
  • the python object is never added to the list of active actors

Since removing an actor from the actor list will trigger sanity checks in the core, we cannot remove a python actor if it is not also in the core. We also cannot keep an invalid actor in the core. LB and EK actors still rely on the old workflow.

[jngrad]

Does anyone remember why in the original implementation, we need to re-tune ScaFaCoS before each integration step? Maybe it has to do with side effects when the node_grid or number of particles changes, or when we reload from a checkpoint (tuned parameters cannot be checkpointed). Commenting out line 84 doesn't affect the testsuite, but the edge case it's supposedly guarding against is not documented in the original implementation.

[reinaual]

LGTM