Current Interpolation: During add to EMF

[ax3l]

This commit adds the general infrastructure to call functors before adding FieldJ to the electro-magnetic fields (in Yee: before manipulating the E field).

The previous method None without interpolation is added as a default in fieldSolver.param. Additionally, a simple second order Binomial filter for 2D and 3D is added to test the new functionality.

Run time tests:

• KHI 3D e+/e- devvs. with/without Binomial filter (with ZigZag & TSC)

[ax3l]

@psychocoderHPC: for loop over simDim -> avoids second impl

[ax3l]

@psychocoderHPC event system updates added.

[ax3l]

Should be ok for NULL values when unused.

Also see: #532

[psychocoderHPC]

Yes __delete() checks for NULL

[ax3l]

even delete() itself checks for null (nullptr and NULL are well-defined input for the std delete), that is why __delete() should not double-check it (see #532)

[ax3l]

Yee solver, KHI default but with e+/e-:
dev vs pull with None vs pull with Binomial

-d 2 8 1 -g 192 512 16 -s 2000 (hdf5 every 250)
Courant c*dt <= 1.02768 ? 1
full simulation time: 556 564 561 sec

t=1000:

[ax3l]

the main difference might be due to the seed, since the species ID is used in it (dev to None).

The binomial image looks "smoother" with less edgy features in the B-fields, which I would expect and is quite physically correct.

I'll check again with the old seeds for a comparison (src/picongpu/include/simulation_types.hpp).

[psychocoderHPC]

Is it correct that the CFL is near 1.0?
You wrote: Courant c*dt <= 1.02768 ? 1

Update: sry for closing, I would only close my comment without saving.

[psychocoderHPC]

Only a note: In general a forward declaration like currentInterpolation/None/None.def and currentInterpolation/Binomial/Binomial.def is not needed because the solver itself is free from dependencies.
It is also possible to write and use only *.hpp because everything is given in as template.

Update: Some solver needs EPS0 therefore it isn't possible to scip the *def files.

[ax3l]

same as before, no hdf5, png every 100 (instead of 10):

else still: binning every 100, energy sum of fields & particles every 10.

full simulation time: 341 341 346 sec (order: dev, none, bin.)

total energy over time:

t=1000: (B red/green, density e- red)

[ax3l]

ok that looks valid (I now keep the seed change as we planned, see first pictures before). the energy and images are identical for dev and the new None. Binomial is with a slightly better smoothed b-field.

the only small difference is a tiny bit different scaling in the png preview, that I don't see an immediate reason for, and the correct smoother b-fields with Binomial.

[psychocoderHPC]

👍