Optimize packing and unpacking particles to pinned memory buffers.

[atmyers]

This PR makes two changes that increase the speed of packing and unpacking the particle data. First, when a transposition from SoA to AoS happens, it is done in shared memory, which is better than doing unordered reads/writes to global memory. Second, when doing a threaded copy from shared to global memory, we make sure each thread accesses memory addresses that are 8 bytes apart.

I have verified that this branch gives the expected error norm for rho on the 2-rank version of the linear wake test w/ GPUs.

Without these changes, on the debugging_particle_communication branch:

TinyProfiler total time across processes [min...avg...max]: 18.43 ... 20.64 ... 22.05

--------------------------------------------------------------------------------------------
Name                                         NCalls  Excl. Min  Excl. Avg  Excl. Max   Max %
--------------------------------------------------------------------------------------------
Hipace::Wait()                                   10  2.087e-05      5.299      16.69  75.66%
DEBUGGING: loop over ptd.packParticleData         7          0      8.907      13.43  60.90%
Hipace::Notify()                                 10  3.942e-05      1.957      7.769  35.23%
DEBUGGING: Loop over ptd.unpackParticleData       7          0      3.685       5.02  22.76%
main()                                            1   0.003096      0.288      1.062   4.81%

With these changes:

TinyProfiler total time across processes [min...avg...max]: 1.241 ... 1.301 ... 1.34

--------------------------------------------------------------------------------------------
Name                                         NCalls  Excl. Min  Excl. Avg  Excl. Max   Max %
--------------------------------------------------------------------------------------------
Hipace::Wait()                                   10  2.068e-05     0.4756     0.7457  55.63%
Hipace::Notify()                                 10  2.544e-05     0.1646     0.5969  44.53%
AnyDST::CreatePlan()                              2      0.411     0.4182     0.4266  31.83%
DEBUGGING: Loop over ptd.unpackParticleData       7          0    0.06304    0.08475   6.32%
DEBUGGING: loop over ptd.packParticleData         7          0    0.05925    0.07905   5.90%
main()                                            1   0.002925    0.03015    0.05634   4.20%

• Small enough (< few 100s of lines), otherwise it should probably be split into smaller PRs
• Tested (describe the tests in the PR description)
• Runs on GPU (basic: the code compiles and run well with the new module)
• Contains an automated test (checksum and/or comparison with theory)
• Documented: all elements (classes and their members, functions, namespaces, etc.) are documented
• Constified (All that can be const is const)
• Code is clean (no unwanted comments, )
• Style and code conventions are respected at the bottom of https://github.com/Hi-PACE/hipace
• Proper label and GitHub project, if applicable

[MaxThevenet]

Looks great, thanks for this PR! Looking forward to seeing production tests. I added a few suggestions:

1. Should we add a few const?
2. I tried to add a few comments. Could you check they are accurate, and fix them if they aren't?

Thanks!

[MaxThevenet]

Could you give a hint of where this number comes from? Is this roughly the largest number of plasma particles that can fit in shared memory?

[atmyers]

Roughly, I think 256 would also work. In principle we could try to tune this.

[SeverinDiederichs]

I did a scaling for production run parameters: (1024**3 grid points, 1 plasma particle per cell).

Additionally, the packing and unpacking is now drastically shortened:

TinyProfiler total time across processes [min...avg...max]: 178.7 ... 178.7 ... 178.7

---------------------------------------------------------------------------------------------------
Name                                                NCalls  Excl. Min  Excl. Avg  Excl. Max   Max %
---------------------------------------------------------------------------------------------------
Hipace::Wait()                                         100  0.0001306      27.49      51.71  28.93%
FFTPoissonSolverDirichlet::SolvePoissonEquation()    64000      49.82      50.23      50.47  28.24%
UpdateForcePushParticles_PlasmaParticleContainer()   51200      14.77      21.89      38.81  21.71%
DepositCurrent_PlasmaParticleContainer()             25612      15.58      19.99      25.97  14.53%
Hipace::Evolve()                                         1      1.005      6.799      12.69   7.10%
Fields::ComputeRelBFieldError()                      25600      7.615      8.016      8.964   5.01%
Hipace::Notify()                                       100  0.0001339      2.892      7.127   3.99%
FillBoundary_nowait()                                76800      4.165      5.266      6.552   3.67%
DEBUGGING: Loop over ptd.unpackParticleData             87          0      2.738      3.956   2.21%
Fields::TransverseDerivative()                      102400      3.767      3.838       3.95   2.21%
DEBUGGING: Loop over ptd.packParticleData               87          0      3.037      3.905   2.18%
MultiFab::LinComb()                                  76800      3.327      3.376      3.467   1.94%
FabArray::setVal()                                   64507      3.293      3.346      3.411   1.91%
MultiFab::Subtract()                                 51200      2.239      2.277      2.347   1.31%
Hipace::PredictorCorrectorLoopToSolveBxBy()          12800      1.961      2.063      2.236   1.25%
Fields::SolveExmByAndEypBx()                         12800      1.967      2.003      2.071   1.16%
ResetPlasmaParticles()                               12900      1.808      1.908      2.004   1.12%

[MaxThevenet]

Awesome, thanks!