Kokkos based P3M

[RudolfWeeber]

• parallelize charge and force assignment
• use efficient parallel FFT library
• parallelize K-space calculations

Charge/force assignment

• This is already a for_each on all particles. A parallel for each can be used.
• For force assignment (read from lattice, write to particle) the calculations are independent for each particle.
• for charge assignment (particle to mesh), atomic adds have to be used when writing to the mesh, as several particles can contribute to the same mesh sites

Parallel FFT

• @jngrad has separated the P3M from the FFT logic
• A parallel FFT implementation has to be wrapped in a sub-class of FFTBackend
• Currently, the data type of the meshes (e.g. res_mesh) is hard-coded in the FFTBackend class. This might have to be revisited.

K-Space calculations

• @jngrad has introduced a for_each_3d algorithm for the lattice loops.
• To my understanding, the iterations (i.e., calculations for different k-vectors) are independent from each other, so parallelization can be achieved by writing a Kokkos aware for_each_3d (if it not already exists in Kokkos)
  • For more details, see Modernize core #4931 (comment)

[RudolfWeeber]

Discussion elsewhere: to parallelize the loops, the meshes would have to be represented as Kokkos views.

I looked into this, briefly.

To bring this in iteratlively, one could begin by creating a Kokkos unmanaged view which points to the memory of the meshes.
In this way, Kokkos parallel loops can be leveraged immediately, and the underlying storage can be replaced separately.