Decide what the data structures will look like

[ajamias]

[eth1030]

• first consider what information we need

• how to store to exploit GPU memory

• can come up with a few options with tradeoffs

[ajamias]

Cell list: group particles spatially in a "Rubiks Cube"

I'm thinking that a cell will have the following properties:
id: int (this can be the cell list index)
particle_list: Particle[N] where $N \stackrel{\text{def}}{=} \texttt{block\_mem} / \texttt{sizeof(Particle)}$
neighboring_cells: Cell[] (if a particle switches cells)

Basically, the size of these lists and structures depend on the hardware parameters

[ajamias]

There will also be a "shared" forces array that all the thread blocks will be able to update in each time step for each particle, indexed by the particle's ID.

the dimensions of the cell array will look like [CELL_LENGTH_X][CELL_LENGTH_Y][CELL_LENGTH_Z]

the dimensions of the thread block array will look like [CELL_LENGTH_X * CELL_LENGTH_Y * CELL_LENGTH_Z][14]

the dimensions of the thread array in each thread block will look like [MAX_PARTICLES_PER_CELL]

I think that there will be enough memory in shared memory for all the structures that we decide to pull from global memory in all generations