Determine how to break up simulation into parts

[eth1030]

• we need to think about when we may need the host to jump between different parts of the simulation

[ajamias]

what do you mean by jump?

[eth1030]

Oh let me update the comment, but what I meant is we need to determine how to break the simulation up ie. into three parts and see if we need the host CPU to save any of the information (unless we only need it to run the GPU code)

[ajamias]

For the "naive" GPU implementation (not geared to any compute capability), we will have something that looks like this pseudocode:

memcpy structures from CPU memory to GPU memory
for each timestep:
    execute parallel force computation on GPU
    execute parallel motion update on GPU
memcpy structures from GPU memory to CPU memory

there may be another way to combine the force computation and motion update into the same kernel code, but that requires research on ways to synchronize across thread blocks. And if we are able to do that, we can also move the timestep loop in the kernel code