Fig. 2.

Diagrammatic depiction of the overall algorithm for one bulk time step. Operations have been batched into a single category for simplicity. The shown structure reflects optimizations to best overlap compute-intensive simulation time steps on the GPU/window with the CPU/bulk. The only synchronization points between the window and bulk parts of the APR model are the bulk-to-window, window-to-bulk communication routines, and checking/re-initialization if the window moves.