FastFlow: GPU Acceleration of Flow and Depression Routing for Landscape Simulation

To test the performance of depression routing we include all terrain categories (synthetic, real, and flat) from our benchmarking suite since the distribution of local minima can have a significant impact in this case. We compare two variants of our algorithm and previous work, as follows:

• • Ours (J) Our algorithm with depression jumping.
• • Ours (C) Our algorithm with depression carving.
• • Barnes [Bar16], who proposed an algorithm optimized for CPU-based parallel or distributed computing.

• • Cordonnier et al. [CBB19], who provided a $O(n)$ solution for flow routing on the CPU with a single-core implementation.

Note that [Bar16] does not solve the depression routing problem as posed here since their goal is instead to fill depressions. Extending this to a more general solution that enables flow routing for discharge computation is far from trivial.

The comparative performance of these depression routing algorithms is shown in Figure 17. In particular, we note that our algorithms outperform previous work by one order of magnitude for terrains up to a resolution of $2048\; \backslash times\; 2048$. However, this gap narrows somewhat for the largest terrains, which indicates non-favorable scaling at the upper end. Also worth noting is that, with the exception of flat terrains, the comparative advantage of using depression jumping increases for larger terrains.