Calypso88

It's not often that my experimental projects line up with my day job, but a couple months ago I talked the higher-ups into letting me do some crossover work with 3D terrain rendering.

The obvious issue with Flash and 3D is performance. The maps we wanted to show we're rectilinear height-maps, 256 vertices to a side. I already knew there was no way to do this in a photo-realistic way (256² is just over 65,000 squares, which would be 130,050 triangles to render...last I checked, Papervision gets unusable around 2000) so I decided to go with a wireframe style instead.

Since I didn't need camera's or any of the weight of the bigger engines, I opted to use Mathieu Badimon's very lightweight and (imo) thoughtfully laid out FIVe3D library as my starting point. From there I got a quick proof of concept up and realized there was still no way I could get away with the full 256² grid so I started cutting down the resolution. Using every third node (64²) struck a good compromise with the fidelity of the original mesh, and the performance limitations I needed to stay in to work for the majority of our users.

Off the shelf, FIVe3D was giving me roughly 12fps and consuming a pretty weighty chunk of system memory, so I started digging in and looking for optimizations I could make. To start with, I created a specific class for my model to eliminate all the cross-class calls between Shape3D and Graphics3D. I also decided the Point3D class had to go since every render pass would create tens of thousands of throwaway objects in order to calculate the z-translations for my nodes. After internalizing those, I brought in the Matrix3D class as well and started combining some of the heavy-use functions so the AVM could stay within one scope for the majority of the calculation and drawing.

That gave me a boost of about 50% but I still needed a little more oomph. Finally, I took the dive and flattened my arrays and ended up restructuring the core FIVe3D engine so it was operating a little more specifically to my needs (mostly hard typing some things and taking out optional calls to the flat-fill shader and the like). At this point, the code is fairly incomprehensible and some of the extensibility of the engine was lost, but it did put my frame rate just under 30fps so I can't complain.

Here is a standalone copy showing the final mesh, and if you want to melt your processor, click for a version with the original full-detail grid.