I’ll admit that I just had a really geeky weekend. I spent most of it coding away and I thought I should share what I’ve learnt. My mini weekend project was rebuilding the low detail model generator in BuildR. It was working fine but for what it does, however it was a little too slow and I felt I should spend some time working out why.
The output needed to be a low poly, single textured version of the high poly model that BuildR outputs. Generating the simplified mesh is super simple. Turning facades into planes and stripping some detail from the roof brought the polycount from 3813 down to 875 in the Paris example above. It was transforming the facade details and textures into a single atlased image that proved time-consuming.
The whole generation process clocked in at about 8000 milliseconds on Saturday morning. This was a little too long, even for doing things in the Unity3D editor. Right now I’ve got it down to 600 without cutting the quality significantly. There is definitely some more research I can do in this area (use RenderTextures maybe?) to reduce it further and 600ms is still far too high for this to be used at runtime.
The key to the problem lay in the powerful function in Texture2D called PackTextures. For most things, this is an awesome, magic function – you just throw your textures in and a single packed texture comes out. Simple right? To do this, PackTextures requires a set of Texture2Ds. That means that BuildR needed to create a new Texture2D object and call Apply for each facade. However, with the large textures that BuildR deals with, these are not only very expensive operations but take up a massive memory footprint.
I figured one way to avoid this would be to use a single Color32 array throughout the whole procedure and use one Texture2D object and Apply call at the end. To do that I needed to build a custom texture packer and texture resize function which I previously allowed Texture2D to handle.
The process I ended up with felt a little backwards at first, but it seemed like the best way. The first thing BuildR now does is build the roof mesh. From this mesh generation, BuildR extracts which textures are used and the world dimensions they occupy so it can atlas an appropriately sized version. The data goes straight into the custom texture packer along with basic facade data.
The texture packer actually only generates bounding rectangles for all the textures and does no actual image manipulation It sorts the textures by width, stacking texture bounds on top of one another. If they don’t fit, they are added to the right creating a new column. There is also no limit at this stage to the size, but the overall dimensions are kept within a square shape as the output would be a power of two.
Once all the rectangle bounds are generated, the whole thing is scaled down to the nearest power of two size and BuildR begins generating facade textures. This involves filling the main colour array with values based on the texture bounds that the texture packer generates.
BuildR needs to resize textures so that they fit the output image size. I ended up implementing a nearest neighbour algorithm to tackle this. This simple algorithm down samples the image into a smaller size giving BuildR reasonable results. I feel there is a lot more research that could be done here, but this ends up being a fairly expensive operation.
At this point it’s worth mentioning a problem with using Texture2D.PackTextures padding. You can specify an amount of padding in this function, but the colour of the background is black. The buildings end up with some nasty black lines along some of the facade edges (as above) and it is only worse with lower mipmaps. To solve this, facades need to have their own padding that reflects the edges of the textures.
I’ve attached the RectanglePack.cs class but I’ll do a separate blog post on texture resizing algorithms.
TL:DR – By dumping multiple instances of Texture2D, calls to Apply, performance is dramatically improved. This necessitated dumping a couple of useful Texture2D functions Resize and PackTextures.