Thursday, May 7, 2009

Framework1 / Quake3 level Bezier patches

Today I got the Bezier patches working in the renderer.
Quake3 uses the simplest patches possible, bi-quadratic, which means 9 control points per patch.

A face in Quake3 can be composed of multiple patches arranged in a grid, when this is the case, each patch shares a line of 3 control points with it's horizontal and vertical neighbor.
To render such a face, all what needs to be taken care of is to correctly setup and share the control points, the resulting tesselated vertices however, are not shared and also not stitched together. The fact that they are not shared is obvious since they will tesselate to different vertices, not having to stitch them together however (based on the references mentioned in previous articles), sounded a bit strange to me, but it seems this really is the case, since the renders look ok.

I implemented two version of the Bezier tesselator, a simple one (as a reference for debugging and testing) that makes a mesh directly out of the control points, and another real tesselator.

The whole implementation took roughly 8 hours of work.

As usual, I took some screenshots, some are quite dark (despite a bit of gimp post-processing), but if you download the image you can see sufficient detail.
I grouped all the shots in one image, the ones the left show the 'reference' control point renderings, while the ones on the right show a tesselation of level 5 (6x6=36 vertices per patch).

In some places, small artifacts due to texture coordinates are visible, (the arched gate), I checked the level with another Quake3 bsp renderer and I observed the same artifacts, so this might be a level design problem, in any case it is not a big deal.

I really got into C# generics to make the tesselator be able to tesselate any kinds of vertices without needing to know their structure, I did this using Generics with 'where' constraints which was needed to be able to weight and add vertices together using functions of a required interface.
I really liked that, specially that implementing an interface is very cheap and does not mean that the functions implemented are 'virtual' and therefore less performant that non-virtual functions. It is simply a 'compile-time' constraint, nice!

This might be premature optimization, but since the whole design is based on streaming, all loading operations (such as loading a face that just became visible and tesselating in on the fly) better be fast enough.

No comments: