5.1 How to Create Optimal Maps for Crystal Space

Written by Jorrit Tyberghein, jorrit.tyberghein@gmail.com.

Note: Creating optimal maps is not very easy as there are a lot of factors to consider. Crystal Space has a lot of tools to offer (like sectors, portals, visibility cullers, etc.) but using these tools effectively is an art. In this chapter we will not talk about how to create maps. For that you use external tools like Blender, QuArK, or some other suitable tool. This chapter focuses on how you should partition your map into sectors, the kind of mesh objects you should use, the visibility cullers, packing textures for more efficient rendering, combining objects, lighting considerations, etc.

For a good discussion about sectors and visibility cullers, see Visibility Culling In Detail.

For more tips about efficient maps, see Some Tips for Efficient Maps.

Sectors and Portals

A sector is the basic building block in a map (see section Visibility Culling In Detail). When you create a map you should decide upon how to partition your map in sectors. The easiest solution is to use a single sector. In many cases that may even be acceptable. Here are some points to consider when deciding upon how to partition your map into sectors:

• Creating multiple sectors complicates map design a bit since you have to think about placing portals between the sectors. On the other hand it can ease map design since you can delegate map creation by assigning different portions of a map to other designers.
• It is easier to implement dynamic loading when you use more sectors since you can easily load/unload a sector.
• Portals help with visibility culling in two ways. When a portal is not visible the destination sector of that portal is ignored completely; the complexity of the portal's destination is totally ignored if the portal is not visible. Even if the portal is visible, the portal still restricts the view frustum (i.e. you can only see what is inside the portal's outline) and this will cause many objects to be culled away very easily. Thus, one good reason to use more sectors is that doing so may improve visibility culling. This works best if the portals are small (i.e. doorways or windows).
• Open space maps are hard to partition into sectors. There is also little benefit in portal culling and dynamic loading with large open spaces. So, in general, it is probably best to use a single sector for an outdoor world.
• The insides of buildings are best represented using distinct sectors. This makes it possible to design them separately and it is also very good for the visibility culler. Doorways and windows make excellent portals.
• Whether or not you should further partition an indoor world into multiple sectors is hard to say and really depends on the complexity of the map.

Visibility Cullers

Every sector has its own visibility culler. Crystal Space currently supports two kinds of visibility cullers: Frustvis and Dynavis. In the future we will have a PVS culler as well (potentially visible set) which basically precalculates visibility in a separate tool. Dynavis attempts to work more on culling but it also has more overhead. So, you should use Frustvis in cases when you have the following kinds of maps:

• Simple sectors that only contain a few objects.
• Large sectors that are mostly open-space (i.e. for a space game). Dynavis works by finding which objects occlude (cover) other objects. In open space sectors this is rare so the extra work needed for this coverage test is usually wasted.
• Complex sectors that nevertheless do not have sufficient objects which can occlude other objects. This is on account of the same reason as described in the previous case.

On the other hand, if you have a complex map with lots of large objects then you should consider using Dynavis. If you do decided to use Dynavis for a sector you should follow the following guidelines for that sector:

• Dynavis works best with closed and mostly convex large occluder objects that do not have a lot of polygons. For example, a big wall is usually a good occluder. However, it is possible that the wall is created using lots of detailed polygons. That does not make the wall a bad occluder but it increases the cost of using that wall as an occluder. To help alleviate the cost, you should add a dummy polygon mesh (this is called an ‘occlusion mesh’) to that object. This occlusion mesh should have roughly the shape of the wall but it should only have a few polygons, which means that Dynavis can use it in a more efficient manner. Below is an example on how to set an occlusion mesh in a map file. Note that you should not replace the occlusion mesh if the base mesh is already good enough.
• If you have complex objects that have very little chance of occluding other objects (for example a star shaped object) then you should disable the visibility culling mesh (see below for example). This is called disabling occlusion writing. Dynavis will still cull the object away if needed but it will not use the object for culling other objects. This helps to improve performance of Dynavis.
• For culling objects Dynavis prefers small objects. Keep in mind that culling always happens on entire objects; that is, either the object is (partially) visible and it will be rendered completely or else the object is invisible. So, a large object has less chance of being culled away since it needs to be covered entirely before it can be considered invisible. Note that this guidelines conflicts with the next guideline, as well as with what OpenGL actually prefers.
• For occluding other objects Dynavis prefers large objects. Large objects are also prefered for other reasons (i.e. OpenGL is most efficient at rendering large objects; more on this later).
• There is a special optimization that follows from the previous two items. If you have a number of smaller objects that are themselves not very good occluders, but together they are very good (for example, a wall made out of separate segments) then it would be best to disable occlusion writing for the individual wall segments and add a new null object with a single occlusion mesh that represents the entire wall.
• It is sometimes useful to add dummy occluders to help with occlusion. For example, if you have a large dungeon-like map then you can put large simple occluding objects in the void space between the dungeon corridors (this is space where the camera should normally not be able to see). When used properly, this optimization can provide a lot of gain for performance. You can use standalone trimesh objects (see example below) to construct this, so you don't need to make a real mesh for it. This optimization can also be used when you have a complex multi-part object. You can then put a dummy occluder inside that object which will help with visibility culling.

Here is an example showing how you can replace the occlusion mesh of some mesh object in a map file:

<meshobj name="complexWall">
  <plugin>thing</plugin>
  <params>
    ...
  </params>
  <trimesh>
    <mesh>
      <v x="-1" y="-1" z="-1" />
      <v x="1" y="-1" z="-1" />
      <v x="1" y="4" z="-1" />
      <v x="-1" y="4" z="-1" />
      <t v1="0" v2="1" v3="2" />
      <t v1="0" v2="2" v3="3" />
    </mesh>
    <id>viscull</id>
  </trimesh>
</meshobj>

And this is an example of how you can disable the occlusion mesh for an object:

<meshobj name="wallSegment">
  <plugin>thing</plugin>
  <params>
    ...
  </params>
  <trimesh>
    <id>viscull</id>
  </trimesh>
</meshobj>

Here is a special note about closed versus non-closed objects in Dynavis. A closed object is an object that has no holes in it (i.e. a cube is a closed object). A cube with one of the six sides removed is not closed. When Dynavis writes an object to the occlusion buffer (coverage buffer) it will choose a technique based upon whether the object is closed or not.

• If the object is closed, then Dynavis will calculate the outline of the object and render that in the coverage buffer. Dynavis also has a depth buffer and, in this case, the maximum depth of the object will be used to update the depth buffer.
• If the object is not closed, then every individual polygon of the object is written separately to the coverage buffer. For depth, the maximum depth of every polygon will be used instead of the maximum depth of the entire object.

Here is an example which illustrates what this means in practice. Assume you have a highly tesselated sphere (a sphere mesh that has a large number of triangles). Also assume that there is another object inside the sphere. If the first technique is used (closed object) then Dynavis will be most optimal for updating the coverage buffer since writing an outline is very cheap. On the other hand it will not be able to cull away the object inside the sphere since the depth buffer is set too deep (i.e. the object in the sphere will be in front of the depth that is set in the depth buffer).

On the other hand, if the second (non-closed) technique is used then Dynavis will need to use a lot of CPU processing in order to update the coverage buffer but it will be able to cull away the object inside the sphere since the depth buffer will now contain accurate values per triangle.

We have not done any performance tests to find out which is better. If you have a really complex object (highly tesselated sphere) then the first technique will probably be better since the cost of writing the object to the coverage buffer will probably be higher then the cost of rendering the object inside the sphere. But we do not know where exactly the threshold is.

Object Types

Regardless of sector partitioning and visibility culling requirements, the choice of objects you use can also be important for performance. Crystal Space supports many mesh objects but the most important ones are:

• thing: A polygon mesh object that supports multiple materials and lightmapped, vertex based, or stencil based lighting. It is not easy to animate them internally so they are usually used for static objects. If you use lightmaps then these also will not be updated if the object moves, so that is another consideration. ‘thing’ objects are heavy-weight (use more memory than many other mesh types). Their main advantage is that they support lightmaps which means that they can cast and receive shadows even if they only have a few polygons. In contrast, an object that only supports vertex lighting can not have shadows on a single polygon or triangle (except if stencil-based lighting is used) (see section Thing Mesh Object).
• genmesh: A triangle mesh object with a single material. It only supports vertex- or stencil-based lighting and it is also not designed for internal animation (although it is possible to do so). Genmeshes are very efficient with regards to memory usage and rendering speed (see section Genmesh Mesh Object).
• sprite3d: Similar to a ‘genmesh’ except that it also supports frame-based animation (see section Sprite3D Mesh Object).
• sprcal3d: Similar to ‘sprite3d’ except that skeletal or bone based animation is supported. (see section SpriteCal3D Mesh Object).

Here are some guidelines on using and choosing these meshes:

• For static geometry, ‘thing’ or ‘genmesh’ objects are ideal. ‘genmesh’ typically is used for high-detail objects that only need a single material (every triangle can use other parts of the texture) and where there is no need for detailed shadows on the individual triangles (usually a ‘genmesh’ object is either small or highly detailed so that shadows are accurately represented using vertex lighting only). ‘thing’ is used in case you want large objects that have little detail but require detailed lighting.
• If you plan on using stencil lighting only, then it is probably best to use mostly ‘genmesh’ objects.
• For moving objects you can also use ‘thing’ or ‘genmesh’ as long as there is no need to change the object's appearance (i.e. no internal animation is required). You have to be aware of lighting problems mentioned above, though.
• There are two flags that can also help performance considerably. If you have a mesh that will never move then you should use the <staticpos> flag on the mesh object in the map (or ‘CS_MESH_STATICPOS’ flag for ‘iMeshObject’). This is a hint for the engine and the mesh object so that it may optimize performance based on the fact that the object will never move. The other flag is <staticshape>. This flag can be set on factories and mesh objects and indicates that the object will not animate (never changes shape). This corresponds with the ‘CS_MESH_STATICSHAPE’ and ‘CS_FACTORY_STATICSHAPE’ flags.
• Game actors and creatures should use ‘sprite3d’ or ‘sprcal3d’.

One important note when using ‘thing’ meshes: it is very important to try to use quads instead of triangles. This is because lightmaps are rectangular. If you have a wall made out of two triangles then you will have two full square lightmaps (one for every triangle) which is a waste given that half of these lightmaps will be unused. If you use a single quad polygon for the wall then only one lightmap will be used. Lightmaps can be big so this really can make a huge difference on performance.

Assisting the Renderer

When considering on how to design your objects you should keep in mind what the renderer prefers. For the renderer a mesh is defined as a polygon or triangle mesh with a single material and/or shader. So, if you are using a ‘thing’ mesh that uses multiple materials then this is actually a set of different meshes for the renderer. To avoid confusion we will call the single-material mesh that the renderer uses a render-mesh.

With OpenGL, and especially if you have a 3D card that supports the VBO (Vertex Buffer Objects) extension, the renderer prefers render-meshes that have a lot of polygons. So, for the renderer it is better to use 10 render-meshes with 10000 polygons each, as opposed to 100 render-meshes with 1000 polygons each, even though the total number of polygons is the same.

On the other hand, this requirement conflicts with some of the guidelines for the visibility culler. Getting an optimal setup depends on the minimum hardware you want to support. If you are writing a game for the future and decide to require VBO support in the 3D hardware then you should use fewer but larger objects. For the current crop of 3D card, finding a good compromize is best.

One other technique you can use to help increase the size of render-meshes is to try to fit several materials on one material. For example, if you have a house with three textures: wall, roof, and doorway then you can create a new texture that contains those three textures. The end result of this is that every house will be a single render-mesh instead of three which is more optimal for OpenGL. There are some disadvantages to this technique, however:

1. You have to be able to fit the smaller textures in the big texture without too much waste. Fitting four 64x64 textures in one 128x128 texture is easy but fitting three 64x64 textures in one 128x128 texture is going to waste some precious texture memory. Of course, you could try to use the remaining 64x64 space for textures on other objects.
2. It is possible that you have to use lower quality textures since combining them on a bigger texture may otherwise overflow hardware limitations.
3. It is harder for the artist to create models with this technique.
4. This technique is not possible if you have a tiling texture; i.e. a wall texture that is repeated accross a large surface. There are a few workarounds for this problem. For example you can pretile the texture on the super-material but that only works if there are not too many tiles required. Also you can split the polygons so that tiling is no longer required.

Lighting Considerations

When designing a map you also have to think about where to place lights. If you plan to use stencil-based lighting or hardware-based vertex lighting then you must be very careful to not exagerate the number of lights. Runtime performance with those two techniques depends on the number of lights and their influence radius. For this reason you are probably better off using lightmapped lighting in case you have a big map with a lot of lights.

With lightmapped lighting there is no runtime cost associated with having multiple lights (there is a slight memory cost associated with having many pseudo-dynamic lightmaps). A higher number of lights simply means that recalculating lighting will take longer.

This document was generated using texi2html 1.76.