Occlusion culling refers to the removal of objects (and faces) as early as possible from the rendering pipeline when they have been determined to be hidden from view (occluded) by geometry that is in front of them. This most often is taken to mean hidden from the point of view of the camera, but occlusion culling can also take place from other viewpoints such as the view of lights, in order to determine which objects can cast shadows.
while it is fairly easy to determine which objects are outside a view frustum (for example using a series of plane checks), determining which objects are in view within that frustum is a far more difficult problem. If we send everything to the GPU to render this will be resolved using brute force using the Z buffer, however if we can use cheaper techniques this can potentially save a lot of processing.
An important factor to note about culling is that it should be conservative. If something is in view, we cannot afford to miscalculate and cull it. This would result in objects popping in and out of view. However, if something is not in view, we can afford to make a mistake and not cull it, because the visual results will be the same. So very rare cases where we fail to cull an object are fine, if they make the overall algorithm faster.
As GPUs have become more powerful over time, the types of fragment shaders used now are far more complex than their counterparts in the early years of GPUs (early 2000s). Modern fragment shaders often use PBR rendering and many calculations, whereas early (fixed function) shaders were simply one or two texture lookups. Due to this, a very common optimization technique to save on fill rate has been to split rendering into 2 passes:
- A simple shader that only writes to the z buffer
- Full shader that tests the pre-created z buffer
While rendering everything twice appears counter-intuitive, it is often faster because the first pass has a very cheap shader, and as a result of the first pass, the expensive second pass only has to render those fragments that have already been determined to be visible.
As a result, if we naively draw an object behind something that will occlude it, it is wasting processing, however it is NOT AS EXPENSIVE as if the object were also put through the second pass (rendered in its entirety).
As such, the savings we get from camera-only occlusion culling depend on
- The proportion of objects we can cull away before the first pass
- The relative cost of the shaders used in pass (1) and pass (2)
As fragment shaders have become more expensive over the years, the cost of pass (1) relative to pass (2) has dropped. The cost of vertex processing relative to fragment shading has also probably dropped over the years. This means that the performance benefits from occlusion culling from the camera view have become less pronounced on high powered PCs and consoles.
However it can still be very important on low power devices such as phones, tablets and PCs that don't feature expensive GPUs, where there is often no choice but to run simpler shaders for high performance. And occlusion culling can be used to reduce lighting calculations, which will often help even in complex shaders.
There are also benefits to visibility determination outside of rendering.
Camera occlusion culling can potentially save performance by:
- Reducing the fragment shader costs of pass (1)
- Reducing the number of draw calls / material changes
- Eliminating some of the transparent objects
- Enabling other processing such as physics / AI to be switched off for invisible objects
It can also save performance for lighting, for instance by:
- Reducing the amount of shadow casters that need to be rendered into shadow maps
- Culling lights entirely that are not affecting the objects in view of the camera
These can be separated into two area, although they can potentially be used in combination:
- Geometrical methods
- Raster based methods
The classical example of geometrical methods is rooms and portals, as used in LPortal. However, there are other geometrical methods, such as anti-portals and tests against occluding geometry. Especially notable is that geometrical methods DO NOT BECOME MORE EXPENSIVE as screen resolution increases. Geometrical methods have become relatively cheaper at runtime over the years, as their processing cost is essentially fixed (depending on the complexity of the game level).
The major disadvantages of some of these methods is the need to build level geometry in order to take advantage, and mark areas such as portals (usually manually). They tend to work better with game levels that contain mostly static geometry.
The simplest raster based occlusion method is the Z Buffer, it determines the visibility on a per pixel basis. But there are raster methods that can achieve faster results than the conventional z buffer.
As GPUs are much faster than CPUs at raster based operations, it would seem to make sense to try and do this type of occlusion culling on the GPU. There is a problem though, if the CPU needs to read back the results of such a query from the GPU in order to decide what to draw. The GPU is typically rendering frames in a queue submitted by the CPU, and as such there can a delay of several frames between making a query and getting result, so the visibility information from say, 3 frames ago can be stale by the time it is read, resulting in objects being incorrectly culled and objects popping into and out of view.
An alterative is to perform raster methods on the CPU. This can be done without any delay, however it is usually necessary to optimize the method because of the lack of brute force available on the CPU. This can for example involve using a lower resolution version than the final frame to determine approximate visibility, or storing minimum and maximum z values within tiles instead of the exact z value at each pixel.
Raster methods have become relatively more common recently as techniques have improved, and the processing horsepower has become available. They have the advantage that in many cases there is less need to manually process game levels to take advantage of the culling. They also work well with totally dynamic worlds.
As the calculations involved in occlusion culling can be quite time consuming, it is worth noting that there is an alternative to making these calculations at runtime. If the level is mostly static, and can be split up into zones or cells of some kind, it is possible to precalculate all the objects / areas that are potentially visible from each cell. Then at runtime instead of a complex calculation, you can simply lookup the objects that are potentially in view, cull them against the view frustum and render!
This is extremely fast, however there are a few disadvantages:
- This only works for static geometry
- The pre-processing to calculate the PVS can be time consuming (but it only need be done at design time)
- The PVS can take significant memory - it may need to be compressed, or low resolution
- As visibility is determined from a cell (with volume), the results cannot be as accurate as visibility determined from a single camera viewpoint
- Without modification it can only give quite approximate culling of dynamic objects (similar problem to (4))
LPortal, as the name implies, is a geometrical solution using rooms and portals. At present it is calculated at runtime but I am planning to add optional PVS (this is unlikely to make a significant difference in most cases, as LPortal is very efficient at runtime).