Question: how to synchronise between the game and render thread?

In today's blogpost, I'd like to do something radically different: I'm going to ask you a question. I know that quite a few of the visitors of my blog are hardcore programmers, and since I have a complex situation for which I would like to improve the performance, I figured I could ask for advice here! :)

The problem is pretty complex, though, so I figured a short blogpost would be needed to really explain it. In short, this is the question:

What is the most efficient way to synchronise renderable objects between the game thread and the render thread?

There are a couple of subtleties you need to know to answer this for our situation, though. Let me start with a scheme that shows the approach to threading in the Ronitech (our in-house multiplatform 2D engine), as we use it in Awesomenauts on the Xbox 360 and Playstation 3:



This works fine and we managed to get 60fps on console this way. However, the copying phase takes relatively long, so I am still wasting a lot of performance there. Since we will also use the Ronitech in all our future games, improving here will pay off greatly in our future projects.

There are a couple of specific requirements that I have to make this work with the rest of our engine:

• The gameplay thread must be able to modify (and even delete!) all renderable objects. At the same time, the renderer must still be able to render each object how it looked during the previous frame.
• Within a single frame, the same object might be rendered to different viewports in different ways. For example, a character might be partially transparent in one viewport and fully visible in another. The game can register itself as a PreViewportRenderListener and thus change objects in between the rendering of two viewports.

The copying phase as I have it right now, works as follows: for every renderable object I have a clone for every viewport it might be shown on. If there are lots of viewports, this can mean lots of clones. I don't have any memory problems, so I don't mind the clones. Now every frame during the copying phase, for all the objects that are actually visible in a viewport, all the properties of the original object are copied to the clone. This is the step that costs the performance, since in three player splitscreen this can mean copying all the properties of 2000 objects every frame, and each object has dozens of properties! I currently just call the PreViewportRenderListeners during the copying phase, in between copying data for one viewport and copying data for the next viewport.

So, what alternatives are there to just copying all the relevant data for all visible objects every frame? Or should rendering in a separate thread be set up differently all together? Please comment below if you have interesting ideas for this!