I've just gone through the nightmare of upgrading from stock Ogre_1_8(pre-baked SDK from about 2 years ago) & CEGUI_0_1_7, to using home-compiled Ogre_1_9/CEGUI_0_8_3 via the latest available mercurial repos.
It's been... emotional.
(I still need to rework all of my CEGUI xml-based data files into the new formats, and change a boat-load of gamecode, but that's another story.)My reason for doing this is because I'm currently in a 'profiling & optimising & cleaning' phase with my game.
Basically, I'm spring-cleaning my entire codebase from the ground up before I add any new features.
The thing that has pushed me into the 'tidy-up' phase: Frame stuttering on adding entities to the scenegraph.
My game relies quite heavily on procedural mesh generation for the landscape. To mitigate the problems which that can cause I have shifted the mesh-gen into an asynchronous boost-thread. Finished meshes are then added to the scenegraph by the main thread during a FrameRenderingQueued event.
The problem is, I always get a frame stutter at this stage.
I get frame stutters even if I am only adding tiny meshes (about 300 verts), and even if I ensure that I only upload 1 tiny vert/index buffer to the GPU per frame.
(The stall is the same if I Lock>memcpy>Unlock OR use the Buffer->writeData method. With tiny meshes or HUGE meshes. The problem isn't the GPU upload.)
It seems that these stalls occur during (or shortly after) attaching the finished mesh to a scenenode, or when creating an instance/entity of the (already GPU resident) mesh.
I've noticed that the landscape demos for Ogre suffer from this exact same problem. I'm not using Ogre's landscape systems, but I have the same frame-stutter problem that they do.
So I was wondering if there are any known bottlenecks inside Ogre's scene management code that I should be looking at and tweaking?
Perhaps the adding of a new bounding box is triggering a huge (N ^ N) sort algorithm or something?
Any advice/tips/help would be hugely appreciated.