The first part of the fix is to assign VBO/EBO's upon loading the array in the Nif reader. This avoids triggering the 'addVertexBufferObjectIfRequired' code path in osg::Geometry which has the race condition when two threads add the same Array at the same time. Essentially, we want the Arrays to be 'const' when they come out of the Nif reader.
The second part of the fix is to make sure not to create empty arrays in the Nif reader (importantly, not assigning a VBO to the empty array). This empty array would be deleted when the NIFFile is cleaned up, and the detachment of the VBO assigned to it (which is still in use by other arrays) would cause threading issues.
This rare crash bug was first introduced with commit a7c5beb7c5. When using OSG dev version 3.5 the crashes were a little more prevalent, because 'addVertexBufferObjectIfRequired' in osg::Geometry is now used even when VBO's are disabled (as part of the VAO support changes).
The previous commit broke some magic effects (e.g. shield). The controller's emit start/stop times should control the emitter, while the absence of the ParticleSystem's autoplay flag should freeze the whole particle system if no valid controller source gets assigned. This revised version works for both the Shield effect and the "dagoth ur death sparkles" effect.
Since commit e8662bea31, we're using OSG functionality that contains an unfixed crash bug in version 3.2. The bug is fixed in version 3.4 (OSG commit 6351e5020371b0b72b300088a5c6772f58379b84)
Instead use getImage and let the caller create the Texture. Sharing of textures is then handled in post by the SharedStateManager.
This is closer to what the OSG serializer does.
Streamlines the TextureManager and will make it easier to multithread.
Practical benefits:
- Filter settings are now applied to native OSG format models. These models do not use TextureManager::getTexture2D since the model itself specifies a Texture.
- The GUI render manager will be able to use its own separate textures, making it easier to turn off filtering for them.
Cache the current position in the animation track and attempt to reuse it in the next frame.
Decent speed up for the Update phase, about 0.3 ms faster in Balmora.
Unlike what I expected, the osgUtil::UpdateVisitor is set to traverse all children (not only active children). The FrameSwitch was thus traversing both RigGeometries part of the double-buffering scheme, rather than only the one active in the current frame.
Seems wrong to me, but MW appears to do it that way. Without this fix, the light_de_candle_08_64 from http://www.nexusmods.com/morrowind/mods/41654/ has flame particles in the wrong spot.
OSG 3.4 adds the ability to place Drawables directly in the scene graph, without a Geode decorating them. Leveraging this should give a small performance boost, because the redundant Geodes increase culling overhead.
There is still an oustanding issue with the RemoveDrawableVisitor no longer working correctly, because Drawables can have multiple parents.
Works around a compiler warning with OSG 3.4:
warning: base class 'class osg::Callback' should be explicitly initialized in the copy constructor [-Wextra]
With no default argument for osg::CopyOp&, the compiler no longer sees the function as a real copy constructor and stops warning about the missing virtual initializations.
We don't care about this warning because there is nothing interesting to initialize in the osg::NodeCallback base anyway.
A proper fix for the warning would require to inserting OSG_VERSION conditional compiling all over the place, that is as long as we are still supporting OSG 3.2.
