Hi, I investigated this a bit more. I believe I have arrived at the bottleneck.
I stepped through the BufferObject code (which ImageStream uses via the derived class PixelBufferObject) and found that every time the dirty() method is called on osg::Image, in the osg::Texture that uses this osg::Image calls applyTexImage2D_load which in turn calls BufferObject::compileBuffer. Over here, I found that each time the texture's data is updated, it updates the PBO, by calling _extensions->glBufferSubData. While this works fine for texture updates that are not so frequent or for updates that don't involve changing the whole buffer, for updates as frequent as 25-50 FPS (as in my case), and ones involving change of the whole buffer (1920x1080 pixels), glBufferSubData starts showing drastic performance degradation as opposed to using glMapBuffer to get the PBO buffer and directly updating it. There are some posts out there which talk about the performance of glMapBuffer and using that buffer directly vs using glBufferSubData. One notable post is this one here (http://www.stevestreeting.com/2007/03/17/glmapbuffer-vs-glbuffersubdata-the-return/) from the Ogre3D dev lead. So, I decided to do a quick and dirty hack of the same into OSG. Please find attached a patch that does two things: a) (Hackily) exposes the buffer object's buffer b) Modifies the FFMPEG plugin to write to this buffer directly You should be able to apply the patch by simply doing: Code: patch -p1 < pboPatch on a folder containing the source of OSG 3.2.0 release. The speed improvements are quite spectacular! On the multimovie sample using the osgDB FFMPEG plugin, if we just remove all the nodevisitors and rendercallbacks to show stats (since these kill framerates), and simply log FPS while rendering all 10 videos, the frame rate improvements for 10 Full HD videos (1920 x 1080) videos are around 2000%(!!!). You can try the same by toggling the m_useDirectPBOMapping flag that I added to FFMPEGImageStream class from false to true. Without direct buffer mapping (i.e. using glSubBufferData) I get a niggardly ~20 FPS on an Intel Xeon (2.4 GHz), 8 GB RAM with AMD Radeon HD 6800. With the direct buffer mapping turned on, I get over 500 FPS. So, then, what I would like to know now, is whether there is a way of accomplishing the same direct mapping to PBO buffer (result of glMapBuffer) in OSG without having to resort to the hacks that I have. Thank you! Cheers, Abhishek[/code] ------------------ Read this topic online here: http://forum.openscenegraph.org/viewtopic.php?p=57681#57681 Attachments: http://forum.openscenegraph.org//files/3155_1387188273._194. _______________________________________________ osg-users mailing list [email protected] http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
