[osg-users] osgCuda: Using CUDA in osg - Problem with Threading
mick.svt at googlemail.com
Wed Mar 11 09:43:57 PDT 2009
my name is Mick an I'm currently involved in a project we call osgCuda.
osgCuda is based on osgPipe which can be added as a leaf node
anywhere into the osg-scenegraph and which enables you to jump to the
GPU for any kind of calculations using Nvidia's CUDA. The data
manipulated by CUDA is then afterwards available to the scenegraph for
further processing (e.g. rendering).
We plan to make this nodekit available to the public very soon so
everybody who might be interested in connecting osg and cuda is
welcome to test/use it.
At the moment we still have one problem to be solved:
In a single threaded application (i.e. osgViewer::Viewer::SingleThreaded)
everything works fine. In a multithreaded environment (e.g. DrawThreadPerContext/
CullDrawThreadPerContex/CullThreadPerCameraDrawThreadPerContext) we have problems
to clean up the memory allocated by CUDA at the end of the application.
The reason for this is obvious: any CUDA resources created through the runtime
in one host thread cannot be used by the runtime from another host thread.
In our case:
We allocate memory within a GraphicsThread. When the application closes the main
thread which also runs ViewerBase::stopThreading() then prevents CUDA from
releasing its resources because the GraphicsThread associated with CUDA doesn't exist
anymore. That is why we are not able to clean up the CUDA resources and why the application
crashes at the end while running the osg application multithreaded.
One solution we have verified would be to release the resources at the end of
GraphicsThread::run() by adding OperationThread::_operationQueue->releaseAllOperations().
This would work because we ensure that we have added one osg::Operation per GraphicsThread
which takes care of releasing the resources in its release-method.
But this would require a change of the osg core library within the GraphicsThread.cpp file.
// make the graphics context current.
GraphicsContext* graphicsContext = dynamic_cast<GraphicsContext*>(_parent.get());
Does anyone know another solution to our problem?
Robert, we guess it would not be that easy to change the core in the way we did (just adding
something like a "releaseAllOperations()" to the GraphicsThread's run() method to ensure that
all osg::Operation's will be notified before thread-termination)?
Thank you for your help.
Read this topic online here:
More information about the osg-users