Hi everyone and thank you for your help. I successfully made a little program working with my matrices of real. Like Serge said : For each osg::Geode in my scene, I add it an update callback that compute the MODELVIEW matrix using Matrixr and then, convert this Matrixr to Matrixd before sending it to OpenGL. In this way all matrix computations are made with a great accuracy and OpenGL understand it. But now I have performance problem : OpenSceneGraph continue to do it MODELVIEW computations in double at during the update traversal. I can't stop it !
I already tried to use the depth partition that Mr Wright speaks about, but I met lot of problem with it : - bad clipping, - strange alpha blending, - performance issues (the scene is rendered multiple time each frame), - ... Now I need to find a way to replace the OSG matrix computations by mine... Stephane Lamoliatte a écrit : > Dear mister Osfield, > > I currently develop a space simulator working with very hight dynamic > range data. > During the development I met multiple floatting point number precision > problems. > > For now, I try to solve one of these that come from OpenSceneGraph : > For my scene, I need to use Matrices, Vectors and Quaternions with a > very big accuracy. (I detail an exemple of osg accuracy problem below). > I wrote a Matrixr, a matrix of real, wich is a Matrix using the GMP library. > The problem is that I can't simply replace all matrices in the code of > OpenSceneGraph by my Matrixr. > I try to change the "typedef Matrixd Matrix" by "typedef Matrixr Matrix" > in osg/Matrix, but It does not work because some class don't use the > Matrix but Matrixd or Matrixf. > Another solution would be to develop new osg::Transform and > osg::MatrixTransform using the real type and the Matrixr class. But I > don't know how to make osg considers them as transform node without > inherit them from the original osg::Transform node. > By the way, I want to know if someone have already try to do this kind > of thing, and/or if there is an accuracy node kit in development to help me. > > Best regard. > > > > > Here is a detailed example of my problem : > There is a star far far away from the center of the universe. > For some reasons I MUST have this scene graph : > [rootNode : osg::Group] -> [starTransform : osg::MatrixTransform] -> > [starModel : osg::Geode] -> ... > The Matrix in the osg::MatrixTransform is a Matrixd that contains a very > big translation vector, for example at (x=10^20, y=0, z=0) from the > center of the universe. > I want to move in the universe, close to this star at (x=10^5, y=0, z=0) > from the star and then at (x=10^20 + 10^5, y=0, z=0) from the center of > the universe. > To compute the MODELVIEW matrix, osg implicitly compute (x=10^20 + 10^5, > y=0, z=0) - (x=10^20, y=0, z=0) . > But in fact, osg works like if the camera is in the center of the > universe at (x=0, y=0, z=0). > Why ? Because osg computes matrix transforms with float and double > values. In double : 10^20 + 10^5 = 10^20. > So I can't freely move in my universe using Matrixd and that's why I > need more accuracy. > > > > -- > Lamoliatte Stephane. > _______________________________________________ > osg-users mailing list > [email protected] > http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org > _______________________________________________ osg-users mailing list [email protected] http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
