Hi,
thank you for your quick answer. In fact, my project is part of a GIS and these
big transforms are are actually part of such a "local origin" approach.
However, I still have not understood where the inaccuracies that I see come
from. I have, however, reduced the critical part of my scene graph construction
to just a few lines of code:
// let bigMatrix1 and 2 be my transforms with numbers in the order of magnitude
of a million
// let scene be the subgraph that contains visible geometry
// let root be the top of my scene graph
// approach #1 (working fine)
ref_ptr<MatrixTransform> bigTransform1 = new MatrixTransform(bigMatrix1);
ref_ptr<MatrixTransform> bigTransform2 = new MatrixTransform(bigMatrix2);
bigTransform1->addChild( scene );
bigTransform2->addChild(bigTransform1);
root = bigTransform2;
// approach #2 (visible inaccuracies)
ref_ptr<MatrixTransform> singleTransform = new MatrixTransform(bigMatrix1 *
bigMatrix2);
singleTransform->addChild( scene );
root = singleTransform;
I have searched my entire project for Vecf's and Matrixf's. I use none of them.
The matrix I set in my subclass of CameraManipulator is also a Matrixf. As
another experiment, I have casted the big matrices down to Matrixf once, and
multiplied them with test vectors. The floating point errors that came out of
this where pretty much in the order of magnitude of the inaccuracies that I
witness when using the above approach #2 (with doubles). So, it really looks as
if OSG was using float when doing my transforms in 2 steps.
I am somewhat desperate for an answer to this... Do you have any ideas what
could go wrong here, or how I could investigate the problem any deeper? Is it
possible to apply a MatrixTransform node to a Vec3 'by hand'?
Greets,
Simon
PS: All of this is targeting iOS. Just in case this is important.
________________________________
Von: [email protected]
[[email protected]]" im Auftrag von "Robert Osfield
[[email protected]]
Gesendet: Montag, 16. September 2013 19:52
An: OpenSceneGraph Users
Betreff: Re: [osg-users] MatrixTransforms and MatrixD
Hi Simon,
The OSG by default uses double Matrices (osg::Matrix is typedef'd to
osg::Matrixd) for all internal transforms and camera matrices. The matrices
are all accumulated in doubles and passed to OpenGL as doubles. Most OpenGL
drivers will then cast the matrices down to floats when passing down to the
GPU, but as the OSG passes a fully accumulated modelview matrix the precision
is at least the best you can get.
To best handle scenes with large extents one breaks the scene into regional
tiles, each of which has a local origin and then a MatrixTransform above the
tile subgraph to place it in it's final world coordinates. The OSG is used
widely in the GIS market with many users handling whole earth database without
precision issues.
So... for you as long as you haven't deliberately compiled the OSG to use
osg::Matrix and a osg::Matrixf then you'll be uses doubles. How you manage
your scene graph and internal transforms will be the key, get it right as I
suggest above then you shouldn't have issues.
Robert.
On 16 September 2013 17:48, Voelcker, Simon
<[email protected]<mailto:[email protected]>>
wrote:
Hi,
I am using MatrixTransform nodes with matrices that contain quite large
numbers. When I stack two of these transforms, I notice severe floating point
inaccuracies in the scene, although I use double precision matrices to set up
my transform nodes. Is it possible that OSG uses only single precision here
(internally), and if so, how can I force it to use double precision?
I know I could multiply my matrices directly and use a single MatrixTransform
node, but I want to avoid this since it breaks my architecture.
Thanks in advance,
Simon
_______________________________________________
osg-users mailing list
[email protected]<mailto:[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
