[osg-users] KdTree support now checked in

James Killian James_Killian at hotmail.com
Wed Jul 16 23:24:57 PDT 2008


I've got some numbers that are interesting:

In a build before KDTree.. Matrixd::mult was the most used method:
1367 Matrixd::Mult
1080 Matrixd::makeidentity
982   Group Traverse

With KDTree on here is how the numbers stand:
1305 Group::Traverse
1228 Matrix::mult
990 Compute Bound
776 Transform::Compute Bound
728 MatricsD::Make Identity

So we can see how the KDTree has brought down the number of Matrixd::Mults, 
and so now Group Traverse can do more iterations.


James Killian
----- Original Message ----- 
From: "Robert Osfield" <robert.osfield at gmail.com>
To: "OpenSceneGraph Users" <osg-users at lists.openscenegraph.org>
Sent: Saturday, July 12, 2008 10:46 AM
Subject: Re: [osg-users] KdTree support now checked in


Hi Adrian,

It's great to hear you getting better performance.  How much is down
to more efficient intersections vs other changes to the OSG I cannot
say - you'd need to profile the relative costs of intersections/cull
and draw traversals etc.

KdTree are only used in intersections, so it doesn't change the cull
and draw traversals at all.  It's not something that is likely to help
much anyway, as the OpenGL pipeline is geared up to handling a
moderate number of relatively large sets of geometry, and rarely is
the actual geometry a bottleneck so it doesn't pay to just a do fine
grained view frustum culling.  Intersections a completely different
matter - ray intersection strongly favour fine grained
representations, hence why KdTree work so well.

Robert.

On Sat, Jul 12, 2008 at 4:41 PM, Adrian Egli OpenSceneGraph (3D)
<3dhelp at gmail.com> wrote:
> Hi Robert,
>
> i tested the whole afternoon with a really big terrain database (city of
> cannes). i rebuild our internet based visualisation plugin with latest
> openscenegraph (SVN of this morning) and the performance with
> the kdTree datastructure enabled is much better than with elder OSG 
> version.
> i don't yet understand why the performance is much faster, of course the 
> new
> database pager concept could be responsable for the behaviour but i don't
> think so. now the frame rate is really constant, no longer view dependent
> for our motion model, also for the terrain manipulator. each of this 
> motion
> model are using intersection visitor, may this was the reason why the 
> frame
> rate was never constant, was view angle dependant. one pixel changed and 
> the
> frame could dorp down or vise versa. but know the frame rate is nearly
> constant, for this huge database. of course still view dependant, but 
> little
> change doesn't no longer change the frame rate (60/20) between small
> changes, this isn't question of culling,... i guess this was the
> intersection test, if we passed through a huge object, the fps drops down.
> this is greate.
>
> next question. do you use the KDTree to cull objects against the frustum.
> big objects (Triangles could better culled) if we can intersect easely a
> plane against the kdtree.
>
> /adrian
>
> 2008/7/12 Robert Osfield <robert.osfield at gmail.com>:
>>
>> Hi J-S,
>>
>> On Sat, Jul 12, 2008 at 1:09 AM, Jean-Sébastien Guay
>> <jean-sebastien.guay at cm-labs.com> wrote:
>> > I think I may be able to help a bit regarding the higher-level setup 
>> > and
>> > bookkeeping changes needed to speed things up on that regard. One thing
>> > I
>> > noticed before is that creating a new Intersector and
>> > IntersectionVisitor
>> > each time is costly, and instead keeping static or cached instances and
>> > using the reset() and setStart()/setEnd() methods is faster. There may
>> > be
>> > some other similar things that can be done too, we'll see what I can 
>> > dig
>> > up.
>>
>> It's interesting how different bottlenecks pop up, in this case
>> creating objects on the heap is clearly showing itself as a bottleneck
>> whereas the old brute force intersection code was so slow that this
>> wasn't a significant factor.
>>
>> Caching visitors and intersectors can certainly help, although one
>> does need to be careful about threading issues with doing this, as the
>> IntersectionVisitor and Intersectors aren't thread safe, so you'd need
>> a cached instance per thread.
>>
>> Optimization could possibly be done with the data structure used to
>> store the intersection results, this would unfortunately break the API
>> compatibility.  Changing that Intersector's are cloned might also help
>> performance when dealing with scene graphs have transform - the clone
>> is required to move the intersector into the local coordinate system
>> of transforms subgraph.
>>
>> There other things one can do like grouping intersectors, or using
>> intersection coherency.
>>
>> Robert.
>> _______________________________________________
>> osg-users mailing list
>> osg-users at lists.openscenegraph.org
>> http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
>
>
>
> --
> ********************************************
> Adrian Egli
> _______________________________________________
> osg-users mailing list
> osg-users at lists.openscenegraph.org
> http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
>
>
_______________________________________________
osg-users mailing list
osg-users at lists.openscenegraph.org
http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org




More information about the osg-users mailing list