next »

[rox]

Hello!

I'm porting my Java3D software to use Xith3D. Everything seems to be working pretty well. However, I've got some serious performace issues. I checked out the version from SVN and noticed thaht it seems to be under constant evolution. That's nice, and I'm hoping these issues to be resolved soon

Anyway, browsing through the code and profiling it, I noticed that if the tranform of a group is changed, the bounding sphere is updated by scanning through all the vertices in the children. While it always produces correct results and "tight" bounding volumes, it also effectively kills the performance. The better solution would be just to transform also the bounding volumes and trust that the volumes are as "tight" as possible. So as I understand, there doesn't seem to be a mechanism to make a difference between the dirty bounds and transformed bounds case. In the end, they are different things, right?

Any plans on making improvements on this soon, or should I make a "quick" fix for myself?

- rox

[rox]

Ok, seems like a simple fix works at least in my case. In Shape3D.java (SVN Feb 16.2007) in method updateBounds(boolean onlyDirty) I replaced:

Code:

bsph.compute( geom );

with:

Code:

if (getTransformGroup() == null)
    bsph.compute( geom );
else
    bsph.transform(getTransformGroup().getTransform());

Got immediately LOTS more of FPS I don't know if this is a generally valid fix... Maybe someone with more insight to the engine knows.

- rox

[horati]

I don't know that much about the internals of the picking library; however, that change feels dangerous to me for general use.  When you created your shapes, did you try making them Shape3D.Types.STATIC?  The default is Shape3D.Types.DYNAMIC which has the least optimizations.  There are a couple levels in between but your code change clearly implies the STATIC choice.

[rox]

I don't know that much about the internals of the picking library; however, that change feels dangerous to me for general use.  When you created your shapes, did you try making them Shape3D.Types.STATIC?  The default is Shape3D.Types.DYNAMIC which has the least optimizations.  There are a couple levels in between but your code change clearly implies the STATIC choice.

Thanks for the reply. I did try that but it makes no difference. But you are correct, the STATIC should be my choice (which it wasn't previously). I actually figured out the fix to be dangerous as it fails, for example, if the geometry is modified. For my case it is good enough at the moment though..

- rox

[Marvin Fröhlich]

Ok, seems like a simple fix works at least in my case. In Shape3D.java (SVN Feb 16.2007) in method updateBounds(boolean onlyDirty) I replaced:

Code:

bsph.compute( geom );

with:

Code:

if (getTransformGroup() == null)
    bsph.compute( geom );
else
    bsph.transform(getTransformGroup().getTransform());

Got immediately LOTS more of FPS I don't know if this is a generally valid fix... Maybe someone with more insight to the engine knows.

I will definitely implement a fix inspired by your code. But your code can't work for general case. Consider, you're modifying the TransformGroup's Transform3D twice. Then you would first compute the BoundingSphere, then transform it after the TG has changed and then after the other change, it is again transformed.

The "untransformed" BoundingSphere has to be cached and the "real" BoundingSphere has to be built from it.

I don't know that much about the internals of the picking library; however, that change feels dangerous to me for general use.  When you created your shapes, did you try making them Shape3D.Types.STATIC?  The default is Shape3D.Types.DYNAMIC which has the least optimizations.  There are a couple levels in between but your code change clearly implies the STATIC choice.

This won't make a difference.

Thanks for the reply. I did try that but it makes no difference. But you are correct, the STATIC should be my choice (which it wasn't previously).

Oh, I just forgot again to write some lines about Shape3D.Types in XIN. I will do it these days. If I don't, please remind me.

Marvin

['n ddrylliog]

Really looking forward for this fix. I think it'll increase performances a lot..

[Marvin Fröhlich]

The new bounds fix also includes a tweak like above.

Marvin

[rox]

Thanks for the quick fixes. My application works now much better

- rox

[Marvin Fröhlich]

Thanks for the quick fixes. My application works now much better

Cool, you're welcome .

Marvin