Fred Dech (fred@cerulean.bvis.uic.edu)
Fri, 6 Aug 1999 12:40:23 -0500 (CDT)
i guess i still don't follow exactly what is being describing.  if i'm
working with a 2byte 512x512 image list, and i'm using, say, 1 128x128rgba
2D texture, i can get acceptable performance using various different brick-
sizes.  but after i exceed 128 data images, whether with a brick size of
32x512x512 or 64 or 128, performance hits a logjam.  now a 64zsize brick
volume of 129 slices performs alot better than a 128zsize brick for that
number of slices, but i can't live with anything less than about 12 fps,
and i'm getting more like 1.
i tried not running voAppearanceActions::volumeOptimize().  i tried
changing the parameters to INTERLEAVE_2WAY or SHARED_OBJECT.  if anything
performance got worse.
if i remove the 2D textures, i get performance back, but there goes
the visual ID structure of my GUI.  even an itibit 2x2pixel 2D texture will
cause the app to grind to a stand-still.  i have to resolve this.  i can't
make an effective GUI without textures.  any more advice?  
thanks.
--fred
Chikai Ohazama wrote:
> 
> Hmmm... I guess I misunderstood the problem.  If you're willing to live
> with a 32 MB texture and all your other 2D textures fit into 32 MB, then
> that works too.  I just reread Fred Dech, and realized that he's using
> Performer.  For those of you at SIGGRAPH, go to the Friends of Perfomer
> session and you'll find a pleasant surprise.  We've got a Volumizer node
> working in Performer.
> 
> ________________________________________________________________________
> Chikai J. Ohazama, Ph.D.             SGI 
> Member of Technical Staff            2011 N. Shoreline Blvd. ms 525
> Advanced Graphics Division           Mountain View, CA  94043.1389
>                                      Telephone:  (650) 933-6533
>                                      FAX:        (650) 932-0511
> 
> 
> 
> On Thu, 5 Aug 1999, JB West wrote:
> 
> > The following is NOT a great idea. I make a LOT of use of 2D textures in
> > my application. Thus, what I do is cut the Volumizer texture size to 32Megs,
> > interleaved, shared texture objects; works great. With the latest Volumizer
> > & patch, I can get > 200Mbytes/second average download rate on an IR2E/2RM's.
> > 
> > 
> > On Aug 5,  1:20pm, Chikai Ohazama wrote:
> > > Subject: Re: texture quandery.
> > > IR doesn't manage textures well at all, so you have to do it yourself.  So
> > > here are some suggestions on possibly making it faster.  Try taking out
> > > the volumeOptimize command in the init call.  This will make it so that
> > > texture-objects are not used, there forcing a download each frame.  Also
> > > try change it to INTERLEAVE_2WAY.  It will still be slow, but hopefully
> > > not as slow as you're going now.
> > >
> > > Hope this helps.
> > >
> > > ________________________________________________________________________
> > > Chikai J. Ohazama, Ph.D.             SGI
> > > Member of Technical Staff            2011 N. Shoreline Blvd. ms 525
> > > Advanced Graphics Division           Mountain View, CA  94043.1389
> > >                                      Telephone:  (650) 933-6533
> > >                                      FAX:        (650) 932-0511
> > >
> > >
> > >
> > > On Thu, 5 Aug 1999, Fred Dech wrote:
> > >
> > > > hi.
> > > >
> > > > my dev platform is IR w/4 RMs.  displaying large volumes.
> > > > application is Performer-based, using a heavily modified version of
> > > > Volumizer Performer node from the Performer example program.
> > > >
> > > > symptoms:
> > > > 1. 512x512data at 128 (1 brick) or less slices, great performance (~20
> > fps).
> > > > 2. with over 128 slices (>1 brick) performance goes south, big-time
> > > >    (~0.15 fps)
> > > > 3. remove small texture(s) from GUI portion of application and great
> > > >    performance returns to multiple brick volume (512x512ushort) over
> > > >    128 slices.
> > > > 4. i uninstalled patch3696, just to check.  no change.
> > > >
> > > > conclusions:
> > > > 1. i'm sc**wed! :o :^)
> > > > 2. the texture(s) in my GUI are killing Volumizer performance when more
> > > >    than one brick is required to hold the voxel data.  but why?
> > > >
> > > > any suggestions would be greatly appreciated.
> > > >
> > > > --fred
> > > > --
> > > >   Fred Dech   fdech@uic.edu
> > > >   VRMedLab
> > > >   (312) 413-3092: fax (312) 996-8342
> > > >
> > >-- End of excerpt from Chikai Ohazama
> > 
> > 
> > 
> > -- 
> > JB West | jbwest@lgc.com | downsized .sig
> > 
> 
-- Fred Dech fdech@uic.edu VRMedLab (312) 413-3092: fax (312) 996-8342
This archive was generated by hypermail 2.0b2 on Mon Nov 01 1999 - 14:21:40 PST