[osg-users] First steps to improve culling

[Robert Osfield]

Hi Alexandre,

On 12 November 2015 at 20:30, Alexandre Vaillancourt <
alexandre.vaillancourt.list at gmail.com> wrote:

> Thanks for your reply, and of course I should have specified the version.
> It's 3.2.0, we're planning to update to the most recent one within the next
> 3 months.
>
> I have attached the screenshot. I'm not authorized to post anything not
> approved by the higher end so the visuals have been "painted", but the
> stats are there.
>

>From the screenshots the cull and draw dispatch don't look particular high,
but the draw GPU is clear bottleneck and will be the primary cause of slow
down.

The number of objects that the OSG and hence OpenGL is dealing with isn't
massive so optimizing the scene graph structure probably won't gain you too
much.  Not to say you might not be able to optimize things further but for
now I'd suggest you best spend your efforts looking at how to optimize what
is happening on the GPU.

You could look at threading - CullThreadPerCameraDrawThreadPerContext would
be worth trying just to see how it changes things.  As you are GPU limited
I don't expect this to make a big differents but since calling
"viewer.setThreadingModel(osgViewer::Viewer::CullThreadPerCameraDrawThreadPerContext"
is single line addition to your app it shouldn't take to long to test out.

On the GPU optimization side there are lots of different things that can
cause bottlenecks, each one will have a different set of solutions.  It's
rather open ended topic, so you'll need to do a range of benchmark tests to
find out what is the main bottleneck on the GPU.  One easy test would be to
change the resolution of the different windows - such as half in each
dimension to cut the fill requirements to a 1/4.  If you run the benchmarks
in this configuration and look at how the GPU stats change you'll get an
idea of relative load on vertex vs fragment processing.  For instance if
the performance doesn't change much when you change resolution then you are
likely to be vertex processing limited, while if the performance does vary
significantly then you'll be fill limited.

Robert.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openscenegraph.org/pipermail/osg-users-openscenegraph.org/attachments/20151113/17a1ea20/attachment-0003.htm>