Very Computer

Some OpenGL software I have written for Windows NT renders faster in some cases on software only
than with an OpenGL accelerator.  For objects with a large number of polygons the accelerator is
much faster, but for simple objects of fewer than 20 polygons, using software only renders these
faster.  The acceleration is so much slower that a 33 MHz 486 is just as fast.  The machines I
tested on were 200 MHz Pentium Pro's with either Intergraph or a Glint accelerator.  I am trying to
animate some simple objects by repeatedly rendering them.  The rendering uses double buffering and I
call SwapBuffer.  I don't call glFinish, glFlush, or any delays.  Each time I render I call
glRasterPos3f, recompute the near and far clip, and call gluPerspective.

--

Software Consultant     www.vianova.com/~vianova
Via Nova                OpenGL, Open Inventor products for Windows 95/NT
Dallas, Texas

We developed a sofware package which uses OpenGL on a group of Indigo's
running Irix5.3.  Unfortunately, when the same software is run on the
Onyx which is on Irix6.1 there are numerous rendering errors ranging
from vertex normal errors to z-buffer errors to polygons going missing.
I suspect the errors are in the window initialization --maybe something
changed in the window defaults that would require a change in the code.
 Has anybody been through this or even just heard of it?

I forgot to state my question in my previous post.  The question is, why is using software
only sometimes faster than the OpenGL accelerator?  What am I doing wrong?

Thanks:

--

Software Consultant     www.vianova.com/~vianova
Via Nova                OpenGL, Open Inventor products for Windows 95/NT
Dallas, Texas

There are a lot of trade-offs with hardware acceleration that can cause hw to be slower than sw in some
cases, but the example you give doesn't fit any of the ones I'm aware of.

In general, we find that hardware that requires the CPU to do triangle setup will not run faster than
software if the triangles are only a few pixels in size.  This cases, for example, SPEC viewperf benchmarks
to run w/o much speed up -- and in some cases actual slow-down -- with cards like the Millenium.  Hardware
of this kind is much more effective with larger polygons, when the high fill-rate of the hardware dominates
the drawing time; for example, VRML viewing.

The Intergraph machines, however (as well as GLints with the Delta chip) do handle triangle setup on the
board, and I would expect them to outperform software in virtually all contexts.  One thought: there may be
some setup time associated with using the hardware and/or it's associated driver, which is not being
amortized effectively over a display of only 20 polygons.  Which OS and version are you using?

Steve Wright
OpenGL Program Manager
Microsoft Corporation

First, I strongly recommend you upgrade your Onyx to IRIX 6.2.  There
were lots of bug fixes and performance improvements for OpenGL on
RealityEngine between 6.1 and 6.2.

Second, I'm not sure what might be causing the problems you describe,
but your guess about window initialization is a good place to start
looking.  On some Indigo systems you'll get a depth (Z) buffer
automatically, whether or not you request one.  On RealityEngine
systems, you won't get a depth buffer unless you request it
explicitly.  Does your list of parameters to glXChooseVisual() include
an explicit request for a nonzero-size depth buffer?

Allen