blob: a89176a62e6c36589fc22f597c4c46f319925a52 [file] [log] [blame]
Mesa 3.0 MITS Information
This software is distributed under the terms of the GNU Library
General Public License, see the LICENSE file for details.
This document is a preliminary introduction to help you get
started. For more detaile information consult the web page.
Version 0.1 (Yes it's very alpha code so be warned!)
Emil Briggs (
David Bucciarelli (
Andreas Schiffler (
1. Requirements:
Mesa 3.0.
An SMP capable machine running Linux 2.x
libpthread installed on your machine.
2. What does MITS stand for?
MITS stands for Mesa Internal Threading System. By adding
internal threading to Mesa it should be possible to improve
performance of OpenGL applications on SMP machines.
3. Do applications have to be recoded to take advantage of MITS?
No. The threading is internal to Mesa and transparent to
4. Will all applications benefit from the current implementation of MITS?
No. This implementation splits the processing of the vertex buffer
over two threads. There is a certain amount of overhead involved
with the thread synchronization and if there is not enough work
to be done the extra overhead outweighs any speedup from using
dual processors. You will not for example see any speedup when
running Quake because it uses GL_POLYGON and there is only one
polygon for each vertex buffer processed. Test results on a
dual 200 Mhz. Pentium Pro system show that one needs around
100-200 vertices in the vertex buffer before any there is any
appreciable benefit from the threading.
5. Are there any parameters that I can tune to try to improve performance.
Yes. You can try to vary the size of the vertex buffer which is
define in VB_MAX located in the file src/vb.h from your top level
Mesa distribution. The number needs to be a multiple of 12 and
the optimum value will probably depend on the capabilities of
your machine and the particular application you are running.
6. Are there any ways I can modify the application to improve its
performance with the MITS?
Yes. Try to use as many vertices between each Begin/End pair
as possbile. This will reduce the thread synchronization
7. What sort of speedups can I expect?
On some benchmarks performance gains of up to 30% have been
observerd. Others may see no gain at all and in a few rare
cases even some degradation.
8. What still needs to be done?
Lots of testing and benchmarking.
A portable implementation that works within the Mesa thread API.
Threading of additional areas of Mesa to improve performance
even more.
1. This assumes that you already have a working Mesa 3.0 installation
from source.
2. Place the tarball MITS.tar.gz in your top level Mesa directory.
3. Unzip it and untar it. It will replace the following files in
your Mesa source tree so back them up if you want to save them.
4. Rebuild Mesa using the command
make linux-386-glide-mits