OpenGL performance

[degasus]

Hello

I wonder, why the OpenGL performance is so slow, but my slow gfx card is only 20% busy. So I started to debug:

With apitrace (nice tool, also working with d3d, try it), I sniffed all opengl calls while rendering one frame of the mario galaxy intro and save it here: http://markus.members.selfnet.de/dolphin/
There is also an short version with the typical rendering of some polygons.

In this typical rendering, 3 operations are done:
- overriding all texture parameters
- overriding all shader parameters
- rendering mostly 30 vertexes

Ok, I see, the last is importent. The shader parameters also often changes. But why changing the texture every 10 polygons to the same texture?

Is it possible only to set the texture parameters on changing the texture? I think this would speed up the opengl plugin enormously.

Markus

[Starscream]

Trying to get to the bottom of issues like this is a good idea, unfortunately you looked right passed the most important thing, your PC. It's far too weak to play anything well enough.

[delroth]

Trying to get to the bottom of issues like this is a good idea, unfortunately you looked right passed the most important thing, your PC. It's far too weak to play anything well enough.

Not relevant.

In this typical rendering, 3 operations are done:
- overriding all texture parameters
- overriding all shader parameters
- rendering mostly 30 vertexes

Ok, I see, the last is importent. The shader parameters also often changes. But why changing the texture every 10 polygons to the same texture?

I think Dolphin is betting on the graphics driver optimizing redundant state changes. However I'm not sure how common this optimization is and if it is actually safe to assume that most used drivers do it (/me adds "check the Mesa source code" to his TODO list). In theory Dolphin could keep track of the state itself to avoid redundant changes, but this will add a substantial amount of code so I would like to be sure that it is actually useful before implementing this.

Also, I haven't checked but I'm pretty sure DX9 and DX11 do the same.

[neobrain]

All HW accelerated backends depend on the driver optimizing redundant state changes, yes. I'm pretty sure that all relevant graphics drivers are doing this though. AFAIK for d3d, it's not even the drivers doing that work - but in fact the d3d runtime itself.

[delroth]

Ok, according to some searches I did really fast GL drivers actually have a hard time optimizing redundant state changes and a lot of people reported a nice performance boost after tracking changes in their engine instead of letting GL do it.

I'll try to work on some improvements in the next few days. Hopefully it will mean a nice perf boost for the GL renderer (10% would be nice ).