Re: [Dri-devel] Application performance: open source vs proprieta ry?

[Pawel S. <pa...@th...>]

On 2003.11.27 06:20, Alexander Stohr wrote:
> [snip]
> There is a comparison from R.Scheidegger, it is linked at
> the dri web page at Documents page, "Other Documents"
> section comparing several Radeon 9000 capable drivers
> including the DRI drivers. [snip]

This review is very good but I would really like to see the numbers -  
and some analysis! - for recently released game "Savage" that was even  
announced on Slashdot. I tried it on my ATI8500LE (I know it is not a  
state-of-the-art graphics card any more) and I have got mixed feelings.  
Binary ATI drivers appeared to have problems with character rendering  
(but performance was almost decent) and DRI drivers could render only  
single frame every few seconds and had usual problems with s3TC but the  
program can be asked not to compress textures - I wish it could detect  
the missing extension automatically as ID-software games do.

I contacted S2Games and they said that: "The open-source DRI drivers  
don't support the Vertex Buffer Object OpenGL extension [...].  We fall  
back to other methods like VAR and such on cards that don't support  
VBO, but really it comes down to the driver optimizing for high poly  
counts. We use the same code in both Linux and Windows (since it's all  
OpenGL), and in Windows we get fast speeds on both ATI and NVidia.  In  
Linux, ATI seems like a non-starter on both driver sets, whereas the  
binary NVidia drivers work like a charm. In games like RTCW, they deal  
in the realm of 2,000-5,000 polygons per frame.  In more recent games  
like Savage, we push more like 100,000 polygons per frame, since we're  
doing full outdoor scenes. [...] The VAR that we use is the   
NV_vertex_array_range extension.  If the driver don't support that, I  
believe we fall back to the CVA extension (compiled vertex array)."

I can imagine many applications that require many polygons to be  
displayed: CAD, molecular modelling and I understand such extensions  
are not patented, are they? Does anybody know why  
GL_EXT_compiled_vertex_array cannot give same performance as  
GL_EXT_vertex_buffer_object? Would that be much work to implement this  
extension?

Pawel