Re: Bug 895099 proposed fix

[Frank C <li...@si...>]

On 16-Jun-04, at 1:47 PM, Dair Grant wrote:

> Frank C wrote:
>
>> I've gained some speed by removing all NULL binds and wrapping
>> glBindTexture in a GL_Util*** routine that compares the last texture 
>> to
>> the requested one, avoiding redundant uploads.
>
> The texture won't be uploaded when it's bound: unless it got bumped out
> of vram by something else since the last time it was bound, it'll still
> be there.
>
>> This helps the transparent pass quite a bit since it often
>> binds/unbinds the same texture over and over (for every triangle) when
>> drawing a mostly self-sorted mesh.
>
> I guess it removes the need to call GL, but I would expect 
> glBindTexture
> to not do anything if the same object is already bound - although I
> suppose it could be implementation dependent (perhaps some drivers do 
> no
> checking at all: seems unlikely though, checking for a no-change is 
> such
> an obvious thing to do if you were writing glBindTexture).

I would hope drivers would do this as well, but I was under the 
impression you shouldn't rely on this.

I've done some testing with OpenGL Profiler on Mac OS X 10.3.4 with a 
GeForce3 64MB card. The scene consisted of one model with three 
textures (i.e. 3 trimeshes) with approximately 1500 triangles total. 
Below are the average timings for glBindTexture from the "% GL Time" 
column:

CVS source 06/17/04
  All solid textures: 1.3 %
  All transparent textures: 8.6 %

CVS source 06/17/04, with redundant calls to glBindtexture and NULL 
binds removed)
  All solid textures: 0.5 %
  All transparent textures: 1.7 %

This doesn't translate to a huge speed up for the entire application 
but it's measurable.

Frank.