Re: Mipmap filtering

[Joseph J. S. <jo...@st...>]

[Please note: the following is actually from Frank Condello, who 
finds himself unable to successfully post to this list at the moment, 
perhaps due to a sourceforge.net issue like #958076 which I ran into 
last week.  So he sent it to me directly and asked me to post it for 
him.  -- JJS]

From: Frank C <li...@si...>
Date: Thu, 3 Jun 2004 13:57:34 -0400

On 2-Jun-04, at 10:53 PM, Joseph J. Strout wrote:

>OK, let's talk about this issue:
>
>>1. Proper mipmap filtering 
>><http://sourceforge.net/tracker/index.php?func=detail&aid=905678&group_id=45158&atid=442052>. 
>>There are several ideas in this request, but the most important one 
>>is to automatically generate mipmaps from a Pixmap texture, just as 
>>QD3D 1.6 does.  (And then of course if you don't want mipmaps 
>>generated for you, you can use a MipMap texture instead.)
>
>My understanding is that in QD3D 1.6, when you use a PixmapTexture, 
>mipmaps are generated under the hood for you, and the renderer then 
>selects the one most suited to the object's usage in the scene.
>...
>That suggests that maybe the renderer code is the right place to 
>look -- perhaps IRTexture.c?  Perhaps modify 
>ir_texture_convert_pixmap so that it generates an OpenGL mipmap set, 
>as ir_texture_convert_mipmap does?  If we do it here, is all that 
>scaling and conversion going to happen on every frame, or just once 
>for each time the texture is edited?  And if there are multiple 
>objects sharing the same texture, will it happen just once for the 
>whole lot?

Easiest way is to replace the glTexImage2D call with a call to 
gluBuild2DMipmaps and they'll be created auto-magically. You can 
minimize them manually and call glTexImage2D multiple times as well, 
which is often preferred since the result will be more predictable 
regardless of OS, hardware, or OGL implementation. A custom algorithm 
can also be tweaked to give better results than gluBuild2DMipmaps.

Creating the mipmaps is only the first step however. The 
cachedTexture->qualityFilter settings in the ir_texture_set_state 
function have to be changed to take advantage of mipmaps. This can be 
done in a couple ways, though I'd recommend the following:

Create a new private struct:

typedef struct TQ3QualityFilter {
	GLuint					magFilter;
	GLuint					minFilter;
} TQ3QualityFilter;

Then replace the qualityFilter member of the TQ3CachedTexture struct 
with TQ3QualityFilter and modify the affected routines accordingly; 
ir_texture_convert_rave_filter, ir_texture_set_state, 
IRRenderer_Texture_Rebuild, and ir_texture_cache_add.

The actual values of the mag/min filters would follow the 
recommendations outlined in the bug report.

I've actually gone ahead and did this on a local copy and it seems to 
work just fine (currently using gluBuild2DMipmaps).

Frank.