[Dri-devel] Selecting texture formats (was Re: CVS Update: xc (branch: trunk))

[Ian R. <id...@us...>]

Felix Kuehling wrote:

> Log message:
>   Improved internal texture format selection in mga, r128, r200 and radeon:
>   - try to choose a texture format with same color depths as application supplied
>     data in order to avoid loss of color information (mga, r200 and radeon only)
>   - don't override application's choice of internal color depth unless force16bpt
>     is GL_TRUE. For now it is always GL_FALSE, but will be configurable soon

I've always felt like the various ChooseTextureFormat routines could be 
refactored into a single generic routine.  The way you restructured the 
code give me an idea how to do it.  For a card that supports every 
possible texture format, we would have a 2D matrix.  One axis is the 
internalFormat as a concrete type (i.e., GL_RGB8), not as a generic type 
(i.e., GL_RGB).  The other axis is the type as a packed format (i.e., 
GL_UNSIGNED_SHORT_5_6_5), which might be derrived from the application 
supplied type / format.  There are some made-up types on that axis as 
well.  For example, {GL_RGB, GL_UNSIGNED_BYTE} maps to a nonexistant 
GL_UNSIGNED_24_RGB.  Each cell in the matrix is a MESA_FORMAT_ value 
that meets the requirements and has the least conversion cost.

Obviously, no card supports all of the formats.  Each cell then becomes 
a sorted list of MESA_FORMAT_ values.  The added input to the generic 
function is a list (probably as a bit-mask) of the formats supported by 
the hardware.  The generic routine would run through the array in the 
selected cell until it finds a format that is suppored by the card.

The value of DRI_CONF_TEXTURE_DEPTH32 would change the way generic 
internal formats are converted to concrete formats, and 
DRI_CONF_TEXTURE_DEPTH_FORCE_16 would select a subset of the available 
texture formats (i.e., a bit-mask that only had 16-bit formats).

The other missing piece is support in Mesa for the byte-swapped formats 
and CI4 (which is supported by Rage128, i830, & MGA hardware).

Anyone have time to work on such a generic routine? :)