[Mesa-dev] gallium & texture rectangles

[Keith Whitwell]

On Wed, 2010-08-18 at 08:01 -0700, Luca Barbieri wrote:
> > I appreciate all the work you've put into looking at alternatives, but
> > at this stage I'm going to be firm - if PIPE_TEXTURE_RECT can be made to
> > work, that's the direction we should be taking.  I haven't seen anything
> > so far that suggests it isn't a feasible approach.
> Yes, it seems feasible, it's just a matter of choosing between tradeoffs.
> Right now I think it's more important to actually make any choice
> rather than making a specific choice.

Agreed.

> Effectively, implementing this is a matter of changing the resource flag to be a
> target, and change all existing code that checks for PIPE_TEXTURE_2D to check
> for PIPE_TEXTURE_RECT as well.

Sounds perfect.

> > What about this for a starting poing:
> >   - PIPE_TEXTURE_RECT texture target, following GL semantics
> OK.
> 
> >   - PIPE_CAP_TEXTURE_2D_NPOT  (nv30 should not advertise this)
> We already have this right now, it's called PIPE_CAP_NPOT_TEXTURES.

OK, my bad.

> > Is there anything more you actually require right now?
> State trackers must support both normalizations and have some criteria
> to choose between them.
> 
> I see these options:
> 1. Have a cap to decide whether to use PIPE_TEXTURE_RECT for internal
> textures, and use unnormalized if and only if the target is
> PIPE_TEXTURE_RECT

I like this very much.  

If there was a way to simplify it even further I'd like it yet more.  

For instance use 2D/normalized for internal rendering iff the driver
advertises CAP_2D_NPOT, otherwise use RECT/non-normalized.

> 2. Have a cap to decide whether to use PIPE_TEXTURE_RECT for internal
> textures, and use the two preference flags to choose the normalization
> 3. Have state trackers always use PIPE_TEXTURE_2D, and use the two
> preference flags to choose the normalization
> 4. Have state trackers always use PIPE_TEXTURE_RECT, and use the two
> preference flags to choose the normalization

Hmm - my intention was that TEXTURE_RECT implies non-normalized
coordinates -- so this doesn't look like it works.

> 5. Have state trackers use yet another new target, e.g.
> "PIPE_SURFACE", that would be intended for "2D operations", and use
> the two preference flags to choose the normalization
> 
> Instead of two preference flags, only one can be used, but this loses
> the ability for the driver to express no preference, and only saves a
> single line of code.

> My latest patchset uses option 3.
> 
> What is your preference in this?

I like (1) for reasons of interface simplicity.  If there was an option
zero, that would be better still...

Keith