[Mesa-dev] Merging translate and unnormalized-coords-hint?

[Roland Scheidegger]

On 17.08.2010 17:22, Luca Barbieri wrote:
>> What's wrong with addressing these needs respectively by:
>>
>> a) Adding a new pipe_texture_target enum PIPE_TEXTURE_RECT to capture
>> the GL usage (unnormalized, clamp).  Think about CL later.
> 
> I think this is essentially equivalent to my latest patchset in the
> following sense:
> - Your PIPE_TEXTURE_RECT is equivalent to my
> PIPE_BIND_SAMPLER_VIEW_UNNORMALIZED_CLAMP
> - Your PIPE_TEXTURE_2D is equivalent to my PIPE_BIND_SAMPLER_VIEW_NORMALIZED
> 
> I think a bind flag is the more appropriate way to represent this, because:
> 1. If we don't ever create a sampler view (e.g. render target) this
> doesn't a matter
> 2. Most checks for PIPE_TEXTURE_2D don't care about this, but only
> about the physical dimensions. Drivers that have multiple layouts
> already have internal flags for them.
> 
>> b) Adding a pipe cap to determine hardware preference for state-tracker
>> generated rendering (prefer TEXTURE_RECT vs TEXTURE_2D).  For API
> 
> I think a single global cap is insufficient:
> - r300g and i965 always prefer normalized coordinates for internal
> drawing (since they emulate unnormalized with a multiply) even for
> GL_TEXTURE_RECTANGLE
> - llvmpipe might perhaps always prefer unnormalized for internal drawing
> - nv30 requires normalized or requires unnormalized depending on texture layout
> 
> Hence, you at least need two global caps: one for the preference
> related to TEXTURE_RECT/BIND_UNNORMALIZED_CLAMP, and one for the
> preference related to TEXTURE_2D/BIND_NORMALIZED.
> 
> However, the state tracker itself has multiple preferences (it either
> wants unnormalized to avoid CPU divides if it uses a subrectangle, or
> it wants normalized to use 0, 1 constants if it uses the whole
> textures), so the driver ideally should be able to specify "no
> preference".
> 
> Thus, we need 4 bits of global caps at least.
> 
> IMHO using global caps leads to a messier situation that is instead
> cleanly avoid with per-resource caps.
> Are there any downsides to per-resource "caps"?

I'm not sure it's really worth bothering about all those cases to
optimize some divs or muls away. Hence one flag (for tex rect) would be
sufficient, otherwise just use normalized.

Roland