[Mesa-dev] gallium-array-textures branch

[Roland Scheidegger]

Corbin Simpson wrote:
> On Mon, Jun 14, 2010 at 4:09 AM, Roland Scheidegger <sroland at vmware.com> wrote:
>> I'm planning to merge this branch soon, though I'll fix some broken things
>> first. In any case, the probably most fundamental change is that
>> get_tex_surface not only got renamed (to create_surface) but also moved to
>> context. In short, pipe_surface is not (or no longer as there were free
>> standing surfaces once upon a time in gallium) some kind of "2d surface"
>> which should be used in (non-rendering) state trackers to move any kind of
>> images around. pipe_surface is meant to be used to attach resources to
>> depthstencil and color render target outputs (similar to DX10 DepthStencil
>> and RenderTarget Views).
>> One thing though I'm not quite sure is if for array textures, the same field
>> should be used in the resource as for storing depth of 3d textures (the
>> depth0 field). Using the same "layer" field both for faces, array index and
>> 3d depth seems to make a lot of sense for transfers, the sampler views and
>> in pipe_surface, but I'm not sure the same field should be used in the
>> resource itself. I've actually converted some code to do that (but not quite
>> all for instance the mesa state tracker is broken there).
>> The reason for using the same field would be that it's mutually exclusive
>> anyway (unless someone introduces 3d array textures...), hence not really 2
>> fields are necessary. But OTOH depth behaves differently to arraysize (and
>> cube faces) because it doesn't get minified for smaller mips. So if someone
>> has convincing reasons what should be preferred I'll follow it :-).
>> One this is for sure though we need some way to store array size, and cube
>> maps will follow that pattern since for resource handling they are pretty
>> much the same as a 2d array texture with array size of 6.
> 
> This all seems fine. Drivers for pre-texture-array chipsets can handle
> all this by forcing off trilinear filtering, right?
This is not connected to trilinear filtering.
But yes, basically drivers not being able to do this should not need 
(further) changes, since the state tracker should not give them state 
they can't handle (like binding multiple layers to a rendertarget view) 
as usual.
I'll do some further changes though, pipe_surface is still not quite 
what we need.

Roland