[Mesa-dev] separate depth and stencil buffers in gallium

[Alex Deucher]

On Wed, Sep 22, 2010 at 6:15 AM, Dave Airlie <airlied at gmail.com> wrote:
> On Wed, Sep 22, 2010 at 7:50 PM, Keith Whitwell
> <keith.whitwell at gmail.com> wrote:
>> On Wed, Sep 22, 2010 at 10:30 AM, Dave Airlie <airlied at gmail.com> wrote:
>>> So Evergreen hardware appears to have only completely separate depth
>>> and stencil buffers and doesn't natively support a combnined DS buffer
>>> from what I can see. I'm awaiting clarification from AMD.
>>>
>>> Now gallium and st/mesa seem to be quite dedicated to the whole
>>> combined DS cause.
>>
>> What formats exactly does it support?
>
> It looks like Z16,24 and a 32-float, along with I think an S8 stencil
> (its a one bit format).
>
> I'm not sure if the Z24 is packed or its actually X8Z24, this only
> matters when I need to know
> what size buffers to allocate in the DDX for passing over DRI2.
>>
>> It's interesting because DX10 (and presumably 11) always talk about
>> combined buffers - but the abstraction is such (with staging
>> resources, no direct mapping of the buffers, etc) that there's nothing
>> which constrains the layout to be an interleaved depth+stencil.   IE.
>> you could quite happily allocate the combined depth/stencil as a
>> planar depth buffer and a separate planar stencil buffer - both hidden
>> behind the same resource handle.
>
> Yeah thats what I'm wondering if I should just hide this all in the
> driver resource handling.
>
> The hw is a bit messy anyways, we can't attach DB direct to texture,
> we have to flush it out
> via a DB->CB rendering pass, so I think its probably possible with
> sufficent shader to make it
> produce a combined ZS from the separated buffers. I'
>
>
>> I would have hoped we'd have the same flexibility in gallium -
>> basically that nobody should be able to tell whether depth & stencil
>> are swizzled together or separate.
>>
>> The obvious case where the app & state tracker might be alerted to
>> your unusual layout is in transfers.  An interim solution would be to
>> swizzle/unswizzle depth buffer transfer data (or organize for the card
>> to do so for you).
>>
>>
>>> I'm mainly posting just wondering if anyone else has considered this
>>> or any other hardware this might be useful for exists, or if anyone
>>> can speak to the pitfalls I'll face.
>>>
>>> I've got some initial done in 30 mins hacks
>>> http://cgit.freedesktop.org/~airlied/mesa/log/?h=sep-zs
>
> So I'm guessing I've taken the wrong approach here from reading this,
> I should probably not expose this to gallium, and just make sure the DDX
> allocates a large enough buffer for two planes.
>
> Hopefully Alex can dig out some more info.

The z and stencil buffers are stored independently.  Z24 buffers are
not packed, they are stored using 32 bits.

Alex

>
> Thanks,
> Dave.
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>