[Nouveau] [Mesa-dev] RFC: Fixing nv30 fbo attachments

[Ilia Mirkin]

On Tue, May 20, 2014 at 4:51 PM, Marek Olšák <maraeo at gmail.com> wrote:
> On Tue, May 20, 2014 at 9:58 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
>> Hello,
>>
>> I attempted doing this a while back, before I really understood what
>> was going on. I got some advice that went totally over my head, and I
>> dropped the issue. I think I'm much better-prepared to tackle the
>> issue this time around.
>>
>> To recap, nv30 (and nv40) hw can't handle certain color/depth format
>> combinations. Specifically, a 16-bit color format must be paired with
>> a 16-bit depth format, and a 32-bit color format must be paired with a
>> 32-bit depth format (well, Z24S8). This HW also can't handle different
>> per-attachment sizes, and also the "linearity" of all the attachments
>> must be the same. (Whether a surface is linear or not is _generally_
>> dictated by its w/h, but not always -- POT textures can sometimes end
>> up allocated linearly by e.g. the ddx, or something else.) The
>> different sizes are handled by not exposing ARB_fbo. However the rest
>> of the cases remain.
>>
>> Now that I kinda understand how things are structured, I was thinking
>> of doing the following:
>>
>> When rendering (i.e. draw_vbo & co) and the fbo has changed (and has
>> one of the problems above), I would instead put in a temporary texture
>> as the depth attachment. Before actually drawing, I would blit from
>> the real target texture into the temporary texture, and then when
>> rendering is done, blit back from the temp texture back into the
>> target. This deals with the target texture getting modified between
>> draws with various blits/mapping/whatever.
>>
>> This means that you'll only get 16 bits of depth even if you ask for
>> 24 with a 16-bit color format, but the alternative seems too
>> complex/costly.
>>
>> So there are a few questions from this approach:
>>
>> 1. Where do I get the temporary texture from? (And more importantly --
>> when... what happens if allocation fails?)
>>
>> 2. Having to blit the depth texture back and forth on every draw seems
>> _really_ wasteful... anything I can do about that?
>
> You can do that in set_framebuffer_state. When binding, blit to a
> depth buffer which matches the colorbuffer format. When unbinding,
> blit back.

set_framebuffer_state doesn't allow an error to be returned. Should I
just print a warning and move on?

I guess I'm still not 100% on all the terminology -- what do you mean
exactly by bind/unbind? Do you mean the transfer_map/unmap stuff? So
basically I would blit once on set_framebuffer_state, and then blit
back and forth on resource map/unmap, and only ever render to the
"temporary" buffer without worrying about blitting wihle rendering?

>
> You can also drop support for 16-bit formats.

I assumed that these were required by some GL version... I also
presume that it's faster to use these. BTW, when I say 16-bit, I mean
like B5G6R5 or B5G5R5X1, not R16*.

>
> What about rendering to R16F, RG16F, and RGBA16F? Does R16F have to be
> coupled with a 16-bit depth buffer too?

Gooood question. I assume they want a 32-bit depth buffer. Although I
believe those are actually disabled for now (aka forever, unless
someone cares enough to turn them on), even though it does look like
there's HW support for R16G16B16A16_FLOAT and R32 versions. Or perhaps
they're disabled because they can't be used in conjunction with depth?
Not familiar enough with the hw restrictions, sorry.

  -ilia