[Mesa-dev] [PATCH 2/2] intel/isl/gen4: Make depth/stencil buffers Y-Tiled

[Nanley Chery]

On Tue, Jul 17, 2018 at 01:29:26PM -0700, Kenneth Graunke wrote:
> On Tuesday, July 17, 2018 10:45:28 AM PDT Nanley Chery wrote:
> > On Tue, Jul 17, 2018 at 08:19:30AM -0700, Kenneth Graunke wrote:
> > > Wow, I had no idea we were actually using linear depth buffers.
> > 
> > Neither did I until Mark let me know that I regressed some tests after
> > landing commit fbe01625f6bf2cef6742e1ff0d3d44a2afec003e. Whoops!
> > 
> > I've included the Bugzilla tag locally.
> > https://bugs.freedesktop.org/show_bug.cgi?id=107248
> > 
> > > Is there any compelling reason to use them?  I can't think of one.
> > > 
> > 
> > Yes. They give us the best performance and memory usage for 1D textures
> > (see isl_surf_choose_tiling()).
> 
> While 1D depth textures are certainly supported, I don't think I have
> ever seen one.  shader-db has no instances of them in any shader.
> 
> It doesn't make much sense...when you use a depth buffer, you draw
> objects in a scene and record their distance from the camera.  Doing
> that with a 1D surface just seems...unlikely?
> 

Good to know.

> > > Y-tiling should offer better performance.  PBO upload needs to be
> > > done with a R32_UINT format anyway, as R24_X8 isn't renderable.
> > > 
> > 
> > I don't understand how the PBO format comes into play here. Care to
> > elaborate?
> 
> PBO upload/download of texture data uses linear buffers.  I was thinking
> of glTexImage2D with data in a PBO.  Not sure if you can do that with
> depth...
> 

I'm still a little lost, but maybe we can talk about this point offline.

> > > I'm pretty sure that we used to always use Y-tiling in the old
> > > miptree code...maybe this regressed when we switched to ISL?
> > 
> > Yeah, we did always use Y-tiling. I caused the regression when I pushed
> > that commit.
> > 
> > > The hardware timeline looks like:
> > > 
> > >    BROKEN     /-------------- Works??? -----------\    DISALLOWED
> > > [Broadwater, Crestline, Eaglelake, Cantiga, Ironlake, Sandybridge, ...]
> > > 
> > > Given that it was broken and ultimately became disallowed, I'm a bit
> > > skeptical whether it really works, or is useful, in the meantime.
> > > 
> > 
> > It seems to work. We're getting testing from these piglit tests:
> > * piglit.spec.arb_shader_texture_lod.execution.tex-miplevel-selection *projgradarb 1dshadow
> > * piglit.spec.arb_shader_texture_lod.execution.tex-miplevel-selection *gradarb 1dshadow
> > * piglit.spec.arb_shader_texture_lod.execution.tex-miplevel-selection *lod 1dshadow
> > * piglit.spec.!opengl 1_1.copyteximage 1d
> > * piglit.spec.ext_texture_array.copyteximage 1d_array
> > * piglit.spec.arb_shader_texture_lod.execution.tex-miplevel-selection *projlod 1dshadow
> > 
> > On ILK and g45 they go from fail->pass with the last patch. But g965
> > needed the linear option removed completely.
> 
> I guess if it's working and useful for 1D, while being basically no
> code, we can do it, but...I have a real hard time caring about 1D depth.

I don't blame you. The main reason I care is that I just broke them :/