[Mesa-dev] [PATCH v3 2/5] i965: Faking the ETC2 compression on Gen < 8 GPUs using two miptrees.

[Eleni Maria Stea]

Hi Nanley,

On Thu, 7 Feb 2019 15:46:29 -0800
Nanley Chery <nanleychery at gmail.com> wrote:
 >  
> > @@ -3825,10 +3849,20 @@ intel_miptree_unmap(struct brw_context *brw,
> >     DBG("%s: mt %p (%s) level %d slice %d\n", __func__,
> >         mt, _mesa_get_format_name(mt->format), level, slice);
> >  
> > +   level_w = minify(mt->surf.phys_level0_sa.width,
> > +                    level - mt->first_level);
> > +   level_h = minify(mt->surf.phys_level0_sa.height,
> > +                    level - mt->first_level);
> > +
> >     if (map->unmap)
> >  	   map->unmap(brw, mt, map, level, slice);
> >  
> >     intel_miptree_release_map(mt, level, slice);
> > +
> > +   if (intel_miptree_has_etc_shadow(brw, mt) &&
> > mt->shadow_needs_update) {
> > +      intel_miptree_update_etc_shadow(brw, mt, level, slice,
> > level_w,
> > +                                      level_h);
> > +   }  
> 
> With the next patch applied, the change in this function becomes
> unnecessary. Is there any reason you're leaving it around?

After a second thought, I believe that this change wasn't unnecessary.
There is a problem if we remove it:

When we generate mipmaps we need to update the shadow for each level.
As the update is done per level during unmap, if we remove the call we
end-up with the first level correctly updated but all the others empty.

An example:
git clone https://github.com/hikiko/test-compression.git
make
./test compressed/full.tex

This test loads dumped compressed mipmap levels from the full.tex and
displays them, if you run it with the per level update inside the unmap
you will see all the mipmap levels. Without, you will see only the
first, like here: https://imgur.com/a/VvS0CYC

Do you have any suggestion on how I could bypass this problem?

Thanks again,
Eleni