[Mesa-dev] i965: Use isl for hiz and stencil

[Pohjolainen, Topi]

On Tue, May 09, 2017 at 08:48:08AM -0700, Jason Ekstrand wrote:
> All of the below being said, I'm fine with landing things with a
> BACK_TO_BACK layout and then making the switch to one isl_surf per miplevel
> later if that's an easier path.

I can give the "array of ISL surfaces" approach a spin to see how it looks,
I'm curious to see if few things become a little cleaner.

> 
> On Tue, May 9, 2017 at 8:45 AM, Jason Ekstrand <jason at jlekstrand.net> wrote:
> 
> > On Mon, May 8, 2017 at 11:10 PM, Pohjolainen, Topi <
> > topi.pohjolainen at gmail.com> wrote:
> >
> >> On Mon, May 08, 2017 at 04:51:35PM -0700, Jason Ekstrand wrote:
> >> > On Wed, May 3, 2017 at 2:22 AM, Topi Pohjolainen <
> >> topi.pohjolainen at gmail.com
> >> > > wrote:
> >> >
> >> > > Patches 1-17 are revision that
> >> > >
> >> > >   - rework hiz on gen6 to use on-demand offset calculator allowing
> >> > >     one to drop dependency to miptree structure and
> >> > >   - rework all auxiliary surfaces to be created against isl directly.
> >> > >
> >> > > Patches 18 and 19 introduce new surface layout in ISL. This is called
> >> > > back-to-back and similar to layout ALL_SLICES_AT_EACH_LOD found in
> >> > > i965 for gen6 hiz and stencil. This layout stacks slices for each
> >> level
> >> > > after one and other, or back to back. All slices ate each lod is
> >> almost
> >> > > the same except that it places levels one and two side-by-side trying
> >> > > to preserve space. Back-to-back wastes a little more memory but aligns
> >> > > each level on page boundary simplifying driver logic.
> >> > >
> >> >
> >> > My primary gripe here is that you seem to have half-added back-to-back
> >> to
> >> > ISL.  If this layout is a long-term thing, then we should add a new
> >> > ISL_DIM_LAYOUT_GEN6_BACK_TO_BACK layout and plumb your offset function
> >> > through isl_surf_get_image_offset_sa.  Is this intended to be a
> >> permanent
> >> > solution?  I think eventually, I'd like us to go with one surf per
> >> miplevel
> >> > (which is almost the same) but I can see how this is easier at the
> >> moment.
> >> > However, I think this works sufficiently well that I'm ok with doing the
> >> > back-to-back thing for a while.
> >>
> >> I thought about adding new layout type but couldn't decide which way is
> >> better. It is easy to buy your arguments in favor, and I'm happy to give
> >> it
> >> a go.
> >> If miptree per level is your number one choice, then lets go with that.
> >
> >
> > I said "one surf per miplevel".  I see no reason why we need N miptrees.
> >
> >
> >> I just
> >> need to check a few things first about the actual solution. I would see
> >> something in these lines:
> >>
> >> 1) Add a dynamically allocated array of miptrees into miptree. This would
> >>    contain miptree instance per level.
> >>
> >> 2) Still uses one buffer object containing space for all levels. The
> >> instances
> >>    in the array would either have their ::bo pointer zero or pointing to
> >> the
> >>    parent ::bo. In both cases ::offset would point the start of the level.
> >>
> >
> > Yes
> >
> >
> >> 3) Instances in the array are not reference counted and therefore deleted
> >>    simply by deallocating the malloced chunk underneath.
> >>
> >
> > If we have one isl_surf per miplevel and not a miptree per level, then I
> > don't think this is an issue.
> >
> >
> >> 4) Add similar dynamically allocated array of intel_miptree_aux_buffer
> >>    instances for hiz. Here also use one ::bo which would need to added to
> >>    miptree I think cause there ins't one in miptree. Or perhaps add the
> >>    array of aux buffers to aux buffer?
> >>
> >
> > Looking at intel_miptree_aux_buffer, I think what we would end up with is
> > an array of aux_buffers
> >
> >
> >> 5) ISL doesn't need to know about this and hence we would add the total
> >> space
> >>    calculator along with ::offset initialization in i965 (brw_tex_layout,
> >>    I think).
> >>
> >
> > That's fine.  We already do that in Vulkan with anv_surface.  ::offset
> > calculation can be done easily enough by just adding sizes.
> >
> >
> >> 6) In Vulkan <-> GL interop, we'd pass single level arrays only as ISL
> >> didn't
> >>    know about back-2-back. Or we simply don't care about gen6 as Vulkan
> >>    doesn't support it anyhow?
> >>
> >
> > Yeah, we don't care about gen6.
> >
> > --Jason
> >