Re: [Intel-gfx] [PATCH] drm/i915: Tile F plane format support

["Lisovskiy, Stanislav" <stanislav.lisovskiy@intel.com>]

On Mon, Sep 27, 2021 at 10:24:11PM -0700, Matt Roper wrote:
> On Mon, Sep 27, 2021 at 09:29:07PM +0300, Ville Syrjälä wrote:
> > On Mon, Sep 27, 2021 at 11:23:35AM -0700, Matt Roper wrote:
> > > On Thu, Sep 23, 2021 at 06:49:59PM +0300, Ville Syrjälä wrote:
> > > > On Thu, Sep 23, 2021 at 11:48:58AM +0300, Stanislav Lisovskiy wrote:
> > > > > TileF(Tile4 in bspec) format is 4K tile organized into
> > > > > 64B subtiles with same basic shape as for legacy TileY
> > > > > which will be supported by Display13.
> > > > 
> > > > Why we still haven't done the F->tile64 rename?
> > > >
> > > > This is the last chance to fix this before we bake 
> > > > this into the uapi and are stuck with a name that doesn't
> > > > match the spec and will just confuse everyone.
> > > 
> > > I think you're confusing the formats here.  The bspec uses both terms
> > > "TileF" and "Tile4" for the same format in different places.  There's a
> > > completely different format that's referred to as both "TileS" and
> > > "Tile64" in the bspec that we don't use at the moment.  So tile64
> > > wouldn't be a correct rename, but tile4 could be.
> > 
> > Right, tile64 is the macro tile variant I think. So like Ys
> > which we never bothered implementing, so I guess we''l not bother
> > with tile64 either.
> > 
> > > 
> > > In general Tile4 is much more common in the bspec than TileF is (TileF
> > > terminology is mostly found in the media sections).  And bspec 44917 is
> > > the most authoritative bspec page on the subject, and it refers to it as
> > > Tile4, so I agree that switching over "Tile4" would probably be a good
> > > move.
> > > 
> > > > 
> > > > > 
> > > ...
> > > > > diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
> > > > > index bde5860b3686..d7dc421c6134 100644
> > > > > --- a/include/uapi/drm/i915_drm.h
> > > > > +++ b/include/uapi/drm/i915_drm.h
> > > > > @@ -1522,7 +1522,8 @@ struct drm_i915_gem_caching {
> > > > >  #define I915_TILING_NONE	0
> > > > >  #define I915_TILING_X		1
> > > > >  #define I915_TILING_Y		2
> > > > > -#define I915_TILING_LAST	I915_TILING_Y
> > > > > +#define I915_TILING_F		3
> > > > > +#define I915_TILING_LAST	I915_TILING_F
> > > > 
> > > > fences...
> > > 
> > > Recognizing TileF/Tile4 separately from TileY is important to code
> > > outside of display as well.  There are blitter instructions that require
> > > different settings for TileY vs Tile4/F so if we drop the tracking of
> > > this as a unique tiling type, it will break the blitting/copying and
> > > some of the upcoming local memory support for Xe_HP-based platforms.
> > 
> > These are uapi definitions for set_tiling(). You are not meant to add
> > anything there. Just like we didn't add anything for Yf.
> 
> Yeah, I think that's the real problem --- we define some values here in
> the uapi header, but we also wind up using the same set of values for
> driver-internal non-uapi purposes too rather than having a separate enum
> (containing a superset of the uapi values) that can be used for those
> other things.  Display code can use FB modifiers for some things, but
> core/lmem code needs a way to refer to Tile4 and such and doesn't have a
> good way to do that today.
> 
> I think most (all?) of the non-display code that's relying on a
> definition of I915_TILING_F is in various selftests that are still being
> prepared for upstreaming, so maybe there's a better way to handle the
> selection of possible formats specifically in the selftest code itself.
> That's really the only area of the kernel code that should need to be
> aware of the specific internal layout of various buffers.

So I will proceed with the renaming at least.

Ville, suppose, I still need part of fencing related code?

Stan

> 
> 
> Matt
> 
> > 
> > -- 
> > Ville Syrjälä
> > Intel
> 
> -- 
> Matt Roper
> Graphics Software Engineer
> VTT-OSGC Platform Enablement
> Intel Corporation
> (916) 356-2795