[Intel-gfx] [PATCH] drm/i915: add support for Z-order of planes.

[Matt Roper]

On Fri, Feb 28, 2014 at 06:03:11PM +0200, Ville Syrjälä wrote:
> On Thu, Feb 27, 2014 at 03:44:04PM -0800, Matt Roper wrote:
> > On Thu, Feb 27, 2014 at 02:36:06PM -0800, Yu Dai wrote:
> > > On 14-02-25 04:19 PM, Matt Roper wrote:
> > > 
> > > >On Thu, Feb 20, 2014 at 04:11:14PM -0800, yu.dai at intel.com wrote:
> > > >>From: "Yu(Alex) Dai" <yu.dai at intel.com>
> > > >>
> > > >>Add "zorder" property to crtc to control Z-order of sprite and
> > > >>primary planes. The alpha channel of the planes can be enabled
> > > >>or disabled during Z-order change.
> > > >>
> > > >>Signed-off-by: Yu(Alex) Dai <yu.dai at intel.com>
> > > >It seems like this is pretty VLV-specific at the moment.  Aren't you
> > > >going to be writing bogus register values if you try to set this
> > > >property on a non-VLV platform?
> > > >
> > > >I'm not sure if any of the other non-VLV platforms supported by i915
> > > >today can change plane z-orders or not, but this is likely to be
> > > >possible and useful on future platforms (or non-Intel platforms) as
> > > >well, which may have a different number of overall planes and different
> > > >ways of programming them.
> > > 
> > > Yes, this is for VLV. New patch-set V3 was submitted.
> > > 
> > > >Would it make more sense to move the z-order value to a plane property
> > > >so that the interface scales to any number of planes?  If you just set
> > > >an integer value as a per-plane zorder-value, the code that actually
> > > >programs the hardware can go collect the values for each plane, sort
> > > >them to determine an order, and figure out what that translates to in
> > > >terms of platform-specific register programming.  This would lend itself
> > > >nicely to the "check/calculate" step of the atomic/nuclear operation
> > > >once those interfaces land.
> > > 
> > > The z-order of a plane really has no meaning until the actual composition is
> > > triggered. In Android, the order is determined by HWC when there are changes
> > > in layers. Breaking up this into plane property will evoke more drm IOCTL
> > > calls from user. The current "one-shot" call limits the flexibility but makes
> > > it simple.
> > 
> > That's true today where we only really have the ability to set
> > properties one by one.  But I believe the end goal is to handle this
> > kind of functionality via the atomic/nuclear API's, where you wind up
> > shoving all the various things you want to change/program into a
> > property set (which happens in libdrm userspace) and then issue a single
> > ioctl which hands that property set to the kernel to be processed in a
> > sort of transactional model.
> > 
> > Without the atomic/nuclear operations, I'm not sure how well z-order
> > property will really work in practice.  If your compositor decides that
> > the next frame of content uses multiple hardware planes to display
> > various buffers, you really need to make sure that the z-order register
> > programming happens within the same vblank that the flips on each of
> > those planes take effect, otherwise things aren't going to look good to
> > the end user.  With the non-atomic API's, I don't think there's really
> > any way to guarantee this.
> 
> Yeah w/o the atomic API probably the only semi useable approach is to
> drop out from the sprite path when stuff actually moves around, and
> get back on it after the scene is again steady.
> 
> As far as the z-order property, I did have a sort of similar idea
> originally where we'd define a bunch of different plane z-order
> combinations as an enum property on the crtc, and then the user has
> to choose one. That would make it easier for the user to figure out
> which way the planes can be stacked. This would be especially nice
> with some old Intel hardware where the z-order for the planes was
> specified using a couple of magic bits which severeily limited the
> ways you could stack the half a dozen or so planes.
> 
> But then you'd either need to parse the string enum name to make any
> sense of it, or we'd need to encode the order in the value itself. But
> to be hardware agnostic, we'd need to use plane IDs in there, and those
> can in theory be up to 31bits, so we couldn't guarantee that the property
> value could hold more than two planes. Although I suppose we could use
> the plane index instead, which would then put the upper limit at 16 planes.
> Well, possibly more in case not all planes can be assigned to the same
> crtc.
> 
> The simple approach is of course to have a range property per plane
> which defines the z-order. But then we need to worry about conflicts
> between z-order assignment, and it's harder for the user to figure out
> which combinations are legal. But the benefit is that this is the most
> obvious way to present this information. IIRC some people were
> suggesting we pick this scheme despite the limitations.
> 
> -- 
> Ville Syrjälä
> Intel OTC

Conflicts between z-order assignments don't seem too worrisome to me;
I figure most display servers would probably (re-)program z-order values
for all planes when they want to make any kind of plane ordering change
at all, so you generally wouldn't accidentally hit the case where, for
example, you set a plane to level "1" while another plane was already
sitting at level "1."

Figuring out which orderings are legal on the given hardware is somewhat
more of a problem.  It seems like the natural solution is to expect the
compositor or DDX run through a few plane ordering possibilities with
the atomic "test only" flag set when it starts up; it can then adjust
its runtime plane ordering logic according to what it determines is
possible.


Matt


-- 
Matt Roper
Graphics Software Engineer
ISG Platform Enabling & Development
Intel Corporation
(916) 356-2795