Re: [PATCH 06/19] drm/blend: Add a generic alpha property

[Laurent Pinchart]

Hi Daniel,

On Thursday, 11 January 2018 18:36:10 EET Daniel Vetter wrote:
> On Thu, Jan 11, 2018 at 4:58 PM, Maxime Ripard wrote:
> > On Tue, Jan 09, 2018 at 03:28:34PM +0100, Daniel Vetter wrote:
> >> On Tue, Jan 09, 2018 at 02:53:22PM +0100, Maxime Ripard wrote:
> >>> On Tue, Jan 09, 2018 at 01:32:41PM +0100, Daniel Vetter wrote:
> >>>> On Tue, Jan 09, 2018 at 11:56:25AM +0100, Maxime Ripard wrote:
> >>>>> Some drivers duplicate the logic to create a property to store a
> >>>>> per-plane alpha.
> >>>>> 
> >>>>> Let's create a helper in order to move that to the core.
> >>>>> 
> >>>>> Cc: Boris Brezillon <boris.brezillon@xxxxxxxxxxxxxxxxxx>
> >>>>> Cc: Laurent Pinchart <laurent.pinchart@xxxxxxxxxxxxxxxx>
> >>>>> Signed-off-by: Maxime Ripard <maxime.ripard@xxxxxxxxxxxxxxxxxx>
> >>>> 
> >>>> Do we have userspace for this?
> >>> 
> >>> Wayland seems to be on its way to implement this, with ChromeOS using
> >>> it:
> >>> https://lists.freedesktop.org/archives/wayland-devel/2017-August/034741
> >>> .html
> >>> 
> >>> and more specifically:
> >>> https://chromium.googlesource.com/chromium/src/+/master/third_party/way
> >>> land-protocols/unstable/alpha-compositing/alpha-compositing-unstable-v1
> >>> .xml#118
> >> 
> >> Yay, would be good to include these links in the patch description.
> >> Really happy we're having a real standard now used by multiple people.
> > 
> > I will.
> > 
> >> > > Is encoding a fixed 0-255 range really the best idea?
> >> > 
> >> > I don't really know, is there hardware or formats where there is more
> >> > than 255? Or did you mean less than that?
> >> 
> >> 30bit I'd assume wants more alpha. In the past we've done some
> >> fixed-point stuff (e.g. for LUT), using the 0.0-1.0 float range. Using
> >> that for the blend equation docs is also what I recommend (and that we
> >> map from 0-255 to 0.0-1.0 logically). Ofc the hw might not do any of that
> >> ... I think 0.16 fixed point, stored in a u16 is probably best. That's
> >> what we're doing for gamma tables already, and that way drivers can
> >> simply throw away the lower bits.
> > 
> > But that would also break the two users of that property that won't be
> > able to move to the generic property (with the same name) without
> > breaking userspace. The point of that patch was to allow some code
> > consolidation, and that would mean failing to do so here :/
> 
> Let me try to clarify:
> - We'll keep the exact existing property semantics with the 0-255
> range for the userspace visible property.
> 
> - But internally, in the decode value that we store into
> drm_plane_state, we'll do the slightly more future proof thing with a
> few more bits.
> 
> That gives us the option of exposing those bits in the future without
> having to change all the drivers again (which we have to do for this
> series here already anyway, since the decoded value moves into
> drm_plane_state from driver subclasses).
> 
> Definitely not asking to break userspace here :-)

As the property range is available to userspace, we could allow drivers to 
expose a driver-dependent number of bits for the alpha value without breaking 
anything. We can even require new drivers to expose 16 bits regardless of the 
number of bits that the hardware can handle, or we could keep that driver-
specific.

Please note, however, that U0.16 can only represent [0.0, 0.999984741...] 
while we need [0.0, 1.0]. As all the hardware I know map the full range of 
values ([0, 255] for 8-bit for instance) to [0.0, 1.0] I wouldn't pretend that 
the 16-bit value exposed to userspace is U0.16.

> >>>> I know other drivers have skimped on the rules here a bit ... But at
> >>>> least internally (i.e. within the drm_plane_state) we probably should
> >>>> restrict ourselves to u8. And this needs real docs (i.e. the full blend
> >>>> equation drivers are supposed to implement).
> >>> 
> >>> You mean straight vs premultiplied? Maybe we should implement this as
> >>> an additional property in read only depending on how the hardware
> >>> behaves?
> >> 
> >> No need for an additional property right now, but definitely document
> >> whether you mean straight or pre-multiplied. Just writing down the blend
> >> equation is probably best.

-- 
Regards,

Laurent Pinchart