I've implemented rotation and reflection support in the core XFree86
server; it was less invasive than I'd feared and seems to work just fine.
For most devices, it does it's work by substituting a whole new rendering
mechanism based on the shadow frame buffer when the screen is rotated or
reflected. Right now, this means that video and DRI won't work right if
the screen is rotated; they'll hit the regular frame buffer and mangle
things badly. It also permits drivers to support rotation themselves as
the siliconmotion hardware seems capable of.
It's possible to provide video support in this rotated world; I've
implemented that for the Mach64 chipset in Tiny-X (mostly because I could),
we could do that for the regular XFree86 servers too with some additional
work. On the other hand, DRI probably won't ever work on a rotated
screen; you'd have to transmit the rotation information to the application
and let them convert transformation matrices and clip lists. That'd be a
real adventure.
I haven't yet committed the changes to CVS as I'm curious of two things:
1) It hooks into some seemingly silly places -- it needs two
initialization steps, one right before acceleration is pushed
on top of fb and another right before the software cursor
is pushed on top of the acceleration. The first is done
by sticking a call inside xf86SetBackingStore, the
second by placing a call inside xf86GetPointerScreenFuncs.
Those two functions happen to need to be called at precisely the
same moment as the RandR hooks, but are obviously not exactly
semantically intended for this purpose.
The alternative is to require every driver to add calls to RandR,
which I'll do if people think it's the right think to do.
2) Permitting rotation incurs an additional layer along several
paths inside the X server -- the usual rendering paths all
need to get redirected when the system switches from hardware to
software rendering, so the screen functions CreateGC, PaintWindow*
and the like all get wrapped by the layer module.
This new layer is implemented by the 'layer' and 'shadow' modules,
both of which must be loaded when the server starts for rotation
to ever be possible.
It seems reasonable, therefore, that this dynamic rotation
mechanism should be selectable. Should I add an option to
enable or disable this? Should it be disabled by default?
Keith Packard XFree86 Core Team HP Cambridge Research Lab
_______________________________________________
Xpert mailing list
[EMAIL PROTECTED]
http://XFree86.Org/mailman/listinfo/xpert
