Xorg ou la pile graphique dominante des OS libres

 
Canonical announced their new display manager yesterday. There’s a section ”Why Not Wayland / Weston?” where they claim:
 
“we consider the shell integration parts of the protocol as privileged and we’d rather avoid having any sort of shell behavior defined in the client facing protocol.”
 
and something similar was written here also:
 
” Wayland .. exposes privileged sections like the shell integration that we planned to handle differently, both for security reasons and as we wanted to decouple the way the shell works on top of the display server from the application-facing protocol”
 
so they would rather have:
 
“An outer-shell together with a frontend-firewall that allow us to port our display server to arbitrary graphics stacks and bind it to multiple protocols.”
 
First of all, there’s nothing privileged about the shell protocol Wayland is exposing. wl_shell and wl_shell_surface (the “shell protocols”) are part of the Wayland core protocol, yes, but as I’ve explained on this post, it’s all customizable for whatever UI needs. Nevertheless, their usage is completely optional and anyone can build a different shell and stack with the rest of Wayland, just like tablet-shell protocol for instance does. Still, this will be Wayland and use the shiny libwayland for IPC.
 
Therefore I don’t think Canonical should justify their new project because Wayland “does not fulfill .. requirements completely”. There are no technical reasons Ubuntu cannot use Wayland in principle. What they wrote there is a very very mean excuse instead.

cairo-xlib tessellates the high-level paths from the user into trapezoids and sends those to the Xserver. The ddx then rasterises the trapezoids into a mask and composites that onto the destination. Both Nvidia and glamor use trapezoid shaders to avoid rasterising with the CPU, SNA uses the same high speed scanline rasteriser as cairo-image (both try to eliminate the intermediate mask), and EXA uses the slow pixman trapezoid rasterisation routines and the extra compositing step. (For -intel the CPU is faster at generating the RLE opacity mask and sending it as geometry to the GPU than the current GPUs are at executing the branch heavy trapezoid shader. The ultimate question is whether we can tolerate using MSAA and have GPUs sufficiently fast enough...)
 
cairo-image rasterises directly from the general complex polygon computed for the path (convert the curves into straight lines, convolve with a pen etc). This essentially folds the two passes peformed by cairo-xlib into one and eliminates the very computationally expensive Bentley-Ottmann routine for tessellating trapezoids. On the downside, cairo-image only uses a single core (and no GPU offload) for its rasterisation. Also, more work can be done for cairo-image to process the path without requiring an intermediate polygonisation (e.g. walk splines within the scanline rasteriser, use a hairline renderer for thin pens, compute offset curves, etc).
 
The next step to speed up cairo-xlib would be to eliminate the trapezoids and send paths directly to X - fix the protocol to be more useful for cairo, and also coincidentally would enable separate render threads within cairo. For Nvidia, they would then couple up their driver to use their existing NV_path acceleration, and I would do something similar for SNA (as usual, look at the early experiments in cairo-drm) if the GPU was not the bottleneck.