32-bit visuals and transparent windows

[Antoine Martin]

Needed for things like #77 (see there for backing patch)

We need to:

• add client caps (ie: "transparency"?)
• detect when the window uses 32-bit visuals
• make sure we use a 32-bit corral window (if client supports it and window uses 32-bit visuals)
• pass the window depth to clients that support it
• client windows should be created as 32-bit rgba windows if needed
• return 32-bit pixel data as needed from get_rgb_rawdata, and the depth used so callers know about it
• switch to encoders that support transparency (as some won't support it)
• store transparency with the encoding (png should handle it ok, we should add a rgb32 mode, x264?, vpx?)

Extra difficulty:

• pixbuf.get_from_drawable does not seem to handle things properly.. maybe we can use gdk_pixbuf_get_from_image instead?
• security: we don't want to have a remote transparent window capable of stealing all input

[Antoine Martin]

adds basics for supporting transparency

[Antoine Martin]

The biggest stumbling block is going to be the X11 server once again.

Neither Xvfb nor the dummy driver support 32-bit visuals, for whatever reason.
See ​this ticket for Xvfb.

I've tried to patch dummy to allow it but only got as far as this:

[128046.305] (II) DUMMY: Driver for Dummy chipsets: dummy
[128046.305] (WW) Falling back to old probe method for dummy
[128046.305] (WW) VGA arbiter: cannot open kernel arbiter, no multi-card support
[128046.305] (II) DUMMY(0): Chipset is a DUMMY
[128046.305] (II) DUMMY(0): Creating default Display subsection in Screen section
	"dummy_screen" for depth/fbbpp 32/32
[128046.305] (**) DUMMY(0): Depth 32, (--) framebuffer bpp 32
[128046.305] (EE) DUMMY(0): Weight given (000) is inconsistent with the depth (32)

Patch attached. Note: module loading is ​complicated and I hacked the ABI number rather than trying to figure out why the srpm did not install out of the box.

[Antoine Martin]

naive hack to try to get 32-bit modes with dummy driver (not working)

[Antoine Martin]

Ignore comment:4, no changes to the X11 server were needed.
What we want are 32-bit visuals, you can get that from a 24-bit dummy server just fine.

r3368 now handles transparency (big changeset - lost of details in commit message)

Remaining work items:

• verify pixel parsing on a big-endian server (and client?)
• verify transparency handling (X11 transparency is pre-multiplied?)
• test transparent system tray forwarding
• servers without PIL should fallback to gtk pixbuf code
• make it work with webp which is better (smaller/faster) than png and does transparency
• change encoding availability detection: can vary depending on client or server, client type, window backing type, etc..
• choose appropriate client window class based on alpha (no GL for now..)
• handle alpha channel with GL client window

[Antoine Martin]

Pretty much all done, see:

• #280 systray
• r3368 gtk2 pixmap backing code (+fixes in r3373 + r3457)
• r3514 (webp)
• r3371 (stub for GL: transparency not handled yet)
• r3772 + r3771 don't bother sending transparency if client cannot handle it

That will do for now, win32 may get transparency if/when we move to gtk3/qt4 (#90).

[Antoine Martin]

Note: X11 pixel data is in pre-multiplied format, so we needed r4152 to display it properly with GTK (#416 to improve on that). I don't think it really matters to the encoders (only png and webp at present) if it's pre-multiplied or not.

[onlyjob]

Since lack of transparency support should not be a problem any more why black rectangle still an issue with Firefox and expanding tree-style-tab?
Antoine, remember in ​https://www.xpra.org/trac/ticket/252#comment:24 you confirmed that you are able to reproduce the problem.
It never improved and as before the issue is only reproducible in active window... Any ideas why it still happening? Thanks.

[Norman Rasmussen]

Should xeyes's transparently work correctly with this? Is there an example that does?

[Norman Rasmussen]

xeyes uses the SHAPE extension, not transparency, so it's probably not expected to work :-)

[Antoine Martin]

(set correct milestone work was completed in)