* = not done / = work in progress ? = should it be done? See the HACKING file for KWin details. The composite-related sources are mostly composite.cpp and scene_opengl.cpp (TODO: add comments and general overview). Sources: - glcompmgr : http://lists.freedesktop.org/archives/xorg/2006-July/017006.html , http://www.mail-archive.com/compiz%40lists.freedesktop.org/msg00023.html - simple and easy to understand - works even without texture_from_pixmap extension - claims to support several different gfx cards - compiz : git clone git://anongit.freedesktop.org/git/xorg/app/compiz - the ultimate - glxcompmgr : git clone git://anongit.freedesktop.org/git/xorg/app/glxcompgr - a rather old version of compiz, but also simpler and as such simpler to understand - libcm (metacity) : cvs -d :pserver:anonymous@anoncvs.gnome.org:/cvs/gnome co libcm - no idea about it - http://opengl.org - documentation - OpenGL Redbook, GLX docs, extensions docs TODO: * Check/make it work with other gfx cards - I've tested only with nvidia with the 9625 beta drivers and 8776 stable drivers so far - I have absolutely no idea about other gfx cards, needs to be tested * ati * intel * who knows what else ? Xgl support - Compiz itself doesn't work when compiled with the libGL from nvidia, it ships its own and links against it ? might be worth trying to use that libGL as well - it may be just because of the special libGL, but when testing with Xgl it even seemed non-conformant - none of the provided configs had GLX_RENDER_TYPE with GLX_RGBA_BIT even though required by GLX and other funny things. Indeed, it may be just me being still pretty clueless about these things. * is there a good reason to support Xgl? With the 9625 nvidia drivers it seems to work fine without them ? AIGLX support - no idea about this at all * find out if it works * XComposite overlay window - X.Org 7.1 and newer have XComposite extension version 0.3, which allows creating an overlay window to paint to instead of painting to the root window - I still run older X.Org here - the code in glcompmgr for this is in HAVE_OVERLAYS, stealing primarily from here should be better since it's simpler and the code so far has been modelled more after glcompmgr - the code in compiz for this is in USE_COW * use double-buffering with the overlay window - kwin already has code for double-buffered destination drawable, but I haven't been able to test it * when relying on extensions, detect that they are available - some of the glcompmgr code probably depends on OpenGL/GLX extensions that may not be always available, probably the same with compiz * find out which code relies on extensions and add checks ? alpha clear hack * find out if it affects performance * if yes, try to find a better way of solving the problem - since kompmgr has an option to make only the decoration transparent, it should be possible to do the same here - if the window has alpha and a decoration or if there should be only the decoration transparent, paint first the contents and then the decoration - this should make it possible to paint the decoration without the random alpha that is right now handled by the alpha hack / design framework for graphical effects - modelling it after compiz seems to make a lot of sense * update only those parts of the screen that have been changed - for paintSimpleScreen() - it currently sets region to the whole screen * instead of using glXSwapBuffers() there should be glCopyPixels() used - compiz has such code in evenLoop() in the if() block for COMP_SCREEN_DAMAGE_REGION_MASK * perhaps syncing to vblank will be needed to avoid flicker * XCopyArea() should not update the whole screen but only the affected areas * sync to vblank - currently the compositing code is run with 20ms timer, i.e. constant 50fps - the GLX_SGI_video_sync extension should be used - compiz uses this, no idea about it though ? GL_ARB_texture_rectangle vs GL_ARB_texture_non_power_of_two - code currently uses GL_ARB_texture_rectangle (GL_TEXTURE_RECTANGLE_ARB), using normal textures when GL_ARB_texture_non_power_of_two is available should(?) be preferred ? in SceneOpenGL::bindTexture() with tfp_mode, with some gfx cards it seems to be faster to not short-circuit the texture binding when there's been no damage * confirm and try to find out when to do it and when not