With QtQuick2 it's possible that the scene graph rendering context either
lives in an own thread or uses the main GUI thread. In the latter case
it's the same thread as our compositing OpenGL context lives in. This
means our basic assumption that between two rendering passes the context
stays current does not hold.
The code already ensured that before we start a rendering pass the
context is made current, but there are many more possible cases. If we
use OpenGL in areas not triggered by the rendering loop but in response
to other events the context needs to be made current. This includes the
loading and unloading of effects (some effects use OpenGL in the static
effect check, in the ctor and dtor), background loading of texture data,
lazy loading after first usage invoked by shortcut, etc. etc.
To properly handle these cases new methods are added to EffectsHandler
to make the compositing OpenGL context current. These calls delegate down
into the scene. On non-OpenGL scenes they are noop, but on OpenGL they go
into the backend and make the context current. In addition they ensure
that Qt doesn't think that it's QOpenGLContext is current by calling
doneCurrent() on the QOpenGLContext::currentContext(). This unfortunately
causes an additional call to makeCurrent with a null context, but there
is no other way to tell Qt - it doesn't notice when a different context
is made current with low level API calls. In the multi-threaded
architecture this doesn't matter as ::currentContext() returns null.
A short evaluation showed that a transition to QOpenGLContext doesn't
seem feasible. Qt only supports either GLX or EGL while KWin supports
both and when entering the transition phase for Wayland, it would become
extremely tricky if our native platform is X11, but we want a Wayland
EGL context. A future solution might be to have a "KWin-QPA plugin" which
uses either xcb or Wayland and hides everything from Qt.
The API documentation is extended to describe when the effects-framework
ensures that an OpenGL context is current. The effects are changed to
make the context current in cases where it's not guaranteed. This has
been done by looking for creation or deletion of GLTextures and Shaders.
If there are other OpenGL usages outside the rendering loop, ctor/dtor
this needs to be changed, too.
The PaintRedirector calls the new method KDecoration::render and passes
it's PaintDevice and the region to update to it. A decoration can
implement this method and provide an optimized implementation for the
painting which does not go through the deco's QWidget at all. In addition
the decoration can invoke an update() slot which will schedule a repaint
in the PaintRedirector and thus completely replaces the need for
intercepting paint events on the QWidget and also allows to add QWindow
based decorations in future.
also work around broken fbo texture clearing on fglrx
so far supports FBO/glClear and resorts to glTexSubImage2D
if the fbo cannot be created or is (in case of fglrx)
known to break, resort to glTexImage2D loading of an
argb array of zeros
BUG: 323065
FIXED-IN: 4.11.2
REVIEW: 112526
Qt 5 only supports raster which means our pixmaps are always non native,
so we don't need the Extension information any more and can drop all
special code handling for mapping a native QPixmap to an X11 pixmap.
This fixes artifacts from over-sampling along the edges of the
decorations when effects such a wobbly windows are active.
The textures are only cleared when framebuffer objects are supported.
All drivers support FBO's now, so in practice this shouldn't be
a problem.
Use two textures per window instead of four, storing the left and
right borders in the first texture, and the top and bottom borders
in the second.
This makes it possible to render the whole decoration with only two
calls to glDrawArrays(). It also reduces the number of texture
allocations while resizing a window.
Add virtual resizePixmaps() and updatePixmaps() methods that resizes and
updates all the pixmaps. Make resize() and paint() regular virtuals
with stub implementations.
The new methods will be used by OpenGLPaintRedirector in the next commit.
The scratch PaintDevice is provided by the PaintRedirector's subclasses
allowing the ones wanting a QImage to provide the QImage directly and the
ones needing a QPixmap to provide a pixmap. This means the copy from
QPixmap to QImage is no longer needed in the OpenGL and XRender/raster
paint redirector.
REVIEW: 109138
Ownership of decoration textures is moved from SceneOpenGL::Window to
OpenGLPaintRedirector. The PaintRedirector is responsible for updating
the textures whenever they change. For this GLTexture is extended by an
update(QImage, QPoint) method which uses glTexSubImage2D to update only
the changed parts.
The big advantage compared to before is that if e.g. only a button is
animated only the button part is updated instead of the complete deco
part.
PaintRedirector is turned into an abstract class providing a factory
method which returns either an instance of
* OpenGLPaintRedirector
* NativeXRenderPaintRedirector
* RasterXRenderPaintRedirector
OpenGLPaintRedirector is basically doing exactly the same as the parent
class used to do before. Though the idea is to extend the functionality
to have the PaintRedirector write directly into OpenGL textures to limit
copying the complete decorations.
NativeXRenderPaintRedirector is similar to OpenGLPaintRedirector by
rendering into a QPixmap and providing the pictureHandle for the QPixmap
to SceneXRender.
RasterXRenderPaintRedirector is providing the functionality for the case
that the QPixmap/XPixmap relationship is not present. From the QPixmap
containing the pending decoration paint a QImage is created and then the
relevent parts are copied directly into the decoration pixmap.
REVIEW: 109074
The only task of the PaintRedirector is to redirect the painting of the
window decorations into Pixmaps. So it should actually do this by also
handling the four pixmaps for the decoration. This simplifies the code
as all the logic concerning redirecting the painting is now grouped
together.
Furthermore the PaintRedirector is now a child of the decoration widget,
which means it gets automatically destroyed whenever the decoration is
destroyed - the Client does not have to care about it.
Also the PaintRedirector gets only created if the Compositor is active as
it is not needed in the non-compositing case.
REVIEW: 106620
The Compositor class actually behaves like a Singleton so it should be
one. Therefore four static methods are added:
* self() to access the Singleton
* createCompositor() to be used by Workspace to create the instance
* isCreated() to have a simple check whether the Singleton is already
created
* compositing() as a shortcut to test whether the compositor has been
created and is active
The isCreated() check is actually required as especially Clients might
be created and trying to access the Compositor before it is setup.
The refactoring of Compositor starting with b1739c3 caused some
regressions due to variables in Workspace and Compositor not
being initialized. Furthermore there was a boolean logic error
in PaintRedirector causing the decorations not to paint.
BUG: 305875
The Scene has always been created and destroyed inside what is
now the split out compositor. Which means it is actually owned
by the Compositor. The static pointer has never been needed
inside KWin core. Access to the Scene is not required for the
Window Manager. The only real usage is in the EffectsHandlerImpl
and in utils.h to provide a convenient way to figure out whether
compositing is currently active (scene != NULL).
The EffectsHandlerImpl gets also created by the Compositor after
the Scene is created and gets deleted just before the Scene gets
deleted. This allows to inject the Scene into the EffectsHandlerImpl
to resolve the static access in this class.
The convenient way to access the compositing() in utils.h had
to go. To provide the same feature the Compositor provides a
hasScene() access which has the same behavior as the old method.
In order to keep the code changes small in Workspace and Toplevel
a new method compositing() is defined which properly resolves
the state. A disadvantage is that this can no longer be inlined
and consists of several method calls and pointer checks.
Instead of scheduling and gathering the Workspace::addRepaint calls at the end of the main loop
we directly emit the signal which ends up in a call to addRepaint. The compositeTimer assures
that the higher number of scheduled repaints are executed in the same rendering pass.
Without Compositing we do not really need the PaintRedirector,
it only adds overhead. This reverts the behavior to how it was before
the PaintRedirector was introduced by just not redirecting the paint
in the event filter when compositing is not active. This should have
the lowest impact on the existing source base.
The change breaks at least oxygen window decoration when compositing
is not active as the decoration is no longer double buffered. But
as written this is exactly the behavior as used before 4.3. Given
the push today there should be enough time for all affected decos
to adjust to the change.
REVIEW: 101413
in case we need it again.
This should improve performance with drivers where creating pixmaps is
expensive.
svn path=/trunk/KDE/kdebase/workspace/; revision=1104069
