Software-based clipping is naive. It maps the clip region from the
global screen coordinates and intersects it with window quads. If the
window is transformed, the quads won't be clipped as expected.
Unfortunately, the OpenGL scene enables software-based geometry clipping
if the screen transformed. It's not clear why it does so, perhaps as a
performance optimization? Either way, this produces incorrect results
when the screen is scaled.
BUG: 440940
This makes shadow texture coordinates consistent with surface and
decoration texture coords and allows reusing window quads between
kwin items and qtquick items more easily.
We use surfaceless contexts with internal windows. We also require
the EGL_KHR_surfaceless_context extension for making context current
without outputs.
Arguably, we could use pbuffers, but since mainstream drivers (Mesa and
NVIDIA) support surfaceless contexts, the extra complexity doesn't buy
us anything.
SceneOpenGLShadow::prepareBackend used to use QImage::Format_Indexed8,
plus a special code path in GLTexture, to create single-channel OpenGL textures.
Now that Qt supports QImage::Format_Alpha8, this workaround can be removed.
Currently, a vertex coordinate is transformed first, then mapped to
the global screen coordinates. This causes a problem if a transform
is applied to the top-most item and child items are not at (0, 0).
For example, scaled windows may have popping out sub-surfaces, etc.
With this change, the item transforms will be computed differently. For
example, if the parent item is transformed, a child's transform will
look as follows
[Parent's translation][Parent's transform][Child's translation]
instead of
[Parent's translation][Child's translation][Parent's transform]
In the future, I'd like to get rid of the Item::setTransform() call in
OpenGLWindow::performPaint() and have either AnimationEffect or
libkwineffects call Item::setTransform().
BUG: 440201
Blending is quite expensive especially with software rendering.
In the case of Firefox on Wayland, it uses a ARGB8888 buffer but marks the
entire surface as opaque, so the alpha channel can be ignored.
CCBUG: 440386
A better alternative is to return the damage region in the beginFrame()
function. This way, the render backend can force full screen repaints or
indicate what parts of the buffer needs to be repainted (for buffer age)
If window quads need to be generated after the wl_surface is destroyed,
the SurfaceItemWayland::mapToBuffer() function will return wrong values.
In order to fix that, we need to store the last surface-to-buffer matrix
in SurfaceItem.
At the moment, we handle window quads inefficiently. Window quads from
all items are merged into a single list just to be broken up again.
This change removes window quads from libkwineffects. This allows us to
handle window quads efficiently. Furthermore, we could optimize methods
such as WindowVertex::left() and so on. KWin spends reasonable amount
of time in those methods when many windows have to be composited.
It's a necessary prerequisite for making wl_surface painting code role
agnostic.
Window quads need to be in some coordinate space. Since we want items to
be used not only for rendering windows, window-local coordinates do not
suffice.
This change makes scene items generate quads in the item-local coordinate
space.
The model matrix is used to map the quads to the global screen coordinate
system from the item coordinate space.
Since the quads are in the item local coordinate space, the mvp matrix
needs to be updated on every draw call. Given the render data, tracking
the last mvp matrix won't result in less glUniform calls. If this indeed
becomes a serious performance bottleneck, we can explore the possibility
of dumping mvp matrices in a UBO, which have been introduced in OpenGL
3.1.
The scene items depend on the scene windows for caching window quads.
The goal of this change is to move window quads management to item.
Merging window quads in one list and then splitting them is inefficient,
it will be highly desirable if window quads are removed from the public
api so we can optimize window quad management.
With this change, the window quad type becomes irrelevant to render
backends for the most part. Note that the Xrender backend is a bit
nitpicky about window quads, so the shadow item doesn't create generic
"WindowQuadShadow" quads anymore.
Drop-shadows with the software render backend impact performance quite
significantly. It also makes it easier to prepare render backends for the
item based design.
So far, we were creating a model view with the complete scene rendered
(even if we didn't render the windows themselves). This required us to
have a big glPerspective spanning the entire scene and we were just
cropping it as we rendered it into a smaller texture.
This changes our scenes so we have the correct matrix set up at all
times.
Specifically in the case of the Pinephone, this solves the following
issue where we were unable to connect external displays because it
exceeded GL_MAX_VIEWPORT_DIMS:
https://invent.kde.org/teams/plasma-mobile/issues/-/issues/11
The Xrender backend was added at the time when OpenGL drivers were not
particularly stable. Nowadays though, it's a totally different situation.
The OpenGL render backend has been the default one for many years. It's
quite stable, and it allows implementing many advanced features that
other render backends don't.
Many features are not tested with it during the development cycle; the
only time when it is noticed is when changes in other parts of kwin break
the build in the xrender backend. Effectively, the xrender backend is
unmaintained nowadays.
Given that the xrender backend is effectively unmaintained and our focus
being shifted towards wayland, this change drops the xrender backend in
favor of the opengl backend.
Besides being de-facto unmaintained, another issue is that QtQuick does
not support and most likely will never support the Xrender API. This
poses a problem as we want thumbnail items to be natively integrated in
the qtquick scene graph.
We used to get a weird line around the window decoration because the
clamping would hit the outside of the rendered decoration.
To make sure we fall inside take the displayPixelRatio into account.
Also on non-integer scales, make sure we are actually falling inside.
It was needed to work around visual glitches in the wobbly windows
effect. Since the wobbly windows effect renders the animated window into
an offscreen texture, we don't need this workaround anymore.
Furthermore, rather than using half-pixel correction, it is more
desirable to use an offscreen texture as it results in simpler design.
Performance-wise, it's not that bad that we need to start looking for
other ways to get rid of the seams between window contents and deco.
Currently, the implementation of the DecoratedClient and the decoration
renderer are strongly coupled. This poses a problem with the item based
design as the ultimate goal is to have scene items construct paint nodes
which are then fed to the renderer. The DecorationItem has to have
control over the decoration texture. Another issue is that the scene
cannot smoothly cross-fade between two window states if the decoration
is removed, e.g. from fullscreen mode to normal and vice versa.
This change moves the decoration renderer to the decoration item. With
the introduction of a generic scene texture atlas, we hope to get rid of
the decoration renderer altogether.
One of the scene redesign goals is to make wayland surface items
re-usable. So we have the same rendering path for drag-and-drop icons,
software cursors, and window surfaces.
The biggest issue at the moment is that window pixmaps are tightly
coupled with scene windows.
This change de-couples window pixmaps from scene windows. In order to
achieve that, some architecture changes were made.
The WindowPixmap class was replaced with the SurfacePixmap class. A
surface pixmap is created by a surface item.
Under the hood, a SurfacePixmap will create a PlatformSurfaceTexture
object, which contains all the information necessary for the renderer.
The SceneOpenGLTexture class was removed. However, the GLX and the EGL
on X11 backends still mess with GLTexture's internals.
