* speeds up incremental builds as changes to a header will not always
need the full mocs_compilation.cpp for all the target's headers rebuild,
while having a moc file sourced into a source file only adds minor
extra costs, due to small own code and the used headers usually
already covered by the source file, being for the same class/struct
* seems to not slow down clean builds, due to empty mocs_compilation.cpp
resulting in those quickly processed, while the minor extra cost of the
sourced moc files does not outweigh that in summary.
Measured times actually improved by some percent points.
(ideally CMake would just skip empty mocs_compilation.cpp & its object
file one day)
* enables compiler to see all methods of a class in same compilation unit
to do some sanity checks
* potentially more inlining in general, due to more in the compilation unit
* allows to keep using more forward declarations in the header, as with the
moc code being sourced into the cpp file there definitions can be ensured
and often are already for the needs of the normal class methods
NVidia GPUs can't render to linear buffers, so we can't force a linear modifier
for CPU rendering. The proprietary NVidia driver also can't map gbm buffers
with all formats and modifiers, so instead of mapping the gbm buffer, this
patch changes KWin to instead use glReadPixels for accessing the source
buffer
The egl import mode ensures that there's a local buffer, which is
preferred as it minimizes the number of data transfers over PCIe.
With dmabuf, it's unclear what the driver will do. But the main takeaway
from discussion with mesa developers is that it's undesired for
gbm_bo_import() to migrate or perform data transfers behind the user's
back, it should be done explicitly.
This change ports the drm backend to the GraphicsBuffer and
GraphicsBufferAllocator.
The main motivation is to unify graphics buffer abstractions across
various backends and to prepare it for output layers, which could be
nicer if we could have direct control over the buffers.
This is done by converting from the sRGB + gamma 2.2 input from clients
to linear with the color space of the output (BT.709 or BT2020 atm) in
a shadow buffer, and then convert from the shadow buffer to the transfer
function the output needs (sRGB or PQ).
If gbm_bo_get_fd_for_plane() or gbm_bo_get_fd() fails, ensure that the
gbm buffer allocator properly handles it and doesn't return a
GraphicsBuffer object.
With the dmabuf multi-gpu path, a buffer is imported to the secondary GPU
and presented directly, but importing a buffer that's usable for scanout
is not possible that way on most hardware. To prevent CPU copy from being
needed in those cases, this commit introduces a fallback where the buffer
is imported for rendering only, and then copied to a local buffer that's
presented on the screen.
CCBUG: 452219
CCBUG: 465809
At the moment, the render backend provides its specific implementation
of LinuxDmaBufV1ClientBuffer. This has some of its limitations. For
example, due to the strong coupling, compositing restarts must be
handled carefully. It's hard to have a generic code path to import
dmabufs, which would be nice to have in order to unify graphics buffer
allocation across various backends; currently, it's all scattered.
To make the code simpler, this change drops the commented out YUV import
code path for now. Given that Mesa implicitly handles it, the need for
it is no longer so urgent.
It makes more sense for an output layer to return the render target fbo.
The user of the render target will then take appropriate steps to bind
the fbo. It reduces the amount of boilerplate code in output layers too.
This rewrite ensures that
- formats are tested for multi-gpu transfer
- dmabuf with either 10bpc or 8bpc buffers are preferred to cpu copy
- formats where drmModeAddFB2 fails are skipped
As the drm format number of 10 bit colors is lower, the fallback sorting
preferred those formats, even if KWIN_DRM_PREFER_COLOR_DEPTH=24 was set.
To fix that, prefer 8 bit colors over something random.
Some legacy drivers either don't accept gbm buffers suitable for cursors,
or don't handle them properly. In order to work around that, always do a
CPU import with legacy and use dumb buffers instead.
BUG: 453860
CCBUG: 456306
When we use a shadow buffer, we always render to the whole surface - setting
the damage region is incorrect and invokes undefined behavior. On the Lima
driver this caused flickering on screen rotation.
To fix this, don't set a damage region when we use a shadow buffer, which is
effectively setting the damage region to the full surface
There are a few benefits to using smart pointers from the standard library:
- std::unique_ptr has move semantics. With move semantics, transfer of ownership
can be properly expressed
- std::shared_ptr is more efficient than QSharedPointer
- more developers are used to them, making contributions for newcomers easier
We're also already using a mix of both; because Qt shared pointers provide
no benefits, porting to standard smart pointers improves consistency in
the code base. Because of that, this commit ports most of the uses of QSharedPointer
to std::shared_ptr, and some uses of QScopedPointer to std::unique_ptr
It appears that importing gbm_bo's using eglCreateImageKHR() doesn't
work on nvidia.
Another issue is that Platform::createDmaBufTexture() uses
gbm_bo_create(). The nvidia driver doesn't like that, it's preferred to
use gbm_bo_create_with_modifiers().
In order to address those issues, this change refactors how gbm_bo
objects are imported and how gbm_bo's are allocated for dmabuf textures,
so the same code can be reused in EglGbmBackend::textureForOutput()
and Platform::createDmaBufTexture().
It can happen that a gbm implementation does not support modifiers, while
the drm driver does. To prevent that from breaking KWin, fall back to creating
a gbm surface without modifiers when creating one with modifiers fails.
BUG: 453320
Xaver Hugl
Laurent Montel
Vlad Zahorodnii
Aleix Pol