Normally all shader programs from materials are cached in a simple table
that is alive as long as the renderer exists, in practice this means being
tied to the QQuickWindow which is often the entire lifetime of an
application.
Use releaseResources(), which is mostly a no-op on the scenegraph level
at the moment, to free this cache.
Task-number: QTBUG-62392
Change-Id: I279e595874fee1ab4fe0dede0843a3686dea5806
Reviewed-by: Gunnar Sletta <gunnar@crimson.no>
Since I362e1cb8e10 the batch renderer defaults to
QSG_SEPARATE_INDEX_BUFFER which broke the batches visualization
Change-Id: If1d51cabb0cc4a3a98ac2c01bd78789d08fe72f7
Reviewed-by: Gunnar Sletta <gunnar@crimson.no>
Reviewed-by: Robin Burchell <robin.burchell@crimson.no>
QSGNode::DirtyForceUpdate is 0x8000, so when you shift it left by 16
positions, it becomes the sign bit. Either the result of that shift
is int or sn->dirtyState is.
qsgbatchrenderer.cpp(302): error #68: integer conversion resulted in a change of sign
Change-Id: I84e363d735b443cb9beefffd14b958f9a622348b
Reviewed-by: Tony Sarajärvi <tony.sarajarvi@qt.io>
0a88774a0f switched from using client side
pointers to VBOs, ostensibly for core profile support. Unfortunately,
some GPUs/drivers seem to suffer quite a bit with these, so only use
them if we must.
Results for changing_over_isolated_with_clip_rotated.qml:
eskil_linux_tx1:
Before: Average: 65.6 frames; using samples; MedianAll=65; StdDev=0.894427, CoV=0.0136346
After: Average: 120 frames; using samples; MedianAll=120; StdDev=0, CoV=0
eskil_linux_focault (nvidia) has a similar regression in all branches
where the original commit is present, too:
5.6 & v5.8.0 tag: 600
5.8, 5.9 & dev branches: 399
Assuming this is the sole regression in that case and we end up at the
original 600 again, that would be a 50% gain there, and an 81% increase
on TX1.
Change-Id: I44af9e67698356200f9587e77c9409fdb756519d
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
Reviewed-by: Gunnar Sletta <gunnar@crimson.no>
This only deals with Qt code. MASM has a lot of those left. We should
just update from upstream instead to get the fixes.
qv4regalloc.cpp:660:52: warning: this statement may fall through [-Wimplicit-fallthrough=]
if (leftSource->type == DoubleType || rightSource->type == DoubleType) {
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
qv4regalloc.cpp:666:13: note: here
case OpBitAnd:
^~~~
Change-Id: I7814054a102a407d876ffffd14b6b0e2d6b03689
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
Unlike renderUnmergedBatches(), renderRenderNode() did not adjust the
projection matrix.
Change-Id: Ib5a7183a3623d35c85af47205cc22187bad89409
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
Otherwise weird things will happen once a node gets turned into a
batch root and the rendernodes in the child subtree still refers to
their old root.
Change-Id: I42b0ba514c2fbeed833f0f665e49b275c19b5686
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
Client-side pointers are not supported. Start using a VBO instead.
Task-number: QTBUG-57768
Change-Id: Ia7ac9b0838d837b02e8bf99fcd22f0373cb357c9
Reviewed-by: Gunnar Sletta <gunnar@sletta.org>
The renderers added in 5.8 had to be adapted to the changed profiling
macros from 5.6.
Conflicts:
src/plugins/scenegraph/d3d12/qsgd3d12renderloop.cpp
src/plugins/scenegraph/d3d12/qsgd3d12threadedrenderloop.cpp
src/quick/scenegraph/adaptations/software/qsgsoftwarerenderloop.cpp
src/quick/scenegraph/adaptations/software/qsgsoftwarethreadedrenderloop.cpp
src/quick/util/qquickprofiler_p.h
tests/auto/qml/qjsengine/tst_qjsengine.cpp
tests/auto/qml/qqmlvaluetypes/tst_qqmlvaluetypes.cpp
Change-Id: Icb370b7c95aab12589ad73881ac6d178759a5c6b
The profiler can be switched on in the middle of a frame. In that case
the last offset into the timing data would be some random number, which
may lead to a crash when recording the sample.
However, as we know all the data points we are going to record, we can
as well specify where they are supposed to go. The timings themselves
may still be random for frames of which we only recorded parts, but
the clients can deal with this.
Task-number: QTBUG-57304
Change-Id: I1d507f2591516e43d5b3cd25f7939716f2b64ed9
Reviewed-by: Simon Hausmann <simon.hausmann@qt.io>
Since eea8fa64ab, we're seeing increased
times in QSGNode::removeChildNode(). The reason for this seems to be
that iteration through the linked list is significantly slower than
iteration through a QList<> due to that each node needs to be loaded
in memory to iterate to the next, compared a more plain sequential
pointer compare with QList<>.
This implementation changes the nodes to use a circular double-linked
list so we can drop the iteration when removing nodes. This brings us
slightly better performance than the original QList based
implementation while still using the same amount of memory as the
single-linked list one.
Change-Id: I0b693730f5b82ad7767e507accafd90551e03fbc
Reviewed-by: Erik Verbruggen <erik.verbruggen@qt.io>
Having rendernodes triggering fullscreen updates with the software backend
is not ideal. Therefore, introduce the option of reporting that the rendernode
is well-behaving, meaning it only writes inside the reported bounding rectangle.
Similarly, the OpenGL batch renderer can keep using the depth buffer when the
rendernode is known to behave like the renderer expects.
Change-Id: I6acc434432c2504776f26e8917b5434b44be293d
Reviewed-by: Gunnar Sletta <gunnar@sletta.org>
Have to change getResource() a bit since it turns out it is not suitable
currently for backends that do not have a per-window rendercontext and do
not implement the interface on the rendercontext. Pass in the window to
make sure it can always figure out which window we want the resources for.
(we do not want rendererInterface() to return window-specific instances
created on the fly, with ownership issues, so stick with the simple model
where backends implement the interface on one of their existing classes)
To support clipping, QSGRenderNode::RenderState is extended accordingly.
Also updated the docs since some claims in the rendernode docs are not true
since Qt 5.3.
Change-Id: I34779c83926f5231b888fcab7131e873ae97964f
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
This also shrinks the class by a tiny bit, because the second bool fits
inside the padding that was added to the first bool.
Change-Id: Ifb3f3f165c5e1666e2787118de5b5fb579a573b9
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
If one overriding virtual method is marked as such, all of them should
be marked.
Change-Id: I29a19d2196a7f1afbb31fba8a4a6f2b345489e3c
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
And start using them from the batch renderer as well.
Change-Id: I20adc3962b06ac712267dc2be95e9e28cea82e21
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
This reverts commit 26a230ee0e.
This optimization breaks when we have a single material change
in the scene within a batch as the geometry node does not get
visited on the next render.
Task-number: QTBUG-52983
Change-Id: Ib385407a9fc35ca03ab18727d1e7b550431416f1
Reviewed-by: Michael Brasser <michael.brasser@live.com>
Reviewed-by: Jocelyn Turcotte (Woboq GmbH) <jturcotte@woboq.com>
For non-OpenGL APIs the primary (and likely the only) way to add custom
rendering into the Qt Quick scene is via the render node. Other
approaches,like the before/afterRendering signals,
QQuickFramebufferObject, remain OpenGL-only. (although QQuickFramebufferObject
may get a multi-API replacement based on QSGRenderNode at a later time)
Note that this is not a generic 3D content integration enabler. It targets
creating 2D and 2.5D Quick items with custom rendering via the graphics
API in use.
Make QSGRenderNode public, enhance the docs a bit and add a releaseResources().
Add a QSGRendererInterface with a query function in QQuickWindow and
QSGEngine. The scenegraph adaptation can then return a custom implementation of
the interface. This will be necessary to query API-specific values, f.ex. the
ID3D12Device and ID3D12CommandList when running with the d3d12 backend.
The interface allows querying the API and void* resources. Resources that
we know about in advance are enum-based to prevent the
QPlatformNativeInterface-like ugliness of string keys.
Support is there in the batch renderer already, fix this up according to
the new public API, and implement the corresponding bits for the D3D12
renderer.
For D3D12, fix also an issue with QSGNode destruction where graphics resources
in use were attempted to be final-released without a proper wait.
The semantics of changedStates() in QSGRenderNode is changed so that it can be
called at any time, including before render(). This is very useful since we can
implement some state restoring in a more efficient manner.
Added a new example as well. Documentation for QSGRenderNode is heavily
expanded.
Change-Id: I4c4a261c55791d0e38743a784bc4c05a53b3462d
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
This change also fixes the build of two benchmarks, tst_affectors
and tst_emission.
Conflicts:
src/plugins/qmltooling/qmldbg_native/qmldbg_native.pro
src/qml/qml/ftw/qhashfield_p.h
tests/benchmarks/particles/affectors/tst_affectors.cpp
tests/benchmarks/particles/emission/tst_emission.cpp
tests/benchmarks/qml/pointers/pointers.pro
tests/benchmarks/qml/pointers/tst_pointers.cpp
tests/benchmarks/qml/qmltime/qmltime.pro
tests/benchmarks/qml/qquickwindow/qquickwindow.pro
Change-Id: I595309d1e183c18371cb9b07af6e4681059de3b2
Currently the Qt Quick module depends on either the OpenGL or OpenGLES
headers being available at build time. Since we are adding support for
adaptations that do not depend on OpenGL, it should be possible to build
Qt Quick in environments that do not have OpenGL development headers.
This does present many challenges though because in some cases GL types,
and classes that require OpenGL are part of the public APIs. However
since these classes were never available when QT_NO_OPENGL was defined,
it should be possible to redefine the function signatures under this
scenario, since it's not possible to break binary compatibility if there
never were any binaries to break compatibility with.
One of the bigger changes that was necessary to facilitate this change
is creating interfaces out of QSGContext and QSGRenderContext. Here the
default behavior was usage of OpenGL directly, even though subclasses
could override all OpenGL usage. Making them interfaces should bring
QSGContext and QSGRenderContext more in line with the other classes
present in the adaptation layer.
Change-Id: Iaa54dc0f6cfd18d2da1d059548abf509bd71f200
Reviewed-by: Laszlo Agocs <laszlo.agocs@theqtcompany.com>
From Qt 5.7 -> LGPL v2.1 isn't an option anymore, see
http://blog.qt.io/blog/2016/01/13/new-agreement-with-the-kde-free-qt-foundation/
Updated license headers to use new LGPL header instead of LGPL21 one
(in those files which will be under LGPL v3)
Change-Id: Ic36f1a0a1436fe6ac6eeca8c2375a79857e9cb12
Reviewed-by: Lars Knoll <lars.knoll@theqtcompany.com>
Fixes: "** CID 156593: Control flow issues (DEADCODE)".
Child can not be null do to while (child) { ...
Change-Id: I46a4f23f24bf7e6b71e72b177249587eea57d5ce
Reviewed-by: Joerg Bornemann <joerg.bornemann@theqtcompany.com>
Reviewed-by: Simon Hausmann <simon.hausmann@theqtcompany.com>
Reviewed-by: Robin Burchell <robin.burchell@viroteck.net>
This reduces transient allocations by 10mb on the Rectangle creation benchmark,
at the cost of an additional two pointers per shadow node (and just that pointer --
no additional allocation ala QList is required, so this will actually end up
around the same size if not better.)
Change-Id: I6c25c93fe31ed892e3c42ed66d6c6656e5d7b38a
Reviewed-by: Gunnar Sletta <gunnar@sletta.org>
Added a check that Batch::drawSets is not empty.
Task-number: QTBUG-48439
Change-Id: Ica76363be8c770240dc69c669815a60904e26988
Reviewed-by: Gunnar Sletta <gunnar@sletta.org>
Where possible, use qEnvironmentVariableIsSet()/
qEnvironmentVariableIsEmpty() instead of checking on the
return value of qgetenv().
Where the value is required, add a check using one of
qEnvironmentVariableIsSet()/Empty().
Move QSGAtlasTexture::qsg_envInt() to qsgrenderer.cpp
for reuse as qt_sg_envInt() and add qt_sg_envFloat().
Change-Id: I4c93f16c228d4f537154f389a0fa1427654485f7
Reviewed-by: Gunnar Sletta <gunnar@sletta.org>
In the uncommon case (m_context->hasBrokenIndexBufferObjects()
|| m_visualizeMode != VisualizeNothing) of mapping a buffer, malloc is
called without first freeing the previous malloc.
Regression was introduced with:
qt5 commit: 9347499e78f03710eaf24af3c1e7ac650d0ef81d
qtdeclarative commit: a371bac3fb
[ChangeLog][QtQuick][Scene Graph] Fixed memory leak in
QSGBatchRenderer::Renderer::map()
Task-number: QTBUG-48799
Change-Id: I5ef4b7301d390463845aeb192851f86655962499
Reviewed-by: Gunnar Sletta <gunnar@sletta.org>
Just not requesting a depth buffer is insufficient since there's nothing
guaranteeing that the EGL or other implementation will return depthless
configs. When the depth buffer is always there, setting the environment
variable is futile. To do what the user has requested, the value has to
be checked in the renderer too.
Change-Id: I1f572bc6f2f5b5aa94070a239d6e871e3421a51f
Reviewed-by: Louai Al-Khanji <louai.al-khanji@theqtcompany.com>
Reviewed-by: Gunnar Sletta <gunnar@sletta.org>
Conflicts:
src/qml/debugger/qv4debugservice.cpp
src/qml/jsruntime/qv4value_inl_p.h
src/qml/jsruntime/qv4value_p.h
src/qml/memory/qv4mm.cpp
src/qml/memory/qv4mm_p.h
src/qml/qml/qqmlnotifier_p.h
src/qml/qml/qqmlproperty.cpp
src/quick/items/qquickflickable.cpp
src/quick/items/qquicktextedit.cpp
tests/auto/quick/qquickwindow/BLACKLIST
The extra changes in qqmlbinding.cpp are ported from changes to
qqmlproperty.cpp that occurred in parallel with writeBinding() being
moved to qqmlbinding.cpp.
Change-Id: I16d1920abf448c29a01822256f52153651a56356
Animating a complete batch of geometries bound by property
bindings will cause an unnecessary rebuild of the batch on
each animation step even though it will end up with the
same batch as in the previous frame.
Since the invalidation happens in nodeChanged, any node
change in an updatePaintNode might trigger an invalitation
if it is compared with a material that hasn't been updated yet.
Delay the verification of the DirtyMaterial flag until all
materials have been updated, later on in visitNode, to
make sure that we call compare only on up-to-date materials.
Change-Id: I03c095efc20817813508d959c74b41eae57beedc
Reviewed-by: Gunnar Sletta <gunnar@sletta.org>
No reason to iterate through all items to create a temporary
QList we don't need.
Change-Id: I3c5388f026ded0290afb9dd35fc7934011c4998d
Reviewed-by: Marc Mutz <marc.mutz@kdab.com>
Reviewed-by: Olivier Goffart (Woboq GmbH) <ogoffart@woboq.com>
Reviewed-by: Gunnar Sletta <gunnar@sletta.org>
Found with GCC's -Wcast-qual.
Change-Id: Ia0aac2f09e9245339951ffff13c9589afabc7ade
Reviewed-by: Alan Alpert (Personal) <416365416c@gmail.com>
Reviewed-by: Simon Hausmann <simon.hausmann@theqtcompany.com>
Adding a QSGRenderNode to the scene permanently disabled opaque batches
by disabling depth buffer usage. Reset the depth buffer usage to the
default value once last QSGRenderNode has been removed from the scene.
Change-Id: I760afde83ae9eaaf1b5571c37fd0081eb23b1f20
Reviewed-by: Gunnar Sletta <gunnar@sletta.org>
Laszlo Agocs
Gunnar Sletta
Berthold Krevert
Thiago Macieira
Robin Burchell
Ulf Hermann
Liang Qi
Gunnar Sletta
Erik Verbruggen
Laszlo Agocs
Andy Nichols
Frank Meerkoetter
Erik Verbruggen
Liang Qi
Giuseppe D'Angelo
Andy Nichols
J-P Nurmi
Jani Heikkinen
Robin Burchell
Jason Erb
Friedemann Kleint
Martin Banky
Ulf Hermann
Gunnar Sletta
Jocelyn Turcotte
Sérgio Martins
Simon Hausmann
Mikko Harju
Frederik Gladhorn