Just to make sure instanced drawing does not regress. Relevant
particularly with Metal.
Fixes: QTBUG-114885
Change-Id: Ib39066d32985bf25ca02d5aa54d9cf654772be9a
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
Some shortcomings and unexpected problems are not unlikely.
The basic feature, with 2 views going to a texture array's
0 and 1 elements, seems to be working with macOS, also with
MSAA. Instanced drawing has not been verified. (relevant
because layered rendering works via instancing in Metal
and the QRhi backend has to adjust the instance count
in every draw call)
Fixes: QTBUG-114774
Change-Id: I3655e0d2c658b88c4cd6b52a32f94134324e4ac9
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
This should not be there because 'u' is passed to beginPass().
Change-Id: I95ba8ed400baa06948b4d4c6bbf7ca2d07a5480f
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
This relies on qsb being able to invoke dxc instead of fxc when the
request HLSL (shader model) version is 6.1. (6.1 is required for
SV_ViewID) This currently works only when conditioning offline with
qsb (or via CMake), because qsb can easily invoke dxc instead of
fxc. When shipping HLSL inside the .qsb package (so when -c is not
specified or running the external tool fails), this won't work since
the D3D12 backend still uses D3DCompile(), not IDxcCompiler. Support
for that will be investigated separately.
We also need to bump to ID3D12Device2 and ID3D12GraphicsCommandList1.
With Windows 10 version 1703 being quite old now, this should not be a
problem at run time.
There are however issues at build time, namely that MinGW and
MinGW/LLVM and similar seems to have ancient Windows SDK headers
according to the CI test runs. None of the MSVC configurations have
this in the CI, they have reasonable versions of d3d12.h and similar.
Therefore, one important aspect of this change is that the D3D12
backend of QRhi will only be available from now on when the SDK
headers are new enough (meaning ID3D12Device2 is declared, which is a
several years old type now). Otherwise, QRhi::create() will simply
fail when asking for D3D12 with a helpful warning message about the Qt
build being crippled.
Implementation-wise, there are surprises in store as well:
The way the PSO is created needs to be revamped to follow the
extensible approach that uses a pipeline state stream
description. Both the graphics and compute pipeline creation is
changed to use CreatePipelineState() and the associated
machinery. This is only really essential for graphics pipelines since
we know have to include data for view instancing (multiview). For
compute the result is the same as before.
Additionally, the view count must now be baked into the
QRhiGraphicsPipeline. This means that applications must call
setMultiViewCount() with the same value (typically 2) *both* on the
render target's color attachment and on the pipeline. Backends that do
not care about the pipeline's view count (GL, Vulkan) will of course
ignore it, but if it's not set correctly D3D12 will fail. The manual
test is updated accordingly.
Fixes: QTBUG-114772
Change-Id: I93db7313377e711c2faeb956815899b12132d23b
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
qrhi.h, qshader.h, qshaderdescription.h (and qshaderbaker.h from
shadertools; done separately) become "RHI APIs", following the concept
of QPA APIs.
Mirror completely what is done for QPA headers, but using the "rhi"
prefix for the headers. This involves updating syncqt to handle the
new category of headers. (a note on the regex: matching everything
starting with "qrhi" is not acceptable due to incorrectly matching
existing and future headers, hence specifying the four header names
explicitly)
There is going to be one difference to QPA: the documentation for
everything RHI is going to be public and part of the regular docs, not
hidden with \internal.
In addition to the header renaming and adding the comments and
documentation notes and warnings, there is one significant change
here: there is no longer a need to do API-specific includes, such as
qrhid3d11[_p].h, qrhivulkan[_p].h, etc. These are simply merged into a
single header that is then included from qrhi.h. This means that users
within Qt, and any future applications can just do #include
<rhi/qrhi.h> (or rhi/qshader.h if the QRhi stuff is not relevant), no
other headers are needed.
There are no changes to functionality in this patch. Only the
documentation is expanded, quite a lot, to eliminate all qdoc warnings
and make the generated API docs complete. An example, with a quite
extensive doc page is added as well.
Task-number: QTBUG-113331
Change-Id: I91c749826348f14320cb335b1c83e9d1ea2b1d8b
Reviewed-by: Volker Hilsheimer <volker.hilsheimer@qt.io>
Reviewed-by: Qt CI Bot <qt_ci_bot@qt-project.org>
...so we have core profile compatible GLSL code
in them. Just so one can run e.g. triquadcube with
-g -c (OpenGL with a core profile context).
Pick-to: 6.5
Change-Id: I585d3b4f0c7cd71ce7fae1fff4bf9a84cb7410da
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
Modeled after Metal's cb.GPUStart/EndTime. Implemented with timestamp
queries for other APIs.
Implemented for Metal, D3D11, Vulkan for now. No more callback, just
a getter on the command buffer which returns the latest known value,
referring to some previous frame. This makes it a lot more usable
than the original solution that is not really used anywhere at
the moment.
Now works for offscreen "frames" as well, this was not implemented
before.
Opt in with a new QRhi::create() flag because we cannot tell in
advance if the getter will be called or not, and this way we can
skip recording the timestamps by default. The cost is probably
minimal, though. Qt Quick will set this automatically when running
with QSG_RHI_PROFILE=1.
Change-Id: I903779984a4e0bbf1d03806d04bf61571ce23d72
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
Noticed the warnings when building the manual tests.
Pick-to: 6.5
Change-Id: I7f927f42f11d234ec3c980f36d8e12c0c49be712
Reviewed-by: Joerg Bornemann <joerg.bornemann@qt.io>
Having a simple Dear ImGui bridge is not just useful for the manual
tests, which do not have any other means to displays GUIs, but is
in itself an important exercise for the QRhi machinery.
Have a new manual test that exercises the built-in ImGui demo window.
Then use it in the displacement test for real, to replace the myriads
of key presses with on-screen sliders and checkboxes (with less code).
Change-Id: I296bafae2a5cce6fc7a447d97e68e5bcec15f451
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
This test was created with a work around for a Metal tessellation
pipeline memory alignment issue. The workaround was to specify shader
stage in / out variable uv as vec3 rather than vec2. A recent patch to
correct Metal tessellation pipeline memory alignment has now allowed
this test to use vec2 for variable uv as originally intended.
Change-Id: I6772c0e824e1e4b7e749dafa218f3fd8eba0e541
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
If it's executable, it should specify how it's to be executed.
Change-Id: If5671712da3e1fbc42b15d22c1253129910091bc
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
This does not really belong here as a built-in feature, esp.
considering that such testing is relevant for other backends
as well.
Change-Id: Ifbe3b8c6a430aacb9fcbdabf0e3761b14c48decc
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
There is something odd when running on Metal: note how the uv
is vec3 instead of vec2, in order to make the vertex-tesc-tese
data to look like this:
struct main0_out
{
float3 out_uv;
float3 out_normal;
float4 gl_Position;
};
if out_uv was float2 we'd get some strange rendering results,
perhaps due to something related to alignment. But have no means
to investigate this further.
Change-Id: I79d4edb2ddde3971c599c4326d98e99a49aa7122
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
- The optional nice-to-haves DebugMarkers, Timestamps, PipelineCache
are not yet implemented (features reported as false, to be
implemented later, although buffer/texture resource name setting
already works as-is, regardless of DebugMarkers).
- Mipmap generation for 3D textures is missing. Won't matter much
given that 3D textures are not used in Qt for anything atm. For
generating mipmaps for 2D (or 2D array) textures, the MiniEngine
compute shader and approach is used. 3D support for the mipmap
generator may be added later. 1D textures / arrays are supported
except for mipmap generation, and so the
OneDimensionalTextureMipmaps feature is reported as false.
- Qt Quick and Qt Quick 3D are expected to be fully functional.
(unforeseen issues are not impossible, of course)
- Uses minimum feature level 11.0 when requesting the device. It is
expected to be functional on resource binding tier 1 hardware even,
although this has not been verified in practice.
- 2 frames in flight with the usual resource buffering
(QRhiBuffer::Dynamic is host visible (UPLOAD) and always mapped and
slotted, other buffers and textures are device local (DEFAULT).
Requests 3 swapchain buffers. Swapchains are mostly like with D3D11
(e.g. FLIP_DISCARD and SCALING_NONE).
- The root signature generation is somewhat limited by the SPIR-V
binding model and that we need to map every binding point using the
nativeResourceBindingMap from the QShader. Thus the root signature
is laid out so each stage has its own set of resources, with shader
register clashes being prevented by setting the visibility to a
given stage.
Sampler handling is somewhat suboptimal but we are tied by the
binding model and existing API design. It is in a fairly special
situation due to the 2048 limit on a shader visible sampler heap, as
opposed to 1000000 for SRVs and UAVS, so the approach we use for
textures (just stage the CPU SRVs on the (per-frame slot) shader
visible heap as they are encountered, effectively treating the heap
as a ring buffer) would quickly lead to having to switch heaps many
times with scenes with many draw calls and sampledTexture/sampler
bindings in the srb.
Whereas static samplers, which would be beautiful, are impossible to
utilize safely since we do not have that concept (i.e. samplers
specified upfront, tied to the graphics/compute pipeline) in the
QRhi API, and an srb used at pipeline creation may change its
associated resources, such as the QRhiSampler reference, by the time
the shader resources are set for the draw call (or another,
compatible srb may get used altogether), so specifying the samplers
at root signature creation time is impossible.
Rather, the current approach is to treat each sampler as a separate
root parameter (per stage) having a descriptor table with a single
entry. The shader visible sampler heap has exactly one instance of
each unique sampler encountered during the lifetime of the QRhi.
- Shader-wise no different from D3D11, works with HLSL/DXBC 5.0
(i.e. existing .qsb files with DXBC in them work as-is). But unlike
D3D11, this one will try to pick 6.7, 6.6, ..., down to 5.0 from the
QShader, in that order.
- Uses D3D12MA for suballocating. As a result it can report vmem
allocation statistics like the Vulkan backend, and it does more
since the DXGI memory usage (incl. implicit resources) is also
reported. This is optional technically, so we also have the option
of going straight with the heavyweight CreateCommittedResource()
instead. That is what we do if the adapter chosen reports it's
software-based or when QT_D3D_NO_SUBALLOC=1 is set.
- PreferSoftwareRenderer (picking the WARP device) and the env.var.
QT_D3D_ADAPTER_INDEX work as with the D3D11 backend.
- It is not unexpected that with large scenes that generate lots of
draw calls with multiple textures/samplers per call the performance
may be slightly below D3D11 (probably mostly due to descriptor
management). Similarly, the reported memory usage will be higher,
which is partly natural due to creating heaps, descriptor pools,
staging areas, etc. upfront. Will need to be evaluated later how
these can be tuned.
Change-Id: I5a42580bb65f391ebceaf81adc6ae673cceacb74
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
Reviewed-by: Qt CI Bot <qt_ci_bot@qt-project.org>
This way we can run the test on macOS with the version 2.1
OpenGL context.
Amends 85a1663eb1
Change-Id: I8ec122fefaab54b35613e226e3937f4b51a7ea5a
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
Support for 1D textures on Vulkan, OpenGL, Metal, and D3D.
Change-Id: Ie74ec103da9cfcbf83fa78588cf8cfc1bd6e104f
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
Setting the flag QSurfaceFormat::StereoBuffers does not actually do
anything, because we do not utilize the extra buffers provided. We need
to expose setting the correct buffers using glDrawBuffers between draw
calls.
Change-Id: I6a5110405e621030ac3a2886fa83df0cfe928723
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
Interesting on its own just because it exercises stencil testing,
unlike any of the other existing manual tests.
In addition it serves as a base example for how outlines could be
done, it is one possible approach at least. (render with stencil
write, then render again slightly scaled up with a solid color with
testing against the stencil buffer content)
Change-Id: I0c845a9004136f229cab037f6f0aab2f772bdd76
Reviewed-by: Christian Strømme <christian.stromme@qt.io>
And include qcore_mac_p.h where needed.
Task-number: QTBUG-99313
Change-Id: Idb1b005f1b5938e8cf329ae06ffaf0d249874db2
Reviewed-by: Tor Arne Vestbø <tor.arne.vestbo@qt.io>
The caveat being having to manually create HLSL versions of the hull,
domain, and geometry shaders in parallel with the Vulkan GLSL ones,
while keeping the interfaces intact (stage inputs and outputs, cbuffer
layouts, binding points/registers). This is not always trivial but
typically doable in not very complicated case after inspecting the
SPIRV-Cross-generated vertex/fragment code in the .qsb files. Once
written, the HLSL files can be injected into a .qsb file with qsb -r.
or the corresponding CMake syntax. Conceptually this is no different
from how samplerExternalOES support is implemented for Multimedia.
(there the problem is that the shaders cannot be compiled to SPIR-V
to begin with, here it is that we cannot translate from SPIR-V, but
in the end the workaround for both problems is effectively the same)
The manual tests demonstrate this, both the tessellation and geometry
apps work now with D3D out of the box.
On the bright side, the implementation here in the the D3D backend of
QRhi does not need to know about how the shaders got there in the
QShader. So none of the implementation is dependent on this manual
process. If some day qsb would start translating to these kind of
shaders as well, it would all still work as-is.
Change-Id: I32d9ab94e00174e4bd5b59ac814dfedef9f93ad1
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
CMakeLists.txt and .cmake files of significant size
(more than 2 lines according to our check in tst_license.pl)
now have the copyright and license header.
Existing copyright statements remain intact
Task-number: QTBUG-88621
Change-Id: I3b98cdc55ead806ec81ce09af9271f9b95af97fa
Reviewed-by: Jörg Bornemann <joerg.bornemann@qt.io>
...by removing the entire adjustedFormat() helper.
Qt Quick has never used this, which indicates it is not that
useful. Same goes for Qt Multimedia or Qt 3D. Ensuring depth and
stencil is requested is already solved by using
QSurfaceFormat::setDefaultFormat() or by adjusting the formats
everywhere as appropriate.
The helper function's usages are in the manual tests that use it as a
shortcut, and in the GL backend itself. Remove it and leave it up the
client to set the depth or stencil buffer size, typically in the
global default surface format. (which in fact many of the mentioned
manual tests already did, so some of calls to
window->setFormat(adjustedFormat()) were completely unnecessary)
By not having the built-in magic that tries to always force depth and
stencil, we avoid problems that arise then the helper cannot be easily
invoked (thinking of widgets and backingstores), and so one ends up
with unexpected stencil (or depth) in the context (where the GL
backend auto-adjusts), but not in the window (which is not under
QRhi's control).
It was in practice possible to trigger EGL_BAD_MATCH failures with the
new rhi-based widget composition on EGL-based systems. For example, if
an application with a QOpenGLWidget did not set both depth and stencil
(but only one, or none), it ended up failing due to the context -
surface EGLConfig mismatches. On other platforms this matters less due
to less strict config/pixelformat management.
Pick-to: 6.4
Change-Id: I28ae2de163de63ee91bee3ceae08b58e106e1380
Fixes: QTBUG-104951
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
Some of the offsets are already quint32 in the API (vertex input
attributes, dynamic offsets, offsets in draw calls), matching the
reality of the underlying 3D APIs, but many buffer-related functions
use int as of now, simply because that used to be the default choice,
and the same goes for sizes (such as buffer or range sizes). This is
not quite consistent and should be cleaned up if for nothing else then
just to make the classes consistent, but also because no 3D API use a
signed type for offsets, sizes, and strides. (except OpenGL for some)
When it comes to strides (for vertex inputs and raw image texture
uploads), those are already all quint32s. This is straightforward
because most of the 3D APIs use 32-bit uints for these regardless of
the architecture.
Sizes and offsets are often architecture-dependent (Vulkan, Metal),
but there is at least one API where they are always 32-bit even on
64-bit Windows (UINT == unsigned int, D3D11). In addition, we do not
really care about buffer or texture data larger than 4 GB, at least
not without realistic use cases and real world testing, which are
quite unlikely to materialize for now (esp. since we still have the
width/height of 2D textures limited to 16 or 32K in many cases even on
desktops, whereas 2GB+ buffers are not guaranteed in practice even
when an API seemingly allows it).
In any case, the important change here is the signed->unsigned
switch. A number of casts can now be removed here and there in the
backends, because the offsets and sizes are now unsigned as well,
matching the underlying API reality. The size can be potentially
increased later on with minimal effort, if that becomes necessary for
some reason.
Change-Id: I404dbc365ac397eaeeb3bd2da9ce7eb98916da5f
Reviewed-by: Inho Lee <inho.lee@qt.io>
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
...that uses the old name after a recent change in the
name of a function.
Change-Id: Ife36fbb0c5d28b350cb1cfc48625528a205af8f9
Reviewed-by: Christian Strømme <christian.stromme@qt.io>
This convenience should be, according to the Apple docs, equivalent to
calling present from a scheduled handler. (which on its own makes it
unclear why we switched in the first place)
In practice it seems the two approaches are not identical. It looks
like that once a frame is submitted earlier than the next display link
callback, the throttling behavior we implement in beginFrame()
(waiting on the semaphore for the completion of the appropriate
command list etc.) starts exhibiting unexpected behavior, not
correctly throttling the thread to the refresh rate. Changing back to
presentDrawable does not exhibit this at all.
The suspicion is that presentDrawable is probably doing more than what
the docs suggest, and so is not fully equivalent to calling present
manually from a scheduled handler.
Therefore, switch to presentDrawable now, which restores the expected
cross-platform behavior, but make a note of the oddity, and also
prepare the hellominimalcrossgfxtriangle manual test to provide an
easy, self-contained application to allow experimenting in the future,
if needed.
This allows Qt Quick render thread animations to advance at the
expected speed (because the render thread is correctly throttled to
the refresh rate), even if the render thread decides to generate a new
frame right away, without waiting for the next display link update.
Without this patch, attempting to get updates not via requestUpdate(),
but by other means (timer etc.) leads to incorrect throttling, and so
the triangle in the test app is rotating faster than expected - but
only with Metal. Running with OpenGL on macOS or with any API on any
other platform the behavior will be correct. Even if scheduling
updates without display link is not efficient, and should be
discouraged, not doing so cannot break the core contract of vsync
throttling, i.e. the thread cannot run faster just because it renders
a frame not in response to an UpdateRequest.
Amends 98b60450f7 (effectively reverts
but keeps the code and the notes because we might want to clear this
up some day)
Pick-to: 6.4 6.3 6.2
Fixes: QTBUG-103415
Change-Id: Id3bd43e94785384142337564ce4b2644bf257100
Reviewed-by: Tor Arne Vestbø <tor.arne.vestbo@qt.io>
Replace the current license disclaimer in files by
a SPDX-License-Identifier.
Files that have to be modified by hand are modified.
License files are organized under LICENSES directory.
Task-number: QTBUG-67283
Change-Id: Id880c92784c40f3bbde861c0d93f58151c18b9f1
Reviewed-by: Qt CI Bot <qt_ci_bot@qt-project.org>
Reviewed-by: Lars Knoll <lars.knoll@qt.io>
Reviewed-by: Jörg Bornemann <joerg.bornemann@qt.io>
QPlatformTextureList holds a QRhiTexture instead of GLuint. A
QPlatformBackingStore now optionally can own a QRhi and a
QRhiSwapChain for the associated window. Non-GL rendering must use
this QRhi everywhere, whereas GL (QOpenGLWidget) can choose to still
rely on resource sharing between contexts. A widget tells that it
wants QRhi and the desired configuration in a new virtual function in
QWidgetPrivate returning a QPlatformBackingStoreRhiConfig. This is
evaluated (among a top-level's all children) upon create() before
creating the repaint manager and the QWidgetWindow.
In QOpenGLWidget what do request is obvious: it will request an
OpenGL-based QRhi. QQuickWidget (or a potential future QRhiWidget)
will be more interesting: it needs to honor the standard Qt Quick
env.vars. and QQuickWindow APIs (or, in whatever way the user
configured the QRhiWidget), and so will set up the config struct
accordingly.
In addition, the rhiconfig and surface type is (re)evaluated when
(re)parenting a widget to a new tlw. If needed, this will now trigger
a destroy - create on the tlw. This should be be safe to do in
setParent. When multiple child widgets report an enabled rhiconfig,
the first one (the first child encountered) wins. So e.g. attempting
to have a QOpenGLWidget and a Vulkan-based QQuickWidget in the same
top-level window will fail one of the widgets (it likely won't
render).
RasterGLSurface is no longer used by widgets. Rather, the appropriate
surface type is chosen.
The rhi support in the backingstore is usable without widgets as well.
To make rhiFlush() functional, one needs to call setRhiConfig() after
creating the QBackingStore. (like QWidget does to top-level windows)
Most of the QT_NO_OPENGL ifdefs are eliminated all over the place.
Everything with QRhi is unconditional code at compile time, except the
actual initialization.
Having to plumb the widget tlw's shareContext (or, now, the QRhi)
through QWindowPrivate is no longer needed. The old approach does not
scale: to implement composeAndFlush (now rhiFlush) we need more than
just a QRhi object, and this way we no longer pollute everything
starting from the widget level (QWidget's topextra -> QWidgetWindow ->
QWindowPrivate) just to send data around.
The BackingStoreOpenGLSupport interface and the QtGui - QtOpenGL split
is all gone. Instead, there is a QBackingStoreDefaultCompositor in
QtGui which is what the default implementations of composeAndFlush and
toTexture call. (overriding composeAndFlush and co. f.ex. in eglfs
should continue working mostly as-is, apart from adapting to the
texture list changes and getting the native OpenGL texture id out of
the QRhiTexture)
As QQuickWidget is way too complicated to just port as-is, an rhi
manual test (rhiwidget) is introduced as a first step, in ordewr to
exercise a simple, custom render-to-texture widget that does something
using a (not necessarily OpenGL-backed) QRhi and acts as fully
functional QWidget (modeled after QOpenGLWidget). This can also form
the foundation of a potential future QRhiWidget.
It is also possible to force the QRhi-based flushing always,
regardless of the presence of render-to-texture widgets. To exercise
this, set the env.var. QT_WIDGETS_RHI=1. This picks a
platform-specific default, and can be overridden with
QT_WIDGETS_RHI_BACKEND. (in sync with Qt Quick) This can eventually be
extended to query the platform plugin as well to check if the platform
plugin prefers to always do flushes with a 3D API.
QOpenGLWidget should work like before from the user's perspective, while
internally it has to do some things differently to play nice and prevent
regressions with the new rendering architecture. To exercise this
better, the qopenglwidget example gets a new tab-based view (that could
perhaps replace the example's main window later on?). The openglwidget
manual test is made compatible with Qt 6, and gets a counterpart in form
of the dockedopenglwidget manual test, which is a modified version of
the cube example that features dock widgets. This is relevant in
particular because render-to-texture widgets within a QDockWidget has
its own specific quirks, with logic taking this into account, hence
testing is essential.
For existing applications there are two important consequences with
this patch in place:
- Once the rhi-based composition is enabled, it stays active for the
lifetime of the top-level window.
- Dynamically creating and parenting the first render-to-texture
widget to an already created tlw will destroy and recreate the tlw
(and the underlying window). The visible effects of this depend on the
platform. (e.g. the window may disappear and reappear on some,
whereas with other windowing systems it is not noticeable at all -
this is not really different from similar situtions with reparenting
or when moving windows between screens, so should be acceptable in
practice)
- On iOS raster windows are flushed with Metal (and rhi) from now on
(previously this was through OpenGL by making flush() call
composeAndFlush().
Change-Id: Id05bd0f7a26fa845f8b7ad8eedda3b0e78ab7a4e
Reviewed-by: Tor Arne Vestbø <tor.arne.vestbo@qt.io>
.. but this will only be supported on Vulkan, OpenGL 3.2+, and Open GL
ES 3.2+ for the time being.
The situation is:
- Vulkan is working. qsb accepts .geom files already, and QShader has
existing geometry shader support.
- OpenGL 3.2 and OpenGL ES 3.2 are working.
- D3D11 is not working. D3D11 supports geometry shaders, but SPIRV-
Cross does not support translating geometry shaders to HLSL.
- Metal is not working. Metal does not directly support geometry
shaders.
Change-Id: Ieb7c44c58b8be5f2e2197bf5133cf6847e6c132d
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
Support for Polygon Mode (Triangle Fill Mode in Metal, Fill Mode in D3D)
in the RHI graphics pipeline.
Options are Fill and Line
Status:
OpenGL - ok
Vulkan - ok
Metal - ok
D3D11 - ok
OpenGL ES - does not support glPolygonMode.
Change-Id: I20b7ef416624700c3dc8d1cbe6474f4ca3889db8
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
...but this will only be supported with Vulkan and OpenGL 4.0+ and
OpenGL ES 3.2+ for the time being.
Taking the Vulkan model as our standard, the situation is the
following:
- Vulkan is ok, qsb secretly accepts .tesc and .tese files as input
already (plus QShader already has the necessary plumbing when it
comes to enums and such) To switch the tessellation domain origin to
bottom left we require Vulkan 1.1 (don't bother with
VK_KHR_maintenance2 on top of 1.0 at this point since 1.1 or 1.2
implementations should be common by now). The change is essential to
allow the same evaluation shader to work with both OpenGL and
Vulkan: this way we can use the same shader source, declaring the
tessellation winding order as CCW, with both APIs.
- OpenGL 4.0 and OpenGL ES 3.2 (or ES 3.1 with the Android extension
pack, but we won't bother with checking that for now) can be made
working without much complications, though we need to be careful
when it comes to gathering and setting uniforms so that we do not
leave the new tessellation stages out. We will stick to the Vulkan
model in the sense that the inner and outer tessellation levels must
be specified from the control shader, and cannot be specified from
the host side, even though OpenGL would allow this. (basically the
same story as with point size in vertex shaders)
- D3D11 would be no problem API-wise, and we could likely implement
the support for hull and domain shader stages in the backend, but
SPIRV-Cross does not support translating tessellation shaders to
HLSL. Attempting to feed in a .tesc or .tese file to qsb with
--hlsl specified will always fail. One issue here is how hull
shaders are structured, with the patchconstantfunc attribute
specifying a separate function computing the patch constant
data. With GLSL there is a single entry point in the tessellation
control shader, which then performs both the calculations on the
control points as well as the constant data (such as, the inner and
outer tessellation factors). One option here is to inject
handwritten HLSL shaders in the .qsb files using qsb's replace (-r)
mode, but this is not exactly a viable universal solution.
- Metal uses a different tessellation pipeline involving compute
shaders. This needs more investigation but probably not something we
can prioritize in practice. SPIRV-Cross does support this,
generating a compute shader for control and a (post-)vertex shader
for evaluation, presumably in order to enable MoltenVK to function
when it comes to tessellation, but it is not clear yet how usable
this is for us.
Change-Id: Ic953c63850bda5bc912c7ac354425041b43157ef
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
Mainly because we do have legacy code in the Qt 5 graphical effects that
tries to dynamically determine the max number of varyings. Make it
easier to port such code.
Change-Id: I846cab2c2fe7b4cd473b5ced0146ca36f1c8169b
Reviewed-by: Qt CI Bot <qt_ci_bot@qt-project.org>
Reviewed-by: Christian Strømme <christian.stromme@qt.io>
The system we inherited from the original Qt 5.14 introduction of QRhi
is a text stream based solution where resource creation and frame
timings are sent in a comma-separated format to a QIODevice.
This, while useful to get insights about the number of resources at a
given time, is not actively helpful. The frameworks built on top (Qt
Quick, Qt Quick 3D) are expected to provide solutions for logging
timings in a different way (e.g. via the QML Profiler). Similarly,
tracking active resources and generating statistics from that is
better handled on a higher level.
The unique bits, such as the Vulkan memory allocator statistics and
the GPU frame timestamps, are converted into APIs in QRhi. This way a
user of QRhi can query it at any time and do whatever it sees fit with
the data.
When it comes to the GPU timestamps, that has a somewhat limited value
due to the heavy asynchronousness, hence the callback based
API. Nonetheless, this is still useful since it is the only means of
reporting some frame timing data (an approx. elapsed milliseconds for
a frame) from the GPU side.
Change-Id: I67cd58b81aaa7e343c11731f9aa5b4804c2a1823
Reviewed-by: Laszlo Agocs <laszlo.agocs@qt.io>
Arrays of textures have always been supported, but we will encounter
cases when we need to work with texture array objects as well.
Note that currently it is not possible to expose only a slice of the
array to the shader, because there is no dedicated API in the SRB,
and thus the same SRV/UAV (or equivalent) is used always, capturing
all elements in the array. Therefore in the shader the last component
of P in texture() is in range 0..array_size-1.
Change-Id: I5a032ed016aeefbbcd743d5bfb9fbc49ba00a1fa
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
QT_VERSION is now at least QT_VERSION_CHECK(6, 3, 0), so remove all
checks against Qt 6.0.0 or earlier. They are superfluous. Tidied up in
some places in the process, particularly #include order.
Change-Id: I2636b2fd13be5b976f5b043ef2f8cddc038a72a4
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
Supported on OpenGL (and ES) 3.0+ and everywhere else.
Can also be a render target, targeting a single slice at a time.
Can be mipmapped, cannot be multisample.
Reading back a given slice from a 3D texture is left as a future
exercise, for now it is documented to be not supported.
Upload is going to be limited to one slice in one upload entry,
just like we specify one face or one miplevel for cubemap and
mipmapped textures.
This also involves some welcome hardening of how texture subresources
are described internally: as we no longer can count on a layer index
between 0..5 (as is the case with cubemaps), simply arrays with
MAX_LAYER==6 are no longer sufficient. Switch to sufficiently dynamic
data structures where applicable.
On Vulkan rendering to a slice needs Vulkan 1.1 (and 1.1 enabled on the
VkInstance).
Task-number: QTBUG-89703
Change-Id: Ide6c20124ec9201d94ffc339dd479cd1ece777b0
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
Add QRhi APIs to retrieve and reload the contents of the "pipeline
cache".
The only API where there is a true pipeline cache is object is Vulkan
(VkPipelineCache). For OpenGL, the other backend where we support this,
it is simulated with program binaries. The Qt 5 style OpenGL program
binary disk cache continues to work like before, but one has now the
option to do things in a more modern, graphics API agnostic way, that
leads to generating a single blob instead of a large set of files in
some system location, allowing easier "pre-baking" of the cache content.
It is expected that Qt Quick exposes the two new functions in form
if QSG_RHI_ environment variables, thus allowing easy testing and
cache file generation.
As an example for the performance improvements this can give, consider
Vulkan, where we do not have any existing persistent caching mechanism
in place:
Running BenchmarkDemoQt6.exe --scene flythrough --mode demo creates 18
QRhiGraphicsPipeline objects from Qt Quick and Qt Quick 3D.
The total time spent in QRhiGraphicsPipeline::create() during application
startup for these 18 pipelines is 35-40 ms on a given Windows (NVIDIA)
system.
When exporting the pipeline cache contents to a file, and then, in a
subsequent run, reloading the cache contents, this is reduced to 5-7 ms
on the same system, meaning we get a 6-7x improvement.
The generated data is always specific to a given Qt version, RHI
backend, graphics device, and driver version. Much of the implementation
consists of adding and verifying the appropriate header to the blobs
retrieved from the driver, to allow gracefully ignoring data that was
generated with a device or driver that differs from the one used at
run time. This should provide robustness, even if the Vulkan or OpenGL
implementation is for some reason not prepared to identity and reject
incompatible cache/program blobs.
Fixes: QTBUG-90398
Change-Id: I67b197f393562434f372c7b7377f638abab85cb3
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
Instead, have a static function in QRhiVulkanInitParams then Qt Quick
and anyone else who creates a QVulkanInstance that is then used in
combination with QRhi can query.
Change-Id: I046e0d84541fc00f5487a7527c97be262221527f
Reviewed-by: Andy Nichols <andy.nichols@qt.io>
