[Interest] [Qt3D] stability and performance issues with Qt 5.8

[Xavier Bigand]

I found some things that can explain my stability issues (Qt 5.8) :

   1. We can not let a parameter with a nullptr value
      - It will lead to an unsupported uniform type
      - I have to detach my texture parameter in this case, but
      add/removeParameter method are slow
      - I had try to use the enable flag instead, but it doenst works
   2. I can't share the QShaderProgram instances between my materials, I
   get an assert that I don't really know I it came from me or just because
   QShaderProgram shouldn't be shared.
      - Having one instance of QShaderProgram per material seems to be
      unoptimal in my case as I can generate a key for our ubershader
and having
      a map


The assert came from the method void
Renderer::prepareCommandsSubmission(const QVector<RenderView *>
&renderViews) :
Q_ASSERT(rGeometry && rGeometryRenderer && shader);


My goal is to be able to change often the ubershader generated code and the
corresponding uniforms of materials. I am trying to optimize it, because it
is easy on our side to do it and avoiding Qt3D thread synchronization,...





2016-10-21 18:15 GMT+02:00 Xavier Bigand <flamaros.xavier at gmail.com>:

> I am actually use Qt 5.8 branch from Git, I have some performances and
> stability regression compared to the 5.7 release.
>
> I can't run my application in Debug mode (tested with mingw and VS 2013)
> with both Qt versions (5.7 and 5.8=), it runs in release but behavior
> change fast when I do very small changes in my code.
> It seems that the version 5.7 is able to support more complexe scenes than
> Qt 5.8 and performances are twice better.
>
> As much as I can tell for the moment stability/performances issues cames
> from threading management.
>
>
> When trying to load a scene in debug mode with Qt 5.8 it often break in T
>  QGenericAtomicOps::load(const T &_q_value) from a pooled thread.
>
> Here is the full call stack :
> > Qt53DRenderd.dll!QGenericAtomicOps<QAtomicOpsBySize<4>
> >::load<int>(const int & _q_value) Line 90 C++
>   Qt53DRenderd.dll!QBasicAtomicInteger<int>::load() Line 99 C++
>   Qt53DRenderd.dll!QtPrivate::RefCount::isShared() Line 101 C++
>   Qt53DRenderd.dll!QVector<Qt3DRender::Render::ClearBufferInfo>::isDetached()
> Line 108 C++
>   Qt53DRenderd.dll!QVector<Qt3DRender::Render::ClearBufferInfo>::detach()
> Line 384 C++
>   Qt53DRenderd.dll!QVector<Qt3DRender::Render::ClearBufferInfo>::begin()
> Line 204 C++
>   Qt53DRenderd.dll!Qt3DRender::Render::FrameGraphVisitor::visit::__l78::<lambda>()
> Line 326 C++
>   Qt53DRenderd.dll!Qt3DRender::Render::GenericLambdaJob<void
> <lambda>(void) >::run() Line 78 C++
>   Qt53DCored.dll!Qt3DCore::AspectTaskRunnable::run() Line 91 C++
>   Qt5Cored.dll!QThreadPoolThread::run() Line 99 C++
>   Qt5Cored.dll!QThreadPrivate::start(void * arg) Line 380 C++
>   [External Code]
>   [Frames below may be incorrect and/or missing, no symbols loaded for
> kernel32.dll]
>
> It breaks because the QBasicAtomicInteger pointer is allready free
> (0xFEEEFEEE).
>
> And for performance issues it seems to come from synchronization between
> thread, so I am trying to reduce the number of Qt3D objects
> (QShaderProgram, Texture,...) by sharing them, but doing this can lead in
> rendering issues or crashs.
>
>
> I want to help to improve stability (first) and performances, but I don't
> have a great experience in thread debugging and on how the Qt3D backend
> works. So I would be happy to get some hints to improve my feedbacks or
> doing more to help.
>
>
> --
> Xavier
>



-- 
Xavier
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.qt-project.org/pipermail/interest/attachments/20161022/7927e3f4/attachment.html>