Performance of QQuickWidget

p3c0

Hi @zepfan, I don't have the exact reason as I don’t have enough experience with using QQuickWidget but the documentation mentions few points related to performance hit that may occur. Check out the 3 notes under that topic. Can you try using QQuickView and createWindowContainer() as according to this blog too:

Well, unless there is a very good reason for using it and you know what you are doing. If the stacking and other limitations do not apply and the application is targeting desktop platforms only and the platform in question is able to use the Qt Quick Scenegraph’s threaded render loop, then having a QQuickView embedded via createWindowContainer() will always lead to better performance when compared to QQuickWidget

zepfan

Thanks for the response. I tried using a QQuickView and yes the performance is better, but it still uses the same CPU % whether there is one meter active or if there are 20 active. So it seems there is still unnecessary repainting happening on the other 19 meters. Is there a way to prevent repainting of elements that don't need it?

Thanks.

p3c0

@zepfan

So it seems there is still unnecessary repainting happening on the other 19 meters.

Are you sure they are repainted everytime ? How did you check ?
Is this VU meter a C++ class derived from QQuickPaintedItem ?

zepfan

No I'm not sure that they're getting repainted, but if I have just one active meter visible and nothing else, the CPU % goes down. But if I have 20 visible and only one active then the CPU is much higher. Since the other 19 shouldn't be getting repainted why would the CPU still be higher? The 19 meters that are inactive just have a black opaque background.
The VU meters are built just using qml. Only the QQuickView derived class is C++ and it has basically nothing in it. This is just a small test app I've built with no frills.

Any ideas?

Thanks.

p3c0

@zepfan I guess there could be some other problem then. May be some connections or properties that are getting updated. Even if it gets repainted I think CPU usage wont go that much high as QML uses scenegraph which in turn uses OpenGL so most of the rendering is done by the hardware.
Can you post a complete minimal runnable example to test ?

zepfan

Here is the qml file that defines the meter. I use a Repeater to create multiple instances of these and give each one a unique index (meterIndex).
In the code you can see I check to see if it is the first meter (meterIndex = 0) and if so then only that one's clip rectangle is active. I've since realised that setting the rest to invisible actually brings down the CPU %. But if I leave them visible then the CPU increases, even though there's no activity. Is it normal to set items to invisible if you don't want anything rendered for them?

import QtQuick 2.0

Rectangle {
  id: myMeter
  width: 20
  height: 100
  property int meterIndex: 0

  color: "black"

  Rectangle {
    id: clipRect
    width: parent.width
    height: 0
    clip: true

    Timer {
      id: myTimer
      interval: 20
      repeat: true
      running: true
      triggeredOnStart: true
      onTriggered: {
        if(myMeter.meterIndex == 0) {
          clipRect.height = myMeter.height * Math.random();
        }
        else {
          clipRect.visible = false;
          myTimer.stop();
        }
      }
    }
    color: "blue"
  }
}

p3c0

@zepfan I see you are starting a Timer of interval 20 and it seems you are starting it for all the other remaining meters. A Timer of such a low interval and that too for 20 items is ofcourse going to be CPU intensive. I would suggest here to not start the Timer for the rest in first place. You can add s condition like:

Timer {
  running: myMeter.meterIndex == 0
}

So the Timer will start only for the Item with meterIndex = 0

Is it normal to set items to invisible if you don't want anything rendered for them?

Yes it reduces the cost of drawing again. See over-drawing-and-invisible-elements.

zepfan

So if I have other elements on screen but they have not changed, does that mean they will be re-rendered every time also? I can't change them to (visible: false) as obviously I need to see them!

Is there a flag/setting that you can set that will make sure only items that are 'dirty' will be rendered?

Thanks

p3c0

@zepfan No change no rendering again.

Is there a flag/setting that you can set that will make sure only items that are 'dirty' will be rendered?

I'm not aware of any such in QML.
But as I explained earlier your timer is eating the CPU. Even if Item is invisible timer will be active. 20 ms timeout for 20 items is CPU intensive.

zepfan

@p3c0 said:

But as I explained earlier your timer is eating the CPU. Even if Item is invisible timer will be active. 20 ms timeout for 20 items is CPU intensive.

True, but if I only have the timer running for one meter there is still quite a difference in CPU % between having the idle meters' clipRect visible or not visible. If they're not being repainted (since they're not active) then there shouldn't be a difference in CPU %?

p3c0

@zepfan How much CPU is being utilized on your system ? Can you post a complete example test it out ?

zepfan

If you use the code above for the actual Meter (MyMeter), and here is its parent which uses a Repeater to construct 20 of them. Try having the timer just running in the first meter, compared with having the timer running in all of them. I don't notice much change in CPU when having one of them getting its clipRect.height changed or having all of them get their clipRect.height changed.

I should state that the reason I've made this test application is because I'm doing something similar in work (which I obviously can't post) and am getting similar results. My suspicion is that the whole scene is getting re-rendered rather than just the item that needs it.

import QtQuick 2.0

Rectangle {
  visible: true
  width: 900
  height: 700
  color: "green"

  Column {
    id: myColumn
    spacing: 10

    Repeater {
      model: 2

      Row {
        id: myRow
        x: 50
        y: 50
        spacing: 40
        property int rowIndex: index

        Repeater {
          id: innerRepeater
          model: 10

          MyMeter {
            objectName: "Meter " + (index + (myRow.rowIndex * innerRepeater.model))
            meterIndex: (index + (myRow.rowIndex * innerRepeater.model))

            width: 15
            height: 150
          }
        }
      }
    }
  }
}

p3c0

@zepfan Just tested. I do notice change in CPU %. For single Meter running it is around 3-3.5% while with all meters running it is around 7.5-8.3%.
How much on your system ?

zepfan

I don't have the exact figure on me at the moment, but the figure itself isn't relevant. For my test app the figure wasn't too high, but my concern is that the CPU % doesn't change much whether I have one meter active or 20 meters active. Once they are all visible it's pretty much the same. And obviously this affects the real project I'm working on even more.
I've actually also just posted in the QtQuick forum just to see if anyone has any thoughts on rendering performance/improvements.

Thanks

p3c0

@zepfan As told earlier the rendering is all done by OpenGL which in turn is handled by the GPU so that wont matter. The Timer is what is affecting the CPU.
QtQuick renderer is capable of handling thousands of items at a time. Please check it here. Also here are some benchmark examples.
Extremetable example loads 100000 items as and when required.
Anyway here are some more link that you may fine useful:
http://doc.qt.io/qt-5/qtquick-performance.html
http://www.sletta.org/apps/blog/show/42708211-looking-at-throughput-in-qt-quick
http://doc.qt.io/qt-5/qtquick-visualcanvas-scenegraph-renderer.html
Try profiling your example.