[launcher] color banding on ARM

[Florian Boucault]

On at least 2 ARM devices the launcher exhibits color banding as shown in the attached screenshot.

On these devices executing xwininfo on the launcher:

xwininfo | grep "Visual Class"
xwininfo | grep "Depth"

report:

 Visual Class: TrueColor
 Depth: 16

whereas 'Depth' should be at least 24.

[Florian Boucault (fboucault)]

A workaround with the current unity-qt package is to start the launcher using the -visual switch:

unity-qt-launcher -visual TrueColor

Investigating whether or not QApplication::setColorSpec(QApplication::ManyColor) is equivalent.

[Florian Boucault (fboucault)]

Qt's source code gives us the answer in:

src/gui/painting/qcolormap_x11.cpp:QColormap::initialize()

        [...]
        else if ((X11->visual_class != -1 && X11->visual_class >= 0 && X11->visual_class < 6)
                   || (X11->visual_id != -1)) {
            // look for a specific visual or type of visual
            d->visual = find_visual(display, i, X11->visual_class, X11->visual_id,
                                    &d->depth, &d->defaultVisual);
        } else if (QApplication::colorSpec() == QApplication::ManyColor) {
            // look for a TrueColor w/ a depth higher than 8bpp
            d->visual = find_visual(display, i, TrueColor, -1, &d->depth, &d->defaultVisual);
            if (d->depth <= 8) {
                d->visual = DefaultVisual(display, i);
                d->defaultVisual = true;
                color_count = 216;
            }
        [...]

where X11->visual_class is the value of the command line switch -visual

Looking at the code after that shows that when using XRender (-graphicssystem native) it defaults to a TrueColor visual.

[Bill Filler (bfiller)]

running unity-qt-launcher -visual TrueColor is much better. Banding of icons is gone. However there is still banding of the launcher panel but it is not terrible. I'll attach a screenshot.

[Florian Boucault (fboucault)]

Not good indeed. Were you observing the same banding when using -graphicssystem native?

[Bill Filler (bfiller)]

no, when using native graphicssystem there is no banding on the panel. I would go ahead and commit the change as it makes the situation better but leave the bug open to continue to investigate the panel banding.

[Florian Boucault (fboucault)]

@Bill: Can you post the result of xdpyinfo please?

[Bill Filler (bfiller)]

attaching xdpyinfo

[Bill Filler (bfiller)]

downgrading and moving to m4 as we can live with banding of panel for now

[Bill Filler (bfiller)]

Testing verion bzr289 on ARM and the dash/places performance is very bad again. I'm suspecting it's because of the QApplication::setColorSpec(QApplication::ManyColor); call that was added to places/apps/places.cpp.

If I run
'qmlviewer /usr/share/unity-qt/places/dash.qml' the performance of the dash and places is good. This defaults to -graphicssystem raster.

But running unity-qt-places directly or with the -graphicssystem raster displays very poor performance, so I'm guessing we are doing something else in unity-qt-places exe that is causing the issue.

[Florian Boucault (fboucault)]

@Bill: after careful investigation I was unable to reproduce the performance issue you are describing on the EFIKA board. I am using unity-qt-places 0.2~bzr289

[Bill Filler (bfiller)]

Are you saying running qmlviewer and unity-qt-places yields identical performance? I am seeing a big difference on Kunlun imx51 arm board.

On Nov 24, 2010, at 9:09 PM, Florian Boucault <email address hidden> wrote:

> @Bill: after careful investigation I was unable to reproduce the
> performance issue you are describing on the EFIKA board. I am using
> unity-qt-places 0.2~bzr289
>
> --
> [launcher] dash performance poor on ARM
> https://bugs.launchpad.net/bugs/674484
> You received this bug notification because you are a member of The
> Upicek team, which is a direct subscriber.

[Florian Boucault (fboucault)]

Exactly, I cannot see any noticeable difference between qmlviewer and unity-qt-places on the EFIKA.

[Aurélien Gâteau (agateau)]

Same thing here: no difference between qmlviewer and unity-qt-places on the EFIKA.

[Ricardo Salveti (rsalveti)]

qmlviewer seems a lit bit faster on Panda, but running unity-qt-places with -graphicssystem raster doesn't get me a very poor performance, actually it's quite near from what I get with qmlviewer.

Also using 0.2~bzr289.