Re: Failing VCL complextext rendering unit test on Windows

On Fri, Mar 24, 2017 at 02:01:34PM +0100, Jan-Marek Glogowski wrote:
> Am 24.03.2017 um 12:32 schrieb Khaled Hosny:
> > On Fri, Mar 24, 2017 at 11:29:52AM +0100, Jan-Marek Glogowski wrote:
> > > Hi everyone,
> > >
> > > when running
> > >
> > > $ /opt/lo/bin/make-4.2.1-msvc -C vcl CppunitTest_vcl_complextext
> > >
> > > on Windows 7, KVM QXL GUI via Spice I get the following failures:
> > […] 
> > > All height values are the same, just the width and x-offsets change.
> >
> > Hmm, might be related to Windows font rendering settings, AFAIK Windows
> > disables anti-aliasing for rdesktop (and I guess this what Spice uses)
>
> No. Spice is a protocol (https://www.spice-space.org/) and has it's own
> graphic driver (QXL). I don't know how rdesktop and Spice work in
> detail, but my guess is it's different. QXL is more in line with the
> vmware and virtualbox video drivers, which allow to resize the viewers
> window and adapt the VMs graphic output size.

I see.

> > and different font rendering strategy can result in different glyph
> > widths. Only bounding box is affected since we still use Windows API
> > for it (unlike glyph widths for layout).
> >
> > > From the documentation, the provided matrix to the function is the
> > > identity matrix, which just leaves some room for different HDC.
> > >
> > > AFAIK HDC is driver dependent (whatever driver is used for SSH) and so
> > > I'm not sure this is fixable at all.
> > >
> > > So I propose disabling it also on Windows.
> >
> > I suppose we have no way to change the font settings during the testing,
> > so disabling this part of the test would be the simplest. Alternatively
> > we can add some tolerance to the test (a pixel or two?) but no idea how.
>
> No idea, if possible.
> https://gerrit.libreoffice.org/#/c/35647/ fixes the problem for me.
>
> Do you remember, if there were many different results on Linux tinderboxes?
> Because your commit message mentioned them as the reason to run the test
> only on Windows and Mac.

The tinderbox emails don’t give a link to the log, but quoting from the
email:

> complextext.cxx:64:Assertion
> Test name: VclComplexTextTest::testArabic
> equality assertion failed
> - Expected: 71x14@(0,1)
> - Actual  : 70x14@(1,1)

Looks similar to yours, which makes me wonder if the failures have a
common cause.

Regards,
Khaled