Review Request 119336: Convert FrameSvg to 9 textures: different approach

[Marco Martin]

> On July 17, 2014, 4:54 p.m., Eike Hein wrote:
> > Unless I'm missing something, this approach seems to defeat the purpose of the exercise: The goal of the original change is to use actually use smaller textures that fit the QSG's "this can use the atlas" heuristic on most platforms. You have one node per patch element here, but they're all calling createTextureFromImage() on a big honking QImage that's not going to be atlased for most frames.
> > 
> > Can you explain your pixel-perfect quality concerns better - you've said the competing approach breaks things there but not mentioned what or why?
> 
> Marco Martin wrote:
>     That's a button that uses composeoverborders: see the corners/sides?
>     http://wstaw.org/m/2014/07/17/plasma-desktopvc1725.png
>     (overlays would give a similar problem, not only the internal is painted over the borders, but it's also cutted with an alpha mask to match)
>     That's why i'm using framesvg to paint with the local qpainter instead of just taking the side image.
>     
>     As for defeating the purpose yeah, it does (what actually does is just being faster duting animated resizes), but in simple cases (no composeoverborders, no overlays) the timer may be disabled altogether.
> 
> Eike Hein wrote:
>     > composeoverborders
>     
>     Right, but doesn't have David's patch a fallback codepath for that? Obviously I'd heavily -1 everything causing visual regressions too.
>     
>     About the symbol exports: I don't think they're a major issue. We're not installing the header, things aren't exposed to developers above the lib, and it's certainly not an uncommon pattern in Qt (there's data() methods all over the place, even in public APIs).
>     
>     On the whole I think David's patch has some compelling advantages in the ideal case that may warrant the complexity (I don't think complexity in low-level components is a big deal if it's Worth It[TM] and just needs to be debugged once; this isn't code that changes often).
>     
>     How do we progress here?
> 
> David Edmundson wrote:
>     To credit Marco, I've only just added that to the fallback path; thanks for pointing it out.
>     We can then recommend themers avoid it; am I right in thinking it was just for QPainter optimisations?

the screenshot i posted is with david's patch, and you can see the button background one pixel outside its corners.
the fallback here is to yes rerender the scaled svg, but with Svg::image(), that only returns the rendered element id, not the composition with alpha masks and all that jazz. Also since it immediately rerenders the svg in that case it loses the advantage when the resize it's animated.


- Marco


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://git.reviewboard.kde.org/r/119336/#review62607
-----------------------------------------------------------


On July 17, 2014, 12:47 p.m., Marco Martin wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://git.reviewboard.kde.org/r/119336/
> -----------------------------------------------------------
> 
> (Updated July 17, 2014, 12:47 p.m.)
> 
> 
> Review request for KDE Frameworks and Plasma.
> 
> 
> Repository: plasma-framework
> 
> 
> Description
> -------
> 
> This is a derivative of https://git.reviewboard.kde.org/r/119330/
> 
> It has a much simpler codebase and it doesn't touch framesvg in the library (and doesn't make things public)
> 
> the important differences are two, that are imo 100% necessary to maintain a pixel perfect rendering (sacrificing that is a regression simply not acceptable in any case, even for non default themes, ever). At the same time avoids duplication of framesvg code in framesvgitem.
> 
> potential issues:
> * e037203748 support for tiling still need porting
> * yes, it uses a qpainter, that means another copy: but again is necessary as framesvg knows how to render the end result pixel perfect, since for many themes the end piece is *not* the simple rendered element id.
> * yes, the final scaled texture is still uploaded as a whole, it only avoids to do it too often (like in animations) with event compression, again, only way to be sure it's correctly rendered all the time.
> 
> The latter two points can have the following optimizations:
> Iff the frame does not have an overlay and does not have composeoverborders set, the following can be done:
> * use Svg::image() to fetch the pixmap of the piece, since in that case would be valid
> * resize the framesvg one single time, at a fixed size , like something > 256x256 (256 is not random thing, since we have 8 bits per channel, usually gradients will have no more than 256 stops) and disable the resize timer, this way we are even sure that only one image per prefix will be stored in cache
> 
> I should add regarding the last two optimization points: i would like to see some real benchmarks about them, or i would not consider then necessary until then.
> 
> 
> Diffs
> -----
> 
>   examples/applets/widgetgallery/contents/ui/Buttons.qml 379585f 
>   examples/applets/widgetgallery/contents/ui/Menu.qml 1336c42 
>   examples/applets/widgetgallery/contents/ui/standalonemain.qml PRE-CREATION 
>   src/declarativeimports/core/framesvgitem.h e155f6a 
>   src/declarativeimports/core/framesvgitem.cpp 8320212 
>   src/declarativeimports/core/svgitem.cpp 1ed0631 
>   tests/dialog.qml PRE-CREATION 
>   tests/testborders.qml PRE-CREATION 
> 
> Diff: https://git.reviewboard.kde.org/r/119336/diff/
> 
> 
> Testing
> -------
> 
> 
> Thanks,
> 
> Marco Martin
> 
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/kde-frameworks-devel/attachments/20140717/356117d4/attachment-0001.html>