Re: [scintilla] Pango support on GTK+ 2.x

[Owen Taylor <otaylor redhat com>]

On Mon, 2003-03-10 at 16:20, Neil Hodgson wrote:
> Owen Taylor:
> 
> >  - I'm not sure it will be faster currently, but I'd definitely
> >    suggest using one layout per line and setting color attributes
> >    on ranges.
> 
>    That turned out to be slightly (~5%) slower. Possibly it was the extra
> allocation and setup of a font description and colour attribute for each
> run.

I'm not completely surprised that it's a little slower ... but that 
can, to some extent, be optimized in Pango, and by not hiding
information from Pango/GDK

 - The rendering will be better

 - GDK can (in the future anyways) be smarter about how it's drawing
   colors - reusing GC's as appropriate, etc.

> >  - Is the problem layout speed or rendering speed? I'd guess that
> >    scintilla can assume constant-height-lines, but perhaps
> >    it still needs to go over the entire file to deal with
> >    line widths. (or if there is wrapping involved)
> 
>    Layout speed appears the big problem. Scintilla performs two runs over
> text to be displayed, the first to measure the text and the second to render
> the text. The measurement is required to perform wrapping as you say. The
> results of the measurement are placed in a 'line layout' cache so most
> editing can use the previously calculated positions. My tests here have
> mostly been with maximum caching turned on but without wrapping. The
> measurement is lazy and normally only occurs when a line needs to be
> measured for display. Wrapping often requires measuring the whole file.
> There is certainly a cost to performing the layout twice and I could attempt
> to defer position calculations under some conditions to render time.

For GtkTextView, we went with the philosophy:

 If we can't do one screenful of layout fast enough for rendering,
 we aren't going to be able do the entire buffer fast enough
 to set up the scrollbars in a useful amount of time.

And it's seemed to work pretty well.

The only things that are cached are:

 - The size (width x height) of each paragraph in the buffer.

 - A single PangoLayout (and associated information) for the
   last paragraph laid out. This helps for things like moving
   a cursor around a paragraph where the same paragraph is
   accessed repeatedly.

It's not really clear to me how you can defer measuring lines
until a line is displayed, even if you base your scrollbar
on lines rather than pixels. How do you handle the horizontal
scrollbar?

> >  - It takes GtkTextView around 1.5s to lay out a 11,000 line
> >    source file on 1 ghz machine.
> 
>    My result was a 190K, 6000 line source code file taking about 8 seconds
> to completely lay out on a 600MHz Athlon running RedHat 8.0 with its default
> GTK+, X and font system.

Please do try Pango CVS, it really has quite different performance
characteristics, it's really hard to speak meaningfully about older
versions at this point.

Regards,
                                         Owen