Re: [MACTCL] ATSUI patches?

[Daniel A. S. <st...@ic...>]

On 21/03/2006, at 10:35, Daniel A. Steffen wrote:

> On 21/03/2006, at 6:01, Jeff Hobbs wrote:
>
>>> So, I have to ask, what exactly are we waiting for before
>>> committing this to cvs (at least to the 8.5 development
>>> branch)? Looking through the tracker, it seems we just wanted
>>> Jeff to sign off on it. Perhaps I'll drop him a line...
>>
>> Jeff is always lurking on the line.  ;)
>>
>> As Vince has given the thumbs up, I will commit to the 8.5 HEAD
>> using Benny's patch tomorrow (PST), if there are no other
>> objections.  This will get it into 8.5a4.
>
> you should test tkchat with this, with large buffers, that's what  
> killed my plan to commit last the atsui patch time around, see  
> below. I'll try to run a few test myself with the new patch later  
> today.

I see the same performance issues as previously with tkchat (latest  
kit from sdarchive): while loading 1000 lines of chat history, text  
layout is so slow that it leads to spinning cursor and loss of app  
responsiveness on my 1GHz G4. This is no doubt due to the continuous  
re-layouts of tkchat's complex text widget (w/embedded smileys, bold  
fonts etc) during history loading, but this is something that the QD  
based text code handles with no trouble.

I sharked an optimized unstripped tktest during the history loading  
phase of tkchat, results below, 92% of total CPU time is spent  
measuring text, with ATSUNextCursorPosition and ATSUOffsetToPosition  
being the individual APIs using the most time. Actually drawing the  
text uses a magnitude less time than measuring it (which is certainly  
at least partially due to caching by ATSUI)

I'll have a more detailed look at TkMacOSXQuarzStartDraw,  
superficially it appears that it should be easy to improve its  
performance somewhat with added caching.

Despite this, I support committing this to HEAD now, so that others  
can test if performance is really a major issue in practice, esp  
since no work has been done on performance since I first reported the  
issue almost 2 years ago... however, I would hold off on committing  
to core-8-4-branch until we have a better idea about the performance  
impact.

I have attached an updated patch for HEAD to SF patch 638966 which  
fixes a number of (mostly cosmetic) small issues with Benny's patch:  
non-stub exports, printing to stderr, function naming conventions,  
whitespace, xcode project additions...

If there are no objections, I'm happy to commit this along with some  
other changes that came up while looking into Benny's patch in  
detail, or Jeff, feel free to commit my patch whenever convenient, my  
other changes are minor...

Cheers,

Daniel

-- 
** Daniel A. Steffen        Dept. of Mathematics  **
** Macquarie University     NSW 2109 Australia    **

Shark: system libraries charged to callers
	Self	Total	Lib	
	42.5%	42.5%	tktest	TkpMeasureCharsInContext	
	31.8%	31.8%	tktest	MeasureStringWidth	
	13.4%	13.4%	tktest	TkMacOSXQuarzStartDraw	
	3.4%	3.4%	tktest	TkpMacOSXLayoutSetString	
	1.3%	1.3%	tktest	TkMacOSXQuarzEndDraw	

Shark: system libraries flattened
	Self	Total	Lib	
	26.1%	26.1%	QD	ATSUNextCursorPosition	
	0.0%	26.1%	tktest	  TkpMeasureCharsInContext	
	23.4%	23.4%	QD	ATSUOffsetToPosition	
	0.0%	23.4%	tktest	  MeasureStringWidth	
	8.8%	8.8%	QD	ATSUBreakLine	
	0.0%	8.8%	tktest	  TkpMeasureCharsInContext