AI's advice how to optimize GTK2 TextOut

[AlexTP]

My question:
Lazarus IDE project from GitLab, https://gitlab.com/freepascal.org/lazarus/lazarus . Analyze the GTK2 widgetset code from folder "lcl/interfaces/gtk2". Interesting part is method ExtTextOut from unit gtk2int, its code is in the file "lcl\interfaces\gtk2\gtk2winapi.inc". This method ExtTextOut is very slow, ie if I call it 200 times per one frame rendering, it spends lot of time. Find out how can I make it faster? Maybe some Pango objects can be cached better? If so, tell me how.

Grok's answer:
https://grok.com/c/80e6b290-418a-4fd2-96ee-979218a494fb?rid=cbf687cf-8884-4a44-b134-fd17dd7c3484

Somebody can add caching of Pango objects? I cannot.

[Thaddy]

Block the frame from paint updating until done? You know what I mean.

[ALLIGATOR]

As far as I know, this person also faced performance limitations. And as a result, he does it like this

UPD:
https://github.com/AlexanderBagel/FWHexView/blob/c3e06b3dc42a494d7683ab8bef694bb89c6c1957/src/FWHexView.Common.pas#L367

[AlexTP]

Thanks @ALLIGATOR. I contact this person now and will use his code. License for this part was changed to 'MPL2 or LGPL' by kind permission of this person.

[JuhaManninen]

The AI is quite impressive! After studying the code for few seconds it already knows it better than I do.
Did somebody try the "Pango Caching Improvements" it suggested? How much they affect?
I guess a similar optimization could be used for LCL-GTK3. That would be even more important for the future.

[ALLIGATOR]

Just for fun, to understand what this is all about, I tried this code:

Code: Pascal  [Select][+][-]

1. program app;
2.  
3. uses
4.   Interfaces, LCLIntf, LCLType, SysUtils, Classes, Graphics, DateUtils;
5.  
6. var
7.   B: TBitmap;
8.   i: Integer;
9.   t: TDateTime;
10.   c: HDC;
11.   s: String = '0123456789';
12.  
13. begin
14.   B := TBitmap.Create;
15.   B.PixelFormat := pf24bit;
16.   B.SetSize(100, 100);
17.  
18.   c := B.Canvas.Handle;
19.  
20.   t:=Now;
21.   for i:=100*1000 downto 0 do
22.     ExtTextOut(c, 0, 0, 0, nil, PChar(s), Length(s), nil);
23.   WriteLn(MilliSecondsBetween(Now, t), 'ms');
24.  
25.   B.Free;
26.  
27.   {$IFDEF WINDOWS}ReadLn;{$ENDIF}
28. end.

Result:
Windows: 500ms
Linux GTK2: 900ms

Am I testing correctly? Is this the difference in performance we are talking about, or is it greater?

[JuhaManninen]

Am I testing correctly? Is this the difference in performance we are talking about, or is it greater?

What does @AlexTP say?
I personally did not realize the code is slow until he mentioned it.

[AlexTP]

@Juha, @Alligator
GTK2 ExtTextOut is slower of coz, than Win32 API. Seems you are testing ok, but I thouhgt that the ratio between gtk2/win32 must be bigger than 900/500, my looking shows it (ratio) is like 2...5. your demo did not pass Dx (inter-char deltas) non-nil value to ExtTextOut.

[ALLIGATOR]

With this code
Windows: 500 ms
Linux GTK2: 14000 ms

Code: Pascal  [Select][+][-]

1. program app;
2.  
3. uses
4.   Interfaces, LCLIntf, LCLType, SysUtils, Classes, Graphics, DateUtils;
5.  
6. const
7.   cs = '0123456789';
8.  
9. var
10.   B: TBitmap;
11.   i: Integer;
12.   t: TDateTime;
13.   s: String = cs;
14.   dx: array [0..Length(cs)] of Integer;
15.  
16.  
17. begin
18.   B := TBitmap.Create;
19.   B.SetSize(100, 100);
20.  
21.   for i:=Low(dx) to High(dx) do
22.     dx[i] := 1;
23.  
24.   t:=Now;
25.   for i:=0 to 100*1000 do
26.     ExtTextOut(B.Canvas.Handle, 0, 0, 0, nil, PChar(s), Length(s), dx);
27.   WriteLn(MilliSecondsBetween(Now, t), 'ms');
28.  
29.   B.Free;
30.  
31.   {$IFDEF WINDOWS}ReadLn;{$ENDIF}
32. end.
33.  

[zeljko]

The AI is quite impressive! After studying the code for few seconds it already knows it better than I do.
Did somebody try the "Pango Caching Improvements" it suggested? How much they affect?
I guess a similar optimization could be used for LCL-GTK3. That would be even more important for the future.

It advised two times to use Qt5 and Qt6 , and I did not see any code for "Pango Caching Improvements"

[JuhaManninen]

It advised two times to use Qt5 and Qt6 ,

That is a good advice.

and I did not see any code for "Pango Caching Improvements"

There is a code snippet although I did not test it.

Code: Pascal  [Select][+][-]

1. // Add fields to device context or a global cache map
2. if not Assigned(CachedLayout) or (FontChanged) then
3. begin
4.   if Assigned(CachedLayout) then g_object_unref(CachedLayout);
5.   CachedLayout := pango_layout_new(PangoContext);
6.   CachedFontDesc := pango_font_description_copy(CurrentFontDesc);
7.   pango_layout_set_font_description(CachedLayout, CachedFontDesc);
8. end;
9. pango_layout_set_text(CachedLayout, PChar(UTF8Str), Count);
10. // Then use for extents and rendering

[zeljko]

That does not smell good, but let someone provide patch which prooves massive improvement.