Performance problems with "photo copy -to"

[Paul Obermeier]

I have a simple tiling procedure and noticed a slowdown when using large images,
i.e. when doing lots of copies.

I noticed the following behaviours (both on Windows and Linux):
1. The time for copying does not grow linearly, as expected, but gets larger the more copies are involved.
2. The copy time of Tcl/Tk 9 is nearly twice the time as when using Tcl/Tk 8.6.15 or the combination
Tcl 8.7/Tk 9. So this performance degradation does not seem to stem from Tk9, but from Tcl9. Strange!

Can anybody confirm my measurements?

Thanks,
Paul


Using Tcl 8.6.15, Tk 8.6.15
Time for 100 copies: 0.69 seconds ( 6 milliseconds / copy)
Time for 400 copies: 10.26 seconds ( 25 milliseconds / copy)
Time for 900 copies: 51.12 seconds ( 56 milliseconds / copy)

Using Tcl 8.7b1, Tk 9.0.0
Time for 100 copies: 0.70 seconds ( 7 milliseconds / copy)
Time for 400 copies: 10.22 seconds ( 25 milliseconds / copy)
Time for 900 copies: 51.35 seconds ( 57 milliseconds / copy)

Using Tcl 9.0.0, Tk 9.0.0
Time for 100 copies: 1.28 seconds ( 12 milliseconds / copy)
Time for 400 copies: 19.45 seconds ( 48 milliseconds / copy)
Time for 900 copies: 98.16 seconds (109 milliseconds / copy)


package require Tk

proc Tile { phImg xRepeat yRepeat } {
set startTime [clock milliseconds]
set w [image width $phImg]
set h [image height $phImg]
set w2 [expr {$w * $xRepeat}]
set h2 [expr {$h * $yRepeat}]

set tileImg [image create photo -width $w2 -height $h2]

for { set x 0 } { $x < $xRepeat } { incr x } {
for { set y 0 } { $y < $yRepeat } { incr y } {
$tileImg copy $phImg -to [expr {$x*$w}] [expr {$y*$h}]
}
}
set endTime [clock milliseconds]
puts [format "Time for %4d copies: %5.2f seconds (%3d milliseconds / copy)" \
[expr {$xRepeat * $yRepeat}] \
[expr { ($endTime - $startTime) / 1000.0 }] \
[expr { ($endTime - $startTime) / ($xRepeat * $yRepeat) }]]
return $tileImg
}

set srcImg [image create photo -width 500 -height 500]

puts "Using Tcl [info patch], Tk [package version Tk]"
Tile $srcImg 10 10
Tile $srcImg 20 20
Tile $srcImg 30 30

exit

[elns]

Post by Paul Obermeier
Can anybody confirm my measurements?

Using Tcl 8.6.15, Tk 8.6.15
Time for 100 copies: 1.31 seconds ( 13 milliseconds / copy)
Time for 400 copies: 20.32 seconds ( 50 milliseconds / copy)
Time for 900 copies: 102.27 seconds (113 milliseconds / copy)

Using Tcl 9.0.0, Tk 9.0.0
Time for 100 copies: 1.17 seconds ( 11 milliseconds / copy)
Time for 400 copies: 16.84 seconds ( 42 milliseconds / copy)
Time for 900 copies: 84.76 seconds ( 94 milliseconds / copy)

So, I see a performance increase with Tcl/Tk 9.0.0 instead of a decrease.

B.t.w.:
- Having replaced in your script [package version Tk] with $tk_patchLevel
to prevent that multiple installed Tk versions are reported.
- The combination Tcl8.7 and Tk9 is not avaiable to me
- Did you perhaps use different compiler optimizations for your builds?
(My builds are all without compiler optimizations.)


Regards,
Erik Leunissen

[elns]

Post by Paul Obermeier
Using Tcl 8.6.15, Tk 8.6.15
Time for  100 copies:  1.31 seconds ( 13 milliseconds / copy)
Time for  400 copies: 20.32 seconds ( 50 milliseconds / copy)
Time for  900 copies: 102.27 seconds (113 milliseconds / copy)
Using Tcl 9.0.0, Tk 9.0.0
Time for  100 copies:  1.17 seconds ( 11 milliseconds / copy)
Time for  400 copies: 16.84 seconds ( 42 milliseconds / copy)
Time for  900 copies: 84.76 seconds ( 94 milliseconds / copy)

This was executed on Linux.

Post by Paul Obermeier
So, I see a performance increase with Tcl/Tk 9.0.0 instead of a decrease.
- Having replaced in your script [package version Tk] with $tk_patchLevel
  to prevent that multiple installed Tk versions are reported.
- The combination Tcl8.7 and Tk9 is not avaiable to me
- Did you perhaps use different compiler optimizations for your builds?
  (My builds are all without compiler optimizations.)
Regards,
Erik Leunissen

[greg]

Post by Paul Obermeier
I have a simple tiling procedure and noticed a slowdown when using large images,
i.e. when doing lots of copies.
1. The time for copying does not grow linearly, as expected, but gets
larger the more copies are involved.
2. The copy time of Tcl/Tk 9 is nearly twice the time as when using Tcl/
Tk 8.6.15 or the combination
   Tcl 8.7/Tk 9. So this performance degradation does not seem to stem
from Tk9, but from Tcl9. Strange!
Can anybody confirm my measurements?
Thanks,
Paul
Using Tcl 8.6.15, Tk 8.6.15
Time for  100 copies:  0.69 seconds (  6 milliseconds / copy)
Time for  400 copies: 10.26 seconds ( 25 milliseconds / copy)
Time for  900 copies: 51.12 seconds ( 56 milliseconds / copy)
Using Tcl 8.7b1, Tk 9.0.0
Time for  100 copies:  0.70 seconds (  7 milliseconds / copy)
Time for  400 copies: 10.22 seconds ( 25 milliseconds / copy)
Time for  900 copies: 51.35 seconds ( 57 milliseconds / copy)
Using Tcl 9.0.0, Tk 9.0.0
Time for  100 copies:  1.28 seconds ( 12 milliseconds / copy)
Time for  400 copies: 19.45 seconds ( 48 milliseconds / copy)
Time for  900 copies: 98.16 seconds (109 milliseconds / copy)
package require Tk
proc Tile { phImg xRepeat yRepeat } {
    set startTime [clock milliseconds]
    set w [image width  $phImg]
    set h [image height $phImg]
    set w2 [expr {$w * $xRepeat}]
    set h2 [expr {$h * $yRepeat}]
    set tileImg [image create photo -width $w2 -height $h2]
    for { set x 0 } { $x < $xRepeat } { incr x } {
        for { set y 0 } { $y < $yRepeat } { incr y } {
            $tileImg copy $phImg -to [expr {$x*$w}] [expr {$y*$h}]
        }
    }
    set endTime [clock milliseconds]
    puts [format "Time for %4d copies: %5.2f seconds (%3d
milliseconds / copy)" \
         [expr {$xRepeat * $yRepeat}] \
         [expr { ($endTime - $startTime) / 1000.0 }] \
         [expr { ($endTime - $startTime) / ($xRepeat * $yRepeat) }]]
    return $tileImg
}
set srcImg [image create photo -width 500 -height 500]
puts "Using Tcl [info patch], Tk [package version Tk]"
Tile $srcImg 10 10
Tile $srcImg 20 20
Tile $srcImg 30 30
exit

Hello,


Using Tcl 8.6.15, Tk 8.6.15
Time for 100 copies: 1.52 seconds ( 15 milliseconds / copy)
Time for 400 copies: 23.17 seconds ( 57 milliseconds / copy)
Time for 900 copies: 117.39 seconds (130 milliseconds / copy)


Using Tcl 9.0.0, Tk 9.0.0
Time for 100 copies: 1.50 seconds ( 14 milliseconds / copy)
Time for 400 copies: 23.18 seconds ( 57 milliseconds / copy)
Time for 900 copies: 117.57 seconds (130 milliseconds / copy)


both Tcl/Tk from debian sid


Gregor

[greg]

Post by greg

Post by Paul Obermeier
I have a simple tiling procedure and noticed a slowdown when using

Hello,
Using Tcl 8.6.15, Tk 8.6.15
Time for  100 copies:  1.52 seconds ( 15 milliseconds / copy)
Time for  400 copies: 23.17 seconds ( 57 milliseconds / copy)
Time for  900 copies: 117.39 seconds (130 milliseconds / copy)
Using Tcl 9.0.0, Tk 9.0.0
Time for  100 copies:  1.50 seconds ( 14 milliseconds / copy)
Time for  400 copies: 23.18 seconds ( 57 milliseconds / copy)
Time for  900 copies: 117.57 seconds (130 milliseconds / copy)
both Tcl/Tk from debian sid
Gregor

Windows 10 with the last Bawt (KVM)
Using Tcl 8.6.13, Tk 8.6.13
Time for 100 copies: 3.15 seconds ( 31 milliseconds / copy)
Time for 400 copies: 42.10 seconds (105 milliseconds / copy)
Time for 900 copies: 201.18 seconds (223 milliseconds / copy)


Windows 10 with the last Magicsplat (KVM)
Using Tcl 9.0.0, Tk 9.0.0
Time for 100 copies: 2.71 seconds ( 27 milliseconds / copy)
Time for 400 copies: 39.80 seconds ( 99 milliseconds / copy)
Time for 900 copies: 184.96 seconds (205 milliseconds / copy)

[saito]

Post by Paul Obermeier
I have a simple tiling procedure and noticed a slowdown when using large images,
i.e. when doing lots of copies.
1. The time for copying does not grow linearly, as expected, but gets
larger the more copies are involved.
2. The copy time of Tcl/Tk 9 is nearly twice the time as when using Tcl/
Tk 8.6.15 or the combination
   Tcl 8.7/Tk 9. So this performance degradation does not seem to stem
from Tk9, but from Tcl9. Strange!
Can anybody confirm my measurements?

I wonder if it has something to do with memory management for images.

On Windows Tcl/Tk 8.6.12, I first got results very close to your Tcl/Tk
9.0.0 versions. I ran it again and it got worse. It was embarrassingly
slow so I won't disclose it here. Let's say the numbers almost tripled,
and it worsened with the size of the images.


But then it errored out with a funny message :-)

Post by Paul Obermeier
not enough free memory for image buffer

The code was not deleting the images so it was running out of memory.

So I changed your code slightly and got these results:

% image delete [Tile $srcImg 10 10]
Time for 100 copies: 1.24 seconds ( 12 milliseconds / copy)

% image delete [Tile $srcImg 20 20]
Time for 400 copies: 19.01 seconds ( 47 milliseconds / copy)

...

[Paul Obermeier]

After some more tests it looks like the source of the problem with issue 1 is the handling of the alpha channel.

I created 4 test images and rewrote the test script.

One image contains just 3 channels (RGB) and three images contain 4 channels (RGBA),
where the alpha channel is either fully opaque, fully transparent or partial transparent.
Only the partial transparent image behaves as expected, i.e. has linear time increase.

Using Tcl 9.0.0, Tk 9.0.0
Using file Balls-3chan.png
Time for 100 copies: 1.36 seconds ( 13 milliseconds / copy)
Time for 400 copies: 20.14 seconds ( 50 milliseconds / copy)
Using file Balls-4chan-transp.png
Time for 100 copies: 1.27 seconds ( 12 milliseconds / copy)
Time for 400 copies: 19.82 seconds ( 49 milliseconds / copy)
Using file Balls-4chan-opaque.png
Time for 100 copies: 1.36 seconds ( 13 milliseconds / copy)
Time for 400 copies: 20.18 seconds ( 50 milliseconds / copy)
Using file Balls-4chan-partial.png
Time for 100 copies: 0.13 seconds ( 1 milliseconds / copy)
Time for 400 copies: 0.56 seconds ( 1 milliseconds / copy)

Find the test images and the script at https://www.tcl3d.org/download/PhotoCopyBug.zip
If someone again can confirm the above behaviour, I would create a Tk ticket.

Regarding issue 2 I get totally different results on various systems:
Fedora 40 : No time differences between Tcl/Tk versions.
Suse 15.6 : No time differences between Tcl/Tk versions.
Ubuntu 24 : Tcl/Tk9 twice as fast as the other combinations.
Debian 12.6: Tcl/Tk9 twice as slow as the other combinations.
Windows 11 : Tcl/Tk8.6.15 twice as fast as the other combinations.

Paul

Post by Paul Obermeier
I have a simple tiling procedure and noticed a slowdown when using large images,
i.e. when doing lots of copies.
1. The time for copying does not grow linearly, as expected, but gets larger the more copies are involved.
2. The copy time of Tcl/Tk 9 is nearly twice the time as when using Tcl/Tk 8.6.15 or the combination
   Tcl 8.7/Tk 9. So this performance degradation does not seem to stem from Tk9, but from Tcl9. Strange!
Can anybody confirm my measurements?
Thanks,
Paul
Using Tcl 8.6.15, Tk 8.6.15
Time for  100 copies:  0.69 seconds (  6 milliseconds / copy)
Time for  400 copies: 10.26 seconds ( 25 milliseconds / copy)
Time for  900 copies: 51.12 seconds ( 56 milliseconds / copy)
Using Tcl 8.7b1, Tk 9.0.0
Time for  100 copies:  0.70 seconds (  7 milliseconds / copy)
Time for  400 copies: 10.22 seconds ( 25 milliseconds / copy)
Time for  900 copies: 51.35 seconds ( 57 milliseconds / copy)
Using Tcl 9.0.0, Tk 9.0.0
Time for  100 copies:  1.28 seconds ( 12 milliseconds / copy)
Time for  400 copies: 19.45 seconds ( 48 milliseconds / copy)
Time for  900 copies: 98.16 seconds (109 milliseconds / copy)
package require Tk
proc Tile { phImg xRepeat yRepeat } {
    set startTime [clock milliseconds]
    set w [image width  $phImg]
    set h [image height $phImg]
    set w2 [expr {$w * $xRepeat}]
    set h2 [expr {$h * $yRepeat}]
    set tileImg [image create photo -width $w2 -height $h2]
    for { set x 0 } { $x < $xRepeat } { incr x } {
        for { set y 0 } { $y < $yRepeat } { incr y } {
            $tileImg copy $phImg -to [expr {$x*$w}] [expr {$y*$h}]
        }
    }
    set endTime [clock milliseconds]
    puts [format "Time for %4d copies: %5.2f seconds (%3d milliseconds / copy)" \
         [expr {$xRepeat * $yRepeat}] \
         [expr { ($endTime - $startTime) / 1000.0 }] \
         [expr { ($endTime - $startTime) / ($xRepeat * $yRepeat) }]]
    return $tileImg
}
set srcImg [image create photo -width 500 -height 500]
puts "Using Tcl [info patch], Tk [package version Tk]"
Tile $srcImg 10 10
Tile $srcImg 20 20
Tile $srcImg 30 30
exit

[greg]

Post by Paul Obermeier
After some more tests it looks like the source of the problem with issue
1 is the handling of the alpha channel.

...
Hello,


photocopy.tcl


Debian Sid, Debian 6.11.6-1

Using Tcl 8.6.15, Tk 8.6.15
Using file Balls-3chan.png
Time for 100 copies: 1.52 seconds ( 15 milliseconds / copy)
Time for 400 copies: 23.30 seconds ( 58 milliseconds / copy)
Using file Balls-4chan-transp.png
Time for 100 copies: 1.51 seconds ( 15 milliseconds / copy)
Time for 400 copies: 23.27 seconds ( 58 milliseconds / copy)
Using file Balls-4chan-opaque.png
Time for 100 copies: 1.52 seconds ( 15 milliseconds / copy)
Time for 400 copies: 23.35 seconds ( 58 milliseconds / copy)
Using file Balls-4chan-partial.png
Time for 100 copies: 0.09 seconds ( 0 milliseconds / copy)
Time for 400 copies: 0.34 seconds ( 0 milliseconds / copy)

Using Tcl 9.0.0, Tk 9.0.0
Using file Balls-3chan.png
Time for 100 copies: 1.52 seconds ( 15 milliseconds / copy)
Time for 400 copies: 23.38 seconds ( 58 milliseconds / copy)
Using file Balls-4chan-transp.png
Time for 100 copies: 1.51 seconds ( 15 milliseconds / copy)
Time for 400 copies: 23.43 seconds ( 58 milliseconds / copy)
Using file Balls-4chan-opaque.png
Time for 100 copies: 1.52 seconds ( 15 milliseconds / copy)
Time for 400 copies: 23.39 seconds ( 58 milliseconds / copy)
Using file Balls-4chan-partial.png
Time for 100 copies: 0.08 seconds ( 0 milliseconds / copy)
Time for 400 copies: 0.35 seconds ( 0 milliseconds / copy)


Windows 10 (KVM)

Bawt Tcl/Tk

Using Tcl 8.6.13, Tk 8.6.13
Using file Balls-3chan.png
Time for 100 copies: 2.06 seconds ( 20 milliseconds / copy)
Time for 400 copies: 35.67 seconds ( 89 milliseconds / copy)
Using file Balls-4chan-transp.png
Time for 100 copies: 2.11 seconds ( 21 milliseconds / copy)
Time for 400 copies: 33.44 seconds ( 83 milliseconds / copy)
Using file Balls-4chan-opaque.png
Time for 100 copies: 2.38 seconds ( 23 milliseconds / copy)
Time for 400 copies: 32.56 seconds ( 81 milliseconds / copy)
Using file Balls-4chan-partial.png
Time for 100 copies: 0.26 seconds ( 2 milliseconds / copy)
Time for 400 copies: 1.15 seconds ( 2 milliseconds / copy)


Magicsplat Tcl/Tk

Using Tcl 9.0.0, Tk 9.0.0
Using file Balls-3chan.png
Time for 100 copies: 1.97 seconds ( 19 milliseconds / copy)
Time for 400 copies: 28.32 seconds ( 70 milliseconds / copy)
Using file Balls-4chan-transp.png
Time for 100 copies: 1.78 seconds ( 17 milliseconds / copy)
Time for 400 copies: 27.21 seconds ( 68 milliseconds / copy)
Using file Balls-4chan-opaque.png
Time for 100 copies: 1.97 seconds ( 19 milliseconds / copy)
Time for 400 copies: 28.12 seconds ( 70 milliseconds / copy)
Using file Balls-4chan-partial.png
Time for 100 copies: 0.28 seconds ( 2 milliseconds / copy)
Time for 400 copies: 1.09 seconds ( 2 milliseconds / copy)



Gregor

[Harald Oehlmann]

The horrible performence of images with an alpha channel on WIndows and
Mac-OS is a long standing bug and subject of many tickets....

It would be great to attack this issue...

Thanks,
Harald

Post by Paul Obermeier
After some more tests it looks like the source of the problem with issue
1 is the handling of the alpha channel.
I created 4 test images and rewrote the test script.
One image contains just 3 channels (RGB) and three images contain 4 channels (RGBA),
where the alpha channel is either fully opaque, fully transparent or partial transparent.
Only the partial transparent image behaves as expected, i.e. has linear time increase.
Using Tcl 9.0.0, Tk 9.0.0
Using file Balls-3chan.png
Time for  100 copies:  1.36 seconds ( 13 milliseconds / copy)
Time for  400 copies: 20.14 seconds ( 50 milliseconds / copy)
Using file Balls-4chan-transp.png
Time for  100 copies:  1.27 seconds ( 12 milliseconds / copy)
Time for  400 copies: 19.82 seconds ( 49 milliseconds / copy)
Using file Balls-4chan-opaque.png
Time for  100 copies:  1.36 seconds ( 13 milliseconds / copy)
Time for  400 copies: 20.18 seconds ( 50 milliseconds / copy)
Using file Balls-4chan-partial.png
Time for  100 copies:  0.13 seconds (  1 milliseconds / copy)
Time for  400 copies:  0.56 seconds (  1 milliseconds / copy)
Find the test images and the script at https://www.tcl3d.org/download/
PhotoCopyBug.zip
If someone again can confirm the above behaviour, I would create a Tk ticket.
Fedora 40  : No time differences between Tcl/Tk versions.
Suse 15.6  : No time differences between Tcl/Tk versions.
Ubuntu 24  : Tcl/Tk9 twice as fast as the other combinations.
Debian 12.6: Tcl/Tk9 twice as slow as the other combinations.
Windows 11 : Tcl/Tk8.6.15 twice as fast as the other combinations.
Paul

Post by Paul Obermeier
I have a simple tiling procedure and noticed a slowdown when using large images,
i.e. when doing lots of copies.
1. The time for copying does not grow linearly, as expected, but gets
larger the more copies are involved.
2. The copy time of Tcl/Tk 9 is nearly twice the time as when using
Tcl/Tk 8.6.15 or the combination
    Tcl 8.7/Tk 9. So this performance degradation does not seem to
stem from Tk9, but from Tcl9. Strange!
Can anybody confirm my measurements?
Thanks,
Paul
Using Tcl 8.6.15, Tk 8.6.15
Time for  100 copies:  0.69 seconds (  6 milliseconds / copy)
Time for  400 copies: 10.26 seconds ( 25 milliseconds / copy)
Time for  900 copies: 51.12 seconds ( 56 milliseconds / copy)
Using Tcl 8.7b1, Tk 9.0.0
Time for  100 copies:  0.70 seconds (  7 milliseconds / copy)
Time for  400 copies: 10.22 seconds ( 25 milliseconds / copy)
Time for  900 copies: 51.35 seconds ( 57 milliseconds / copy)
Using Tcl 9.0.0, Tk 9.0.0
Time for  100 copies:  1.28 seconds ( 12 milliseconds / copy)
Time for  400 copies: 19.45 seconds ( 48 milliseconds / copy)
Time for  900 copies: 98.16 seconds (109 milliseconds / copy)
package require Tk
proc Tile { phImg xRepeat yRepeat } {
     set startTime [clock milliseconds]
     set w [image width  $phImg]
     set h [image height $phImg]
     set w2 [expr {$w * $xRepeat}]
     set h2 [expr {$h * $yRepeat}]
     set tileImg [image create photo -width $w2 -height $h2]
     for { set x 0 } { $x < $xRepeat } { incr x } {
         for { set y 0 } { $y < $yRepeat } { incr y } {
             $tileImg copy $phImg -to [expr {$x*$w}] [expr {$y*$h}]
         }
     }
     set endTime [clock milliseconds]
     puts [format "Time for %4d copies: %5.2f seconds (%3d
milliseconds / copy)" \
          [expr {$xRepeat * $yRepeat}] \
          [expr { ($endTime - $startTime) / 1000.0 }] \
          [expr { ($endTime - $startTime) / ($xRepeat * $yRepeat) }]]
     return $tileImg
}
set srcImg [image create photo -width 500 -height 500]
puts "Using Tcl [info patch], Tk [package version Tk]"
Tile $srcImg 10 10
Tile $srcImg 20 20
Tile $srcImg 30 30
exit