Image::Pngblog

Position with 0.30

Filed under: Uncategorized — adrian @ 2:46 pm

Total Elapsed Time = 34.13125 Seconds
User+System Time = 23.84125 Seconds
Exclusive Times
%Time ExclSec CumulS #Calls sec/call Csec/c Name
25.4   6.069 6.069    638   0.0095 0.0095 Image::Pngslimmer::unfilterpaeth
25.1   6.000 6.000    255   0.0235 0.0235 Image::Pngslimmer::sortonaxes
15.0   3.579 4.568    511   0.0070 0.0089 Image::Pngslimmer::generate_box
8.43   2.010 2.010      2   1.0050 1.0050 Image::Pngslimmer::countcolours
7.54   1.797 12.990      1   1.7973 12.990 Image::Pngslimmer::index_mediancut
5.79   1.380 18.897      1   1.3800 18.897 Image::Pngslimmer::palettize
4.48   1.068 0.916    134   0.0080 0.0068 Image::Pngslimmer::unfilterave
4.31   1.028 1.028 127439   0.0000 0.0000 Image::Pngslimmer::convert_toxyz
1.30   0.309 0.347    256   0.0012 0.0014 Image::Pngslimmer::getcolour_ave
1.09   0.260 0.607      1   0.2596 0.6071 Image::Pngslimmer::getpalette
0.08   0.020 0.020      2   0.0100 0.0100 Compress::Raw::Zlib::inflateStream
::inflate
0.08   0.020 0.020      8   0.0025 0.0025 Image::Pngslimmer::unfiltersub
0.08   0.019 7.024      2   0.0094 3.5120 Image::Pngslimmer::unfilter
0.04   0.010 0.010      1   0.0100 0.0100 Compress::Raw::Zlib::deflateStream
::deflate
0.04   0.010 0.010      3   0.0033 0.0033 vars::BEGIN

Comments Off

Improving the code

Filed under: Uncategorized — adrian @ 11:02 pm

Hopefully I’ve done that, (for instance with alpha channels now handled for the first time) a new version will go to CPAN shortly. In the meantime, for your delight, here is the output for processing a very big file (350K) with the new code
adrian@transpero:~/Image-Pngslimmer$ dprofpp
Total Elapsed Time = 57.98357 Seconds
User+System Time = 24.97357 Seconds
Exclusive Times
%Time ExclSec CumulS #Calls sec/call Csec/c Name
29.0   7.259 7.259    638   0.0114 0.0114 Image::Pngslimmer::unfilterpaeth
23.9   5.990 5.990    255   0.0235 0.0235 Image::Pngslimmer::sortonaxes
15.4   3.865 4.660    511   0.0076 0.0091 Image::Pngslimmer::generate_box
7.97   1.990 1.990      2   0.9950 0.9950 Image::Pngslimmer::countcolours
6.72   1.677 13.013      1   1.6773 13.012 Image::Pngslimmer::index_mediancut
5.49   1.370 19.679      1   1.3700 19.679 Image::Pngslimmer::palettize
5.40   1.348 1.236    134   0.0101 0.0092 Image::Pngslimmer::unfilterave
3.70   0.924 0.924 127694   0.0000 0.0000 Image::Pngslimmer::convert_toxyz
1.20   0.300 0.537      1   0.2996 0.5371 Image::Pngslimmer::getpalette
0.56   0.140 0.140    254   0.0005 0.0005 Image::Pngslimmer::getbiggestbox
0.44   0.109 0.237    256   0.0004 0.0009 Image::Pngslimmer::getcolour_ave
0.16   0.040 0.040      8   0.0050 0.0050 Image::Pngslimmer::unfiltersub
0.08   0.020 0.020      2   0.0100 0.0100 Compress::Raw::Zlib::inflateStream
::inflate
0.08   0.020 0.020     16   0.0012 0.0012 Exporter::export
0.08   0.019 8.554      2   0.0094 4.2770 Image::Pngslimmer::unfilter

And here is the same with 0.25

adrian@transpero:~/Image-Pngslimmer$ dprofpp
Total Elapsed Time = 40.79854 Seconds
User+System Time = 27.58854 Seconds
Exclusive Times
%Time ExclSec CumulS #Calls sec/call Csec/c Name
28.9   7.979 7.979    638   0.0125 0.0125 Image::Pngslimmer::unfilterpaeth
25.5   7.040 7.040    255   0.0276 0.0276 Image::Pngslimmer::sortonaxes
14.1   3.907 4.745    511   0.0076 0.0093 Image::Pngslimmer::generate_box
7.29   2.010 2.010      2   1.0050 1.0050 Image::Pngslimmer::countcolours
6.69   1.847 14.348      1   1.8473 14.347 Image::Pngslimmer::index_mediancut
4.97   1.370 21.414      1   1.3700 21.414 Image::Pngslimmer::palettize
4.60   1.268 1.226    134   0.0095 0.0092 Image::Pngslimmer::unfilterave
3.43   0.947 0.947 127529   0.0000 0.0000 Image::Pngslimmer::convert_toxyz
1.96   0.540 6.147      1   0.5400 6.1469 Image::Pngslimmer::reportcolours
1.27   0.350 0.657      1   0.3496 0.6571 Image::Pngslimmer::getpalette
0.72   0.199 0.307    256   0.0008 0.0012 Image::Pngslimmer::getcolour_ave
0.18   0.050 0.050      8   0.0062 0.0062 Image::Pngslimmer::unfiltersub
0.18   0.050 0.050    254   0.0002 0.0002 Image::Pngslimmer::getbiggestbox
0.07   0.020 0.020      6   0.0033 0.0033 DynaLoader::dl_load_file
0.07   0.020 0.030      2   0.0100 0.0149 Image::Pngslimmer::getuncompressed
_data

NB The difference between total elapsed times seems to be an artefact of screen handling rather than a real measure of performance - it is the total User+System time that is measuring the relative performance of the two lumps of code.

Comments Off

Inlining probably not the answer

Filed under: Uncategorized — adrian @ 3:36 pm

Continuing to search for ways to speed up the code I have tried inlining some “C”, to the small but frequently called convert_toxyz. In pure perl this is as below:

sub convert_toxyz {
        #convert 24 bit number to cartesian point
        my $inpoint = shift;
        return ($inpoint>>16, ($inpoint & 0xFF00)>>8, $inpoint & 0xFF);
}

But now I have replaced it with this C/perl combination:

use Inline C => <<'END_C';

void convert_tocartesian(int bignumb)
{
Inline_Stack_Vars;
Inline_Stack_Reset;

         Inline_Stack_Push(sv_2mortal(newSViv(bignumb>>16)));
         Inline_Stack_Push(sv_2mortal(newSViv((bignumb & 0xFF00) >> ));
         Inline_Stack_Push(sv_2mortal(newSViv(bignumb & 0xFF)));
         Inline_Stack_Done;
}
END_C
sub convert_toxyz {
         #convert 24 bit number to cartesian point
         my $inpoint = shift;
         #use Inline C to speed bit manipulation
         return convert_tocartesian($inpoint);
}

Initailly, when the C code was being built for the first time the performance was porcine:

> perl -d:DProf ./profiletest.pl
> dprofpp tmon.out
Total Elapsed Time = 1.002253 Seconds
User+System Time = 0.352253 Seconds
Exclusive Times
%Time ExclSec CumulS #Calls sec/call Csec/c Name
25.2   0.089 0.128      1   0.0894 0.1283 Parse::RecDescent::_generate
8.52   0.030 0.032    511   0.0001 0.0001 Image::Pngslimmer::generate_box
8.52   0.030 0.030    254   0.0001 0.0001 Image::Pngslimmer::getbiggestbox
5.68   0.020 0.020    424   0.0000 0.0000 Parse::RecDescent::Rule::_contains
5.11   0.018 0.032    565   0.0000 0.0001 Text::Balanced::_match_codeblock
4.83   0.017 0.034    464   0.0000 0.0001 Text::Balanced::_match_variable
2.84   0.010 0.010      1   0.0100 0.0100 Image::Pngslimmer::countcolours
2.84   0.010 0.010      6   0.0017 0.0017 Cwd::abs_path
2.84   0.010 0.010      3   0.0033 0.0033 Inline::C::system_call
2.84   0.010 0.148      1   0.0100 0.1482 Inline::C::ParseRecDescent::get_pa
                                             rser
2.84   0.010 0.010      7   0.0014 0.0014 IO::Handle::BEGIN
2.84   0.010 0.010      7   0.0014 0.0014 Inline::C::BEGIN
2.84   0.010 0.010      8   0.0012 0.0012 Parse::RecDescent::BEGIN
2.84   0.010 0.220      1   0.0100 0.2197 Inline::glue
2.84   0.010 0.010     31   0.0003 0.0003 Image::Pngslimmer::unfilterpaeth

And subsequently, although faster, it seems slower than the pure perl - presumably because the overhead is so great:

> perl -d:DProf ./profiletest.pl
> dprofpp tmon.out
Total Elapsed Time = 0.192022 Seconds
User+System Time = 0.212022 Seconds
Exclusive Times
%Time ExclSec CumulS #Calls sec/call Csec/c Name
13.6   0.029 0.029    507   0.0001 0.0001 Image::Pngslimmer::getRGBbox
9.43   0.020 0.020 10148   0.0000 0.0000 Image::Pngslimmer::convert_tocarte
                                             sian
9.43   0.020 0.020    254   0.0001 0.0001 Image::Pngslimmer::getbiggestbox
8.96   0.019 0.016    256   0.0001 0.0001 Image::Pngslimmer::getcolour_ave
4.72   0.010 0.010      1   0.0100 0.0100 Image::Pngslimmer::unfiltersub
4.72   0.010 0.010      6   0.0017 0.0017 AutoLoader::import
4.72   0.010 0.010     14   0.0007 0.0007 Exporter::export
4.72   0.010 0.010      5   0.0020 0.0020 File::Spec::Unix::catfile
4.72   0.010 0.010      5   0.0020 0.0020 File::GlobMapper::BEGIN
4.72   0.010 0.010     31   0.0003 0.0003 Image::Pngslimmer::unfilterpaeth
4.72   0.010 0.010     30   0.0003 0.0003 IO::Compress::Gzip::Constants::BEG
                                             IN
4.72   0.010 0.020     67   0.0001 0.0003 Exporter::import
4.72   0.010 0.049     18   0.0005 0.0027 Compress::Zlib::BEGIN
4.72   0.010 0.010    255   0.0000 0.0000 Image::Pngslimmer::sortonaxes
4.72   0.010 0.010    255   0.0000 0.0000 POSIX::floor
To test properly I have to slightly reconfigure the apache server, so cannot do that immediately. Update to follow.

Comments Off

Back to this

Filed under: Uncategorized — adrian @ 11:30 pm

After over a year I am back to hacking at this code. I’ve tried a couple of would be optimizations - such as storing all the converted cartesian points in a hash - but while they seem to speed things up on test, they fail in the real world - where the task to serve the “scores on the doors” page of HateMyTory in something less than 23 seconds. Needs more work!

But I will probably post a new version - with whitespace etc tidied in due course regardless.

Comments (0)

Current state of play

Filed under: Uncategorized — adrian @ 10:25 am

Total Elapsed Time = 0.216515 Seconds
User Time = 0.166515 Seconds
Exclusive Times
%Time ExclSec CumulS #Calls sec/call Csec/c Name
18.0   0.030 0.030    255   0.0001 0.0001 Image::Pngslimmer::sortonaxes
18.0   0.030 0.030    507   0.0001 0.0001 Image::Pngslimmer::getRGBbox
15.0   0.025 0.025 10152   0.0000 0.0000 Image::Pngslimmer::convert_toxyz
12.0   0.020 0.020     31   0.0006 0.0006 Image::Pngslimmer::unfilterpaeth
12.0   0.020 0.049     18   0.0011 0.0027 Compress::Zlib::BEGIN
11.4   0.019 0.117      1   0.0191 0.1173 Image::Pngslimmer::index_mediancut
6.01   0.010 0.010      7   0.0014 0.0014 IO::Uncompress::RawInflate::BEGIN
6.01   0.010 0.020      9   0.0011 0.0022 IO::Compress::Gzip::BEGIN
6.01   0.010 -0.000     60   0.0002      - Exporter::import
6.01   0.010 0.010    254   0.0000 0.0000 Image::Pngslimmer::getbiggestbox

Comments Off

References do speed things up after all

Filed under: Uncategorized — adrian @ 12:44 pm

Even with the corrected code the references work faster - each call is about 2/10,000ths of a second faster for the test PNG. Not much in itself, but as the function is being called 254 times, it does make a difference.

Comments (0)

I am an idiot

Filed under: Uncategorized — adrian @ 8:11 pm

Of course, the code below is broken as it searches through an array with zero elements. Updated code etc to follow…

Comments (0)

References do speed things up

Filed under: Uncategorized — adrian @ 6:54 pm

I had tried a few times to speed things up by making better use of references, but before now, haven’t really managed it… but changing the code from this:

sub getbiggestbox {
#return the index to the biggest box
my ($boxesin, @boxesin, $i, $n);
$boxesin = shift;
@boxesin = @$boxesin;
$n = scalar(@boxesin)/4; #each box has 4 items stored about it
my $z = 0;
my $counter = 0;
my $biggest = 0;
for ($i = 0; $i < $n; $i++)
{
#length is 4th item per box
$counter = $i * 4 + 3;
if ($boxesin[$counter] > $z)
{
$z = $boxesin[$counter];
$biggest = $i;
}
}
return $biggest;
}

To this:

sub getbiggestbox {
#return the index to the biggest box
my ($boxesin, @boxesin, $i, $n);
$boxesin = shift;
$n = scalar(@boxesin)/4; #each box has 4 items stored about it
my $z = 0;
my $counter = 0;
my $biggest = 0;
for ($i = 0; $i < $n; $i++)
{
#length is 4th item per

Image::Pngblog

7 September, 2008

Position with 0.30

19 August, 2008

Improving the code

25 July, 2008

Inlining probably not the answer

20 July, 2008

Back to this

4 April, 2007

Current state of play

3 April, 2007

References do speed things up after all

1 April, 2007

I am an idiot

31 March, 2007

References do speed things up