Monday, June 15, 2015

Perl Floating Point-Multiplication Benchmark

I was worried whether I was making basic errors in testing the Perl version, so I decided to use the Benchmark module to get the numbers. I copied the matmult.pl file and added use Benchmark ':all' to the header of the file. The main() routine got changed to :

    my $time = $ARGV[0] || 5;
    my %vars = (   2 => [],
                   5 => [],
                  10 => [],
                  32 => [],
                 100 => [],
    );
    for my $size ( keys %vars ) {
       my $filestub = q{F_} . $size . q{x} . $size . q{.};
       $vars{$size}[0] = readMatrix( $filestub . '1' );
       $vars{$size}[1] = readMatrix( $filestub . '2' );
    }
    
    say "Processing for $time seconds each size ",
        "will take @{[5 * $time]} seconds."; 
    say scalar localtime;
    cmpthese( -$time, {
       'F_2x2'     => sub { matmult( $vars{2}[0],
                                        $vars{2}[1]); },
       'F_5x5'     => sub { matmult( $vars{5}[0],
                                        $vars{5}[1]); },
       'F_10x10'   => sub { matmult( $vars{10}[0],  
                                        $vars{10}[1]); },
       'F_32x32'   => sub { matmult( $vars{32}[0],  
                                        $vars{32}[1]); },
       'F_100x100' => sub { matmult( $vars{100}[0], 
                                        $vars{100}[1]); },
       });
    say scalar localtime;


The user can specify how long each variant should be run, with a default of 5 seconds for each size if no arg is provided.

-➤   perl ./benchmark.pl 2
Processing for 2 seconds each size will take 10 seconds.
Mon Jun 15 17:11:08 2015
              Rate F_100x100   F_32x32   F_10x10     F_5x5     F_2x2
F_100x100   5.39/s        --      -97%     -100%     -100%     -100%
F_32x32      163/s     2931%        --      -97%     -100%     -100%
F_10x10     4909/s    90934%     2904%        --      -85%      -98%
F_5x5      32731/s   606904%    19929%      567%        --      -88%
F_2x2     264660/s  4908131%   161856%     5292%      709%        --
Mon Jun 15 17:11:22 2015

Adjusting for the number of multiplications involved in each size of matrix --- 8 for the 2x2 up to 1,000,000 for the 100x100 --- produces results very similar to what we've already seen ( millions of multiplications per second):

perl -E'say  "100 => 5.36";
        say  "32  => ", 32000 * 154 / 10**6;
        say  "20  => ", 1000 * 4904 / 10**6;
        say  "5   => ", 125 * 32767 / 10**6;
        say  "2   => ", 8 * 263472  / 10**6; 
       '

100 => 5.39
32  => 5.216
20  => 4.909
5   => 4.091375
2   => 2.11728

Rerunning the benchmarks with integer matrices and use integer enabled generated insignificantly, fractionally better numbers; though I wonder if the pragma affected Benchmark's calculations.

2 comments:

Joaquin Ferrero said...

Please, test PDL (Perl Data Language), after.

Tom Legrady said...

Hi Joaquin - Yes, I'm aware of PDL, but I've never used it. I'll see how hard it is.