xref: /aosp_15_r20/external/icu/icu4j/perf-tests/Dataset.pm (revision 0e209d3975ff4a8c132096b14b0e9364a753506e)

#/**
# * © 2016 and later: Unicode, Inc. and others.
# * License & terms of use: http://www.unicode.org/copyright.html
# *******************************************************************************
# * Copyright (C) 2002-2004, International Business Machines Corporation and    *
# * others. All Rights Reserved.                                                *
# *******************************************************************************
# */
package Dataset;
use Statistics::Descriptive;
use Statistics::Distributions;
use strict;

# Create a new Dataset with the given data.
sub new {
    my ($class) = shift;
    my $self = bless {
        _data => \@_,
        _scale => 1.0,
        _mean => 0.0,
        _error => 0.0,
    }, $class;

    my $n = @_;

    if ($n >= 1) {
        my $stats = Statistics::Descriptive::Full->new();
        $stats->add_data(@{$self->{_data}});
        $self->{_mean} = $stats->mean();

        if ($n >= 2) {
            # Use a t distribution rather than Gaussian because (a) we
            # assume an underlying normal dist, (b) we do not know the
            # standard deviation -- we estimate it from the data, and (c)
            # we MAY have a small sample size (also works for large n).
            my $t = Statistics::Distributions::tdistr($n-1, 0.005);
            $self->{_error} = $t * $stats->standard_deviation();
        }
    }

    $self;
}

# Set a scaling factor for all data; 1.0 means no scaling.
# Scale must be > 0.
sub setScale {
    my ($self, $scale) = @_;
    $self->{_scale} = $scale;
}

# Multiply the scaling factor by a value.
sub scaleBy {
    my ($self, $a) = @_;
    $self->{_scale} *= $a;
}

# Return the mean.
sub getMean {
    my $self = shift;
    return $self->{_mean} * $self->{_scale};
}

# Return a 99% error based on the t distribution.  The dataset
# is described as getMean() +/- getError().
sub getError {
    my $self = shift;
    return $self->{_error} * $self->{_scale};
}

# Divide two Datasets and return a new one, maintaining the
# mean+/-error.  The new Dataset has no data points.
sub divide {
    my $self = shift;
    my $rhs = shift;

    my $minratio = ($self->{_mean} - $self->{_error}) /
                   ($rhs->{_mean} + $rhs->{_error});
    my $maxratio = ($self->{_mean} + $self->{_error}) /
                   ($rhs->{_mean} - $rhs->{_error});

    my $result = Dataset->new();
    $result->{_mean} = ($minratio + $maxratio) / 2;
    $result->{_error} = $result->{_mean} - $minratio;
    $result->{_scale} = $self->{_scale} / $rhs->{_scale};
    $result;
}

1;