modelverse/performance/perf2tex.py

"""Converts performance data files (as produced by utils.py) to LaTeX charts."""

import argparse
import colorsys
import utils

# Generated LaTeX is based on the accepted answer to
# http://tex.stackexchange.com/questions/101320/grouped-bar-chart

# pylint: disable=I0011,W0141

COLOR_SCHEME_MAX_COLOR = (16, 59, 176)
COLOR_SCHEME_MIN_COLOR = (216, 33, 0)

DEFAULT_BAR_WIDTH = 14

LATEX_HEADER = r"""\documentclass[12pt,a4paper,onecolumn,openright]{report}
\usepackage[landscape]{geometry}
\usepackage{xcolor}
\usepackage{pgfplots}
\usepackage{tikz}
\usepgfplotslibrary{units}

% Define bar chart colors
%"""

LATEX_DOCUMENT_HEADER = r"""\begin{document}"""

TIKZ_PICTURE_HEADER = r"""\begin{tikzpicture}"""

TIKZ_PICTURE_FOOTER = r"""\end{tikzpicture}"""

LATEX_DOCUMENT_FOOTER = r"""\end{document}"""


def encode_latex_string(value):
    """Encodes the given string as a LaTeX string."""
    # I guess this is good enough for now. This may need to be
    # revisited if we encounter more complicated names.
    return '{%s}' % value.replace('_', '\\_')


def assemble_latex_chart(optimization_levels,
                         color_defs,
                         test_names,
                         data,
                         embed=False,
                         bar_width=DEFAULT_BAR_WIDTH):
    """Assembles a LaTeX chart from the given components."""
    lines = []
    if not embed:
        lines.append(LATEX_HEADER)
        for color_name, (red, green, blue) in color_defs:
            lines.append(r'\definecolor{%s}{HTML}{%02X%02X%02X}' %
                         (color_name, red, green, blue))
        lines.append(LATEX_DOCUMENT_HEADER)
    lines.append(TIKZ_PICTURE_HEADER)
    lines.append(r"""    \begin{axis}[
        width = 0.85*\textwidth,
        height = 8cm,
        major x tick style = transparent,
        ybar=2*\pgflinewidth,
        bar width=%dpt,
        ymajorgrids = true,
        ylabel = {Run time},
        symbolic x coords={%s},
        xtick = data,
        scaled y ticks = false,
        enlarge x limits=0.25,
        ymin=0,
        y unit=s,
        legend cell align=left,
        legend style={
                at={(1,1.05)},
                anchor=south east,
                column sep=1ex
        },
        nodes near coords,
        every node near coord/.append style={rotate=90, anchor=west}
    ]""" % (bar_width, ','.join(map(encode_latex_string, test_names))))
    for color_name, points in data:
        lines.append(r"""
        \addplot[style={%s,fill=%s,mark=none}]
            coordinates {%s};""" % (color_name, color_name, ' '.join(
            [('(%s,%s)' % (encode_latex_string(name), measurement))
             for name, measurement in points])))
    lines.append(r"""
        \legend{%s}""" %
                 ','.join(map(encode_latex_string, optimization_levels)))
    lines.append(r"""
    \end{axis}""")
    lines.append(TIKZ_PICTURE_FOOTER)
    if not embed:
        lines.append(LATEX_DOCUMENT_FOOTER)
    return '\n'.join(lines)


def flip_tuple_keys(tuple_list):
    index_map = {}
    results = []
    for top, top_data in tuple_list:
        for child, child_data in top_data:
            if child not in index_map:
                index_map[child] = len(results)
                results.append((child, []))

            _, child_list = results[index_map[child]]
            child_list.append((top, child_data))

    return results


def assemble_stacked_latex_chart(optimization_levels,
                                 quantity_colors,
                                 test_names,
                                 data,
                                 embed=False,
                                 bar_width=DEFAULT_BAR_WIDTH):
    """Assembles a stacked LaTeX chart from the given components."""
    lines = []
    quantity_color_names = {
        name: 'chartColor%d' % i
        for i, (name, _) in enumerate(quantity_colors)
    }
    quantity_names = [name for name, _ in quantity_colors]
    if not embed:
        lines.append(LATEX_HEADER)
        for quantity_name, (red, green, blue) in quantity_colors:
            lines.append(
                r'\definecolor{%s}{HTML}{%02X%02X%02X}' %
                (quantity_color_names[quantity_name], red, green, blue))
        lines.append(LATEX_DOCUMENT_HEADER)
    lines.append(TIKZ_PICTURE_HEADER)
    lines.append(r"""    \begin{axis}[
        ybar stacked,
        legend style={
            legend columns=%d,
            at={(xticklabel cs:0.5)},
            anchor=north,
            draw=none
        },
        width=0.85*\textwidth,
        height=8cm,
        xtick=data,
        bar width=%dpt,
        ymin=0,
        major x tick style=transparent,
        ymajorgrids=true,
        xticklabels={%s},
        x tick label style={rotate=90,anchor=east,font=\tiny},
        tick label style={font=\footnotesize},
        legend style={font=\footnotesize,yshift=-3ex},
        label style={font=\footnotesize},
        ylabel={Time},
        y unit=s,
        area legend
    ]""" % (len(optimization_levels), bar_width, ','.join(
        [encode_latex_string(name) for name in optimization_levels] *
        len(test_names))))
    for j, (quantity, points) in enumerate(data):
        contents = []
        i = 0
        for _, test_data in flip_tuple_keys(points):
            for _, measurement in test_data:
                contents.append('(%d,%s)' % (i, measurement))
                i += 1
            i += 1

        lines.append(r"""
        \addplot[style={%s,fill=%s,mark=none}%s]
            coordinates {%s};""" % (
            quantity_color_names[quantity], quantity_color_names[quantity],
            ",nodes near coords,"
            "every node near coord/.append style={rotate=90, anchor=west},color=black"
            if j == len(data) - 1 else "", ' '.join(contents)))
    lines.append(r"""
        \legend{%s}""" % ','.join(map(encode_latex_string, quantity_names)))
    lines.append(r"""
    \end{axis}""")
    lines.append(TIKZ_PICTURE_FOOTER)
    if not embed:
        lines.append(LATEX_DOCUMENT_FOOTER)
    return '\n'.join(lines)


def create_latex_chart(perf_data,
                       sorted_opt_levels=None,
                       embed=False,
                       bar_width=DEFAULT_BAR_WIDTH):
    """Creates a LaTeX chart for the given performance data."""
    if sorted_opt_levels is None:
        sorted_opt_levels = sort_by_runtime(perf_data)

    color_scheme = generate_color_scheme(sorted_opt_levels)
    opt_levels = []
    color_defs = []
    test_names = []
    data = []
    for i, optimization_level in enumerate(sorted_opt_levels):
        measurements = perf_data[optimization_level]
        color = color_scheme[optimization_level]
        color_name = 'chartColor%d' % i
        opt_levels.append(optimization_level)
        color_defs.append((color_name, color))
        data.append((color_name, measurements.items()))
        for name, _ in measurements.items():
            if name not in test_names:
                test_names.append(name)

    return assemble_latex_chart(opt_levels, color_defs, test_names, data,
                                embed, bar_width)


def create_stacked_latex_chart(perf_data,
                               sorted_optimization_levels=None,
                               embed=False,
                               bar_width=DEFAULT_BAR_WIDTH):
    """Creates a stacked LaTeX chart for the given performance data."""
    color_scheme = generate_color_scheme([q for q, _ in perf_data])
    opt_levels = []
    quantity_colors = {}
    test_names = []
    for quantity, measurements in perf_data:
        quantity_colors[quantity] = color_scheme[quantity]
        for optimization_level, data_points in measurements:
            if optimization_level not in opt_levels:
                opt_levels.append(optimization_level)

            for name, _ in data_points:
                if name not in test_names:
                    test_names.append(name)

    sorted_data = list(
        sorted(perf_data, key=lambda (_, data): sum(get_mean_runtimes(perf_list_to_dict(data)).values())))

    if sorted_optimization_levels is None and len(sorted_data) > 0:
        sorted_optimization_levels = sort_by_runtime(sorted_data[0])

    sorted_data = [(quantity, list(
        sorted(
            data, key=lambda (opt, _): sorted_optimization_levels.index(opt))))
                   for quantity, data in sorted_data]

    return assemble_stacked_latex_chart(
        sorted_optimization_levels,
        [(q, quantity_colors[q]) for q, _ in sorted_data], test_names,
        sorted_data, embed, bar_width)


def get_mean_runtimes(perf_data):
    """Computes the mean run-time of every optimization level in the given
       performance data."""
    return {
        opt_level: utils.mean(perf_data[opt_level].values())
        for opt_level in perf_data.keys()
    }


def get_baseline_optimization_level(perf_data):
    """Gets a baseline optimization level from the given performance data.
       This baseline optimization level is guaranteed to exist for every test case.
       If no baseline optimization level can be found, then None is returned."""
    # First find the name of all test cases.
    all_test_names = set()
    for optimization_level, measurements in perf_data.items():
        all_test_names.update(measurements.keys())

    # Filter optimization levels which are used for every test case.
    candidate_opt_levels = []
    for optimization_level, measurements in perf_data.items():
        if len(all_test_names) == len(measurements):
            candidate_opt_levels.append(optimization_level)

    if len(candidate_opt_levels) == 0:
        # Looks like there is no baseline optimization level.
        return None

    # Pick the optimization level with the highest total run-time as the baseline.
    return max(candidate_opt_levels,
               key=lambda opt_level: sum(perf_data[opt_level].values()))


def get_relative_measurements(perf_data,
                              baseline_optimization_level,
                              baseline_perf_data=None):
    """Computes a map of measurements that are relative to the given optimization level."""
    if baseline_perf_data is None:
        baseline_perf_data = perf_data

    results = {}
    for optimization_level, measurements in perf_data.items():
        results[optimization_level] = {}
        for test_name, data_point in measurements.items():
            results[optimization_level][test_name] = (
                data_point /
                baseline_perf_data[baseline_optimization_level][test_name])

    return results


def perf_list_to_dict(perf_list):
    """Converts performance data from a list representation to a dictionary representation."""
    return {opt_level: dict(tests) for opt_level, tests in perf_list}


def perf_dict_to_list(perf_dict):
    """Converts performance data from a dictionary representation to a list representation."""
    return [(opt_level, tests.items())
            for opt_level, tests in perf_dict.items()]


def interpolate(value_range, index, length):
    """Uses an index and a length to interpolate in the given range."""
    min_val, max_val = value_range
    if length == 1:
        return max_val
    else:
        return min_val + float(index) * (max_val - min_val) / float(length - 1)


def sort_by_runtime(perf_data):
    """Sorts the optimization levels by mean relative runtimes."""
    baseline_opt_level = get_baseline_optimization_level(perf_data)
    relative_perf = get_relative_measurements(perf_data, baseline_opt_level)
    # Sort the optimization levels by their mean runtimes.
    mean_runtimes = get_mean_runtimes(relative_perf)
    return list(
        sorted(
            mean_runtimes.keys(),
            key=lambda opt_level: mean_runtimes[opt_level],
            reverse=True))


def generate_color_scheme(sorted_opt_levels):
    """Assigns a color to every optimization level in the given performance data."""
    # Assign colors to the optimization levels.
    color_scheme = {}
    min_hue, min_sat, min_val = colorsys.rgb_to_hsv(
        *[c / float(255) for c in COLOR_SCHEME_MIN_COLOR])
    max_hue, max_sat, max_val = colorsys.rgb_to_hsv(
        *[c / float(255) for c in COLOR_SCHEME_MAX_COLOR])
    for i, opt_level in enumerate(sorted_opt_levels):
        hue = interpolate((min_hue, max_hue), i, len(sorted_opt_levels))
        sat = interpolate((min_sat, max_sat), i, len(sorted_opt_levels))
        val = interpolate((min_val, max_val), i, len(sorted_opt_levels))
        color = [
            component * 255 for component in colorsys.hsv_to_rgb(hue, sat, val)
        ]
        color_scheme[opt_level] = color

    return color_scheme


def filter_perf_data(perf_data,
                     filter_quantity=None,
                     filter_opt_level=None,
                     filter_test=None):
    """Applies a number of filters to performance data."""
    if filter_quantity is None:
        filter_quantity = lambda _: True
    if filter_opt_level is None:
        filter_opt_level = lambda _: True
    if filter_test is None:
        filter_test = lambda _: True

    results = {}
    for quantity, quantity_data in perf_data.items():
        if not filter_quantity(quantity):
            continue

        new_quantity_data = []
        for opt_level, opt_data in quantity_data:
            if not filter_opt_level(opt_level):
                continue

            new_opt_data = []
            for test, data in opt_data:
                if filter_test(test):
                    new_opt_data.append((test, data))

            if len(new_opt_data) > 0:
                new_quantity_data.append((opt_level, new_opt_data))

        if len(new_quantity_data) > 0:
            results[quantity] = new_quantity_data

    return results


def subtract_perf_data(lhs_perf_dict, *rhs_perf_dicts):
    """Performs measurement-wise subtraction on performance dictionaries."""
    return {
        opt_level: {
            test_name: measurement - sum(p_dict[opt_level][test_name]
                                         for p_dict in rhs_perf_dicts)
            for test_name, measurement in opt_data.items()
        }
        for opt_level, opt_data in lhs_perf_dict.items()
    }


def main():
    arg_parser = argparse.ArgumentParser()
    arg_parser.add_argument(
        'input', help='The performance data file.', nargs='?', default=None)
    arg_parser.add_argument(
        '-q',
        '--quantity',
        type=str,
        help="The quantity to build a bar chart for. Defaults to '%s'." %
        utils.TOTAL_TIME_QUANTITY,
        default=utils.TOTAL_TIME_QUANTITY)
    arg_parser.add_argument(
        '-O',
        '--opt',
        type=str,
        nargs='*',
        help="Filters on optimization levels.")
    arg_parser.add_argument(
        '-t', '--test', type=str, nargs='*', help="Filters on tests.")
    arg_parser.add_argument(
        '-r',
        '--relative',
        action='store_const',
        const=True,
        help="Produce bars that are relative to some baseline.",
        default=False)
    arg_parser.add_argument(
        '-e',
        '--embed',
        action='store_const',
        const=True,
        help="Don't include a LaTeX document header and footer.",
        default=False)
    arg_parser.add_argument(
        '-b',
        '--bar-width',
        type=int,
        help="The width, in points, of a bar on the bar chart. Defaults to '%s'."
        % DEFAULT_BAR_WIDTH,
        default=DEFAULT_BAR_WIDTH)
    arg_parser.add_argument(
        '-s',
        '--split',
        type=str,
        nargs='*',
        help="Picks other quantities which are subtracted from the "
        "main quantity and then combined in a stacked chart.")

    args = arg_parser.parse_args()

    all_perf_data = utils.parse_perf_data(args.input)
    sorted_opt_levels = None

    opt_level_filter = None
    if args.opt:
        optimization_set = set(args.opt)
        opt_level_filter = lambda x: x in optimization_set
        sorted_opt_levels = list(args.opt)

    test_filter = None
    if args.test:
        test_set = set(args.test)
        test_filter = lambda x: x in test_set

    all_perf_data = filter_perf_data(
        all_perf_data,
        filter_opt_level=opt_level_filter,
        filter_test=test_filter)

    main_perf_data = perf_list_to_dict(all_perf_data[args.quantity])
    baseline_opt_level = get_baseline_optimization_level(main_perf_data)

    split_perf_data = [(s, perf_list_to_dict(all_perf_data[s]))
                       for s in (args.split if args.split is not None else [])]
    sub_perf_data = subtract_perf_data(
        main_perf_data, *[s_data for _, s_data in split_perf_data])

    if args.relative:
        sub_perf_data = get_relative_measurements(
            sub_perf_data, baseline_opt_level, main_perf_data)
        split_perf_data = [(q, get_relative_measurements(
            split_data, baseline_opt_level, main_perf_data))
                           for q, split_data in split_perf_data]

    split_perf_data = [(q, perf_dict_to_list(split_data))
                       for q, split_data in split_perf_data]

    if len(split_perf_data) == 0:
        print(create_latex_chart(sub_perf_data, sorted_opt_levels, args.embed,
                                 args.bar_width))
    else:
        print(create_stacked_latex_chart([
            (args.quantity, perf_dict_to_list(sub_perf_data))
        ] + split_perf_data, sorted_opt_levels, args.embed, args.bar_width))


if __name__ == '__main__':
    main()