abitimer.py

# coding: utf-8
# Copyright (c) Pymatgen Development Team.
# Distributed under the terms of the MIT License.
"""
This module provides objects for extracting timing data from the ABINIT output files
It also provides tools to analye and to visualize the parallel efficiency.
"""

import collections
import logging
import os
import sys

import numpy as np
from monty.string import is_string, list_strings

from pymatgen.util.num import minloc
from pymatgen.util.plotting import add_fig_kwargs, get_ax_fig_plt

logger = logging.getLogger(__name__)


def alternate(*iterables):
    """
    [a[0], b[0], ... , a[1], b[1], ..., a[n], b[n] ...]
    >>> alternate([1,4], [2,5], [3,6])
    [1, 2, 3, 4, 5, 6]
    """
    items = []
    for tup in zip(*iterables):
        items.extend(tup)
    return items


class AbinitTimerParserError(Exception):
    """Errors raised by AbinitTimerParser"""


class AbinitTimerParser(collections.abc.Iterable):
    """
    Responsible for parsing a list of output files, extracting the timing results
    and analyzing the results.
    Assume the Abinit output files have been produced with `timopt -1`.

    Example:

        parser = AbinitTimerParser()
        parser.parse(list_of_files)

    To analyze all *.abo files withing top, use:

        parser, paths, okfiles = AbinitTimerParser.walk(top=".", ext=".abo")
    """

    # The markers enclosing the data.
    BEGIN_TAG = "-<BEGIN_TIMER"
    END_TAG = "-<END_TIMER>"

    Error = AbinitTimerParserError

    # DEFAULT_MPI_RANK = "0"

    @classmethod
    def walk(cls, top=".", ext=".abo"):
        """
        Scan directory tree starting from top, look for files with extension `ext` and
        parse timing data.

        Return: (parser, paths, okfiles)
            where `parser` is the new object, `paths` is the list of files found and `okfiles`
            is the list of files that have been parsed successfully.
            (okfiles == paths) if all files have been parsed.
        """
        paths = []
        for root, dirs, files in os.walk(top):
            for f in files:
                if f.endswith(ext):
                    paths.append(os.path.join(root, f))

        parser = cls()
        okfiles = parser.parse(paths)
        return parser, paths, okfiles

    def __init__(self):
        """Initialize object."""
        # List of files that have been parsed.
        self._filenames = []

        # timers[filename][mpi_rank]
        # contains the timer extracted from the file filename associated to the MPI rank mpi_rank.
        self._timers = collections.OrderedDict()

    def __iter__(self):
        return self._timers.__iter__()

    def __len__(self):
        return len(self._timers)

    @property
    def filenames(self):
        """List of files that have been parsed successfully."""
        return self._filenames

    def parse(self, filenames):
        """
        Read and parse a filename or a list of filenames.
        Files that cannot be opened are ignored. A single filename may also be given.

        Return: list of successfully read files.
        """
        filenames = list_strings(filenames)

        read_ok = []
        for fname in filenames:
            try:
                fh = open(fname)  # pylint: disable=R1732
            except IOError:
                logger.warning("Cannot open file %s" % fname)
                continue

            try:
                self._read(fh, fname)
                read_ok.append(fname)

            except self.Error as e:
                logger.warning("exception while parsing file %s:\n%s" % (fname, str(e)))
                continue

            finally:
                fh.close()

        # Add read_ok to the list of files that have been parsed.
        self._filenames.extend(read_ok)
        return read_ok

    def _read(self, fh, fname):
        """Parse the TIMER section"""
        if fname in self._timers:
            raise self.Error("Cannot overwrite timer associated to: %s " % fname)

        def parse_line(line):
            """Parse single line."""
            name, vals = line[:25], line[25:].split()
            try:
                ctime, cfract, wtime, wfract, ncalls, gflops = vals
            except ValueError:
                # v8.3 Added two columns at the end [Speedup, Efficacity]
                ctime, cfract, wtime, wfract, ncalls, gflops, speedup, eff = vals

            return AbinitTimerSection(name, ctime, cfract, wtime, wfract, ncalls, gflops)

        sections, info, cpu_time, wall_time = None, None, None, None
        data = {}
        inside, has_timer = 0, False
        for line in fh:
            # print(line.strip())
            if line.startswith(self.BEGIN_TAG):
                has_timer = True
                sections = []
                info = {}
                inside = 1
                line = line[len(self.BEGIN_TAG) :].strip()[:-1]

                info["fname"] = fname
                for tok in line.split(","):
                    key, val = [s.strip() for s in tok.split("=")]
                    info[key] = val

            elif line.startswith(self.END_TAG):
                inside = 0
                timer = AbinitTimer(sections, info, cpu_time, wall_time)
                mpi_rank = info["mpi_rank"]
                data[mpi_rank] = timer

            elif inside:
                inside += 1
                line = line[1:].strip()

                if inside == 2:
                    d = {}
                    for tok in line.split(","):
                        key, val = [s.strip() for s in tok.split("=")]
                        d[key] = float(val)
                    cpu_time, wall_time = d["cpu_time"], d["wall_time"]

                elif inside > 5:
                    sections.append(parse_line(line))

                else:
                    try:
                        parse_line(line)
                    except Exception:
                        parser_failed = True

                    if not parser_failed:
                        raise self.Error("line should be empty: " + str(inside) + line)

        if not has_timer:
            raise self.Error("%s: No timer section found" % fname)

        # Add it to the dict
        self._timers[fname] = data

    def timers(self, filename=None, mpi_rank="0"):
        """
        Return the list of timers associated to the given `filename` and MPI rank mpi_rank.
        """
        if filename is not None:
            return [self._timers[filename][mpi_rank]]
        return [self._timers[filename][mpi_rank] for filename in self._filenames]

    def section_names(self, ordkey="wall_time"):
        """
        Return the names of sections ordered by ordkey.
        For the time being, the values are taken from the first timer.
        """
        section_names = []

        # FIXME this is not trivial
        for idx, timer in enumerate(self.timers()):
            if idx == 0:
                section_names = [s.name for s in timer.order_sections(ordkey)]
                # check = section_names
                # else:
                #  new_set = set( [s.name for s in timer.order_sections(ordkey)])
                #  section_names.intersection_update(new_set)
                #  check = check.union(new_set)

        # if check != section_names:
        #  print("sections", section_names)
        #  print("check",check)

        return section_names

    def get_sections(self, section_name):
        """
        Return the list of sections stored in self.timers() given `section_name`
        A fake section is returned if the timer does not have section_name.
        """
        sections = []
        for timer in self.timers():
            for sect in timer.sections:
                if sect.name == section_name:
                    sections.append(sect)
                    break
            else:
                sections.append(AbinitTimerSection.fake())

        return sections

    def pefficiency(self):
        """
        Analyze the parallel efficiency.

        Return: :class:`ParallelEfficiency` object.
        """
        timers = self.timers()

        # Number of CPUs employed in each calculation.
        ncpus = [timer.ncpus for timer in timers]

        # Find the minimum number of cpus used and its index in timers.
        min_idx = minloc(ncpus)
        min_ncpus = ncpus[min_idx]

        # Reference timer
        ref_t = timers[min_idx]

        # Compute the parallel efficiency (total and section efficiency)
        peff = {}
        ctime_peff = [(min_ncpus * ref_t.wall_time) / (t.wall_time * ncp) for (t, ncp) in zip(timers, ncpus)]
        wtime_peff = [(min_ncpus * ref_t.cpu_time) / (t.cpu_time * ncp) for (t, ncp) in zip(timers, ncpus)]
        n = len(timers)

        peff["total"] = {}
        peff["total"]["cpu_time"] = ctime_peff
        peff["total"]["wall_time"] = wtime_peff
        peff["total"]["cpu_fract"] = n * [100]
        peff["total"]["wall_fract"] = n * [100]

        for sect_name in self.section_names():
            # print(sect_name)
            ref_sect = ref_t.get_section(sect_name)
            sects = [t.get_section(sect_name) for t in timers]
            try:
                ctime_peff = [(min_ncpus * ref_sect.cpu_time) / (s.cpu_time * ncp) for (s, ncp) in zip(sects, ncpus)]
                wtime_peff = [(min_ncpus * ref_sect.wall_time) / (s.wall_time * ncp) for (s, ncp) in zip(sects, ncpus)]
            except ZeroDivisionError:
                ctime_peff = n * [-1]
                wtime_peff = n * [-1]

            assert sect_name not in peff
            peff[sect_name] = {}
            peff[sect_name]["cpu_time"] = ctime_peff
            peff[sect_name]["wall_time"] = wtime_peff

            peff[sect_name]["cpu_fract"] = [s.cpu_fract for s in sects]
            peff[sect_name]["wall_fract"] = [s.wall_fract for s in sects]

        return ParallelEfficiency(self._filenames, min_idx, peff)

    def summarize(self, **kwargs):
        """
        Return pandas DataFrame with the most important results stored in the timers.
        """
        import pandas as pd

        colnames = [
            "fname",
            "wall_time",
            "cpu_time",
            "mpi_nprocs",
            "omp_nthreads",
            "mpi_rank",
        ]

        frame = pd.DataFrame(columns=colnames)
        for i, timer in enumerate(self.timers()):
            frame = frame.append({k: getattr(timer, k) for k in colnames}, ignore_index=True)
        frame["tot_ncpus"] = frame["mpi_nprocs"] * frame["omp_nthreads"]

        # Compute parallel efficiency (use the run with min number of cpus to normalize).
        i = frame["tot_ncpus"].values.argmin()
        ref_wtime = frame.iloc[i]["wall_time"]
        ref_ncpus = frame.iloc[i]["tot_ncpus"]
        frame["peff"] = (ref_ncpus * ref_wtime) / (frame["wall_time"] * frame["tot_ncpus"])

        return frame

    @add_fig_kwargs
    def plot_efficiency(self, key="wall_time", what="good+bad", nmax=5, ax=None, **kwargs):
        """
        Plot the parallel efficiency

        Args:
            key: Parallel efficiency is computed using the wall_time.
            what: Specifies what to plot: `good` for sections with good parallel efficiency.
                `bad` for sections with bad efficiency. Options can be concatenated with `+`.
            nmax: Maximum number of entries in plot
            ax: matplotlib :class:`Axes` or None if a new figure should be created.

        ================  ====================================================
        kwargs            Meaning
        ================  ====================================================
        linewidth         matplotlib linewidth. Default: 2.0
        markersize        matplotlib markersize. Default: 10
        ================  ====================================================

        Returns:
            `matplotlib` figure
        """
        ax, fig, plt = get_ax_fig_plt(ax=ax)
        lw = kwargs.pop("linewidth", 2.0)
        msize = kwargs.pop("markersize", 10)
        what = what.split("+")

        timers = self.timers()
        peff = self.pefficiency()
        n = len(timers)
        xx = np.arange(n)

        # ax.set_color_cycle(['g', 'b', 'c', 'm', 'y', 'k'])
        ax.set_prop_cycle(color=["g", "b", "c", "m", "y", "k"])

        lines, legend_entries = [], []
        # Plot sections with good efficiency.
        if "good" in what:
            good = peff.good_sections(key=key, nmax=nmax)
            for g in good:
                # print(g, peff[g])
                yy = peff[g][key]
                (line,) = ax.plot(xx, yy, "-->", linewidth=lw, markersize=msize)
                lines.append(line)
                legend_entries.append(g)

        # Plot sections with bad efficiency.
        if "bad" in what:
            bad = peff.bad_sections(key=key, nmax=nmax)
            for b in bad:
                # print(b, peff[b])
                yy = peff[b][key]
                (line,) = ax.plot(xx, yy, "-.<", linewidth=lw, markersize=msize)
                lines.append(line)
                legend_entries.append(b)

        # Add total if not already done
        if "total" not in legend_entries:
            yy = peff["total"][key]
            (total_line,) = ax.plot(xx, yy, "r", linewidth=lw, markersize=msize)
            lines.append(total_line)
            legend_entries.append("total")

        ax.legend(lines, legend_entries, loc="best", shadow=True)

        # ax.set_title(title)
        ax.set_xlabel("Total_NCPUs")
        ax.set_ylabel("Efficiency")
        ax.grid(True)

        # Set xticks and labels.
        labels = ["MPI=%d, OMP=%d" % (t.mpi_nprocs, t.omp_nthreads) for t in timers]
        ax.set_xticks(xx)
        ax.set_xticklabels(labels, fontdict=None, minor=False, rotation=15)

        return fig

    @add_fig_kwargs
    def plot_pie(self, key="wall_time", minfract=0.05, **kwargs):
        """
        Plot pie charts of the different timers.

        Args:
            key: Keyword used to extract data from timers.
            minfract: Don't show sections whose relative weight is less that minfract.

        Returns:
            `matplotlib` figure
        """
        timers = self.timers()
        n = len(timers)

        # Make square figures and axes
        import matplotlib.pyplot as plt
        from matplotlib.gridspec import GridSpec

        fig = plt.gcf()
        gspec = GridSpec(n, 1)
        for idx, timer in enumerate(timers):
            ax = plt.subplot(gspec[idx, 0])
            ax.set_title(str(timer))
            timer.pie(ax=ax, key=key, minfract=minfract, show=False)

        return fig

    @add_fig_kwargs
    def plot_stacked_hist(self, key="wall_time", nmax=5, ax=None, **kwargs):
        """
        Plot stacked histogram of the different timers.

        Args:
            key: Keyword used to extract data from the timers. Only the first `nmax`
                sections with largest value are show.
            mmax: Maximum nuber of sections to show. Other entries are grouped together
                in the `others` section.
            ax: matplotlib :class:`Axes` or None if a new figure should be created.

        Returns:
            `matplotlib` figure
        """
        ax, fig, plt = get_ax_fig_plt(ax=ax)

        mpi_rank = "0"
        timers = self.timers(mpi_rank=mpi_rank)
        n = len(timers)

        names, values = [], []
        rest = np.zeros(n)

        for idx, sname in enumerate(self.section_names(ordkey=key)):
            sections = self.get_sections(sname)
            svals = np.asarray([s.__dict__[key] for s in sections])
            if idx < nmax:
                names.append(sname)
                values.append(svals)
            else:
                rest += svals

        names.append("others (nmax=%d)" % nmax)
        values.append(rest)

        # The dataset is stored in values. Now create the stacked histogram.
        ind = np.arange(n)  # the locations for the groups
        width = 0.35  # the width of the bars
        colors = nmax * ["r", "g", "b", "c", "k", "y", "m"]

        bars = []
        bottom = np.zeros(n)
        for idx, vals in enumerate(values):
            color = colors[idx]
            bar_ = ax.bar(ind, vals, width, color=color, bottom=bottom)
            bars.append(bar_)
            bottom += vals

        ax.set_ylabel(key)
        ax.set_title("Stacked histogram with the %d most important sections" % nmax)

        ticks = ind + width / 2.0
        labels = ["MPI=%d, OMP=%d" % (t.mpi_nprocs, t.omp_nthreads) for t in timers]
        ax.set_xticks(ticks)
        ax.set_xticklabels(labels, rotation=15)

        # Add legend.
        ax.legend([bar_[0] for bar_ in bars], names, loc="best")

        return fig

    def plot_all(self, show=True, **kwargs):
        """
        Call all plot methods provided by the parser.
        """
        figs = []
        app = figs.append
        app(self.plot_stacked_hist(show=show))
        app(self.plot_efficiency(show=show))
        app(self.plot_pie(show=show))
        return figs


class ParallelEfficiency(dict):
    """
    Store results concerning the parallel efficiency of the job.
    """

    def __init__(self, filenames, ref_idx, *args, **kwargs):
        """
        Args:
            filennames: List of filenames
            ref_idx: Index of the Reference time (calculation done with the smallest number of cpus)
        """
        self.update(*args, **kwargs)
        self.filenames = filenames
        self._ref_idx = ref_idx

    def _order_by_peff(self, key, criterion, reverse=True):

        self.estimator = {
            "min": min,
            "max": max,
            "mean": lambda items: sum(items) / len(items),
        }[criterion]

        data = []
        for (sect_name, peff) in self.items():
            # Ignore values where we had a division by zero.
            if all(v != -1 for v in peff[key]):
                values = peff[key][:]
                # print(sect_name, values)
                if len(values) > 1:
                    ref_value = values.pop(self._ref_idx)
                    assert ref_value == 1.0

                data.append((sect_name, self.estimator(values)))

        data.sort(key=lambda t: t[1], reverse=reverse)
        return tuple(sect_name for (sect_name, e) in data)

    def totable(self, stop=None, reverse=True):
        """
        Return table (list of lists) with timing results.

        Args:
            stop: Include results up to stop. None for all
            reverse: Put items with highest wall_time in first positions if True.
        """
        osects = self._order_by_peff("wall_time", criterion="mean", reverse=reverse)
        if stop is not None:
            osects = osects[:stop]

        n = len(self.filenames)
        table = [["AbinitTimerSection"] + alternate(self.filenames, n * ["%"])]
        for sect_name in osects:
            peff = self[sect_name]["wall_time"]
            fract = self[sect_name]["wall_fract"]
            vals = alternate(peff, fract)

            table.append([sect_name] + ["%.2f" % val for val in vals])

        return table

    def good_sections(self, key="wall_time", criterion="mean", nmax=5):
        """
        Return first `nmax` sections with best value of key `key` using criterion `criterion`.
        """
        good_sections = self._order_by_peff(key, criterion=criterion)
        return good_sections[:nmax]

    def bad_sections(self, key="wall_time", criterion="mean", nmax=5):
        """
        Return first `nmax` sections with worst value of key `key` using criterion `criterion`.
        """
        bad_sections = self._order_by_peff(key, criterion=criterion, reverse=False)
        return bad_sections[:nmax]


class AbinitTimerSection:
    """Record with the timing results associated to a section of code."""

    STR_FIELDS = ["name"]

    NUMERIC_FIELDS = [
        "wall_time",
        "wall_fract",
        "cpu_time",
        "cpu_fract",
        "ncalls",
        "gflops",
    ]

    FIELDS = tuple(STR_FIELDS + NUMERIC_FIELDS)

    @classmethod
    def fake(cls):
        """Return a fake section. Mainly used to fill missing entries if needed."""
        return AbinitTimerSection("fake", 0.0, 0.0, 0.0, 0.0, -1, 0.0)

    def __init__(self, name, cpu_time, cpu_fract, wall_time, wall_fract, ncalls, gflops):
        """
        Args:
            name: Name of the sections.
            cpu_time: CPU time in seconds.
            cpu_fract: Percentage of CPU time.
            wall_time: Wall-time in seconds.
            wall_fract: Percentage of wall-time.
            ncalls: Number of calls
            gflops: Gigaflops.
        """
        self.name = name.strip()
        self.cpu_time = float(cpu_time)
        self.cpu_fract = float(cpu_fract)
        self.wall_time = float(wall_time)
        self.wall_fract = float(wall_fract)
        self.ncalls = int(ncalls)
        self.gflops = float(gflops)

    def to_tuple(self):
        """Convert object to tuple."""
        return tuple(self.__dict__[at] for at in AbinitTimerSection.FIELDS)

    def to_dict(self):
        """Convert object to dictionary."""
        return {at: self.__dict__[at] for at in AbinitTimerSection.FIELDS}

    def to_csvline(self, with_header=False):
        """Return a string with data in CSV format. Add header if `with_header`"""
        string = ""

        if with_header:
            string += "# " + " ".join(at for at in AbinitTimerSection.FIELDS) + "\n"

        string += ", ".join(str(v) for v in self.to_tuple()) + "\n"
        return string

    def __str__(self):
        """String representation."""
        string = ""
        for a in AbinitTimerSection.FIELDS:
            string += a + " = " + self.__dict__[a] + ","
        return string[:-1]


class AbinitTimer:
    """Container class storing the timing results."""

    def __init__(self, sections, info, cpu_time, wall_time):
        """
        Args:
            sections: List of sections
            info: Dictionary with extra info.
            cpu_time: Cpu-time in seconds.
            wall_time: Wall-time in seconds.
        """
        # Store sections and names
        self.sections = tuple(sections)
        self.section_names = tuple(s.name for s in self.sections)

        self.info = info
        self.cpu_time = float(cpu_time)
        self.wall_time = float(wall_time)
        self.mpi_nprocs = int(info["mpi_nprocs"])
        self.omp_nthreads = int(info["omp_nthreads"])
        self.mpi_rank = info["mpi_rank"].strip()
        self.fname = info["fname"].strip()

    def __str__(self):
        string = "file=%s, wall_time=%.1f, mpi_nprocs=%d, omp_nthreads=%d" % (
            self.fname,
            self.wall_time,
            self.mpi_nprocs,
            self.omp_nthreads,
        )
        # string += ", rank = " + self.mpi_rank
        return string

    @property
    def ncpus(self):
        """Total number of CPUs employed."""
        return self.mpi_nprocs * self.omp_nthreads

    def get_section(self, section_name):
        """Return section associated to `section_name`."""
        try:
            idx = self.section_names.index(section_name)
        except Exception:
            raise
        sect = self.sections[idx]
        assert sect.name == section_name
        return sect

    def to_csv(self, fileobj=sys.stdout):
        """Write data on file fileobj using CSV format."""
        openclose = is_string(fileobj)

        if openclose:
            fileobj = open(fileobj, "w")  # pylint: disable=R1732

        for idx, section in enumerate(self.sections):
            fileobj.write(section.to_csvline(with_header=(idx == 0)))
        fileobj.flush()

        if openclose:
            fileobj.close()

    def to_table(self, sort_key="wall_time", stop=None):
        """Return a table (list of lists) with timer data"""
        table = [
            list(AbinitTimerSection.FIELDS),
        ]
        ord_sections = self.order_sections(sort_key)

        if stop is not None:
            ord_sections = ord_sections[:stop]

        for osect in ord_sections:
            row = [str(item) for item in osect.to_tuple()]
            table.append(row)

        return table

    # Maintain old API
    totable = to_table

    def get_dataframe(self, sort_key="wall_time", **kwargs):
        """
        Return a pandas DataFrame with entries sorted according to `sort_key`.
        """
        import pandas as pd

        frame = pd.DataFrame(columns=AbinitTimerSection.FIELDS)

        for osect in self.order_sections(sort_key):
            frame = frame.append(osect.to_dict(), ignore_index=True)

        # Monkey patch
        frame.info = self.info
        frame.cpu_time = self.cpu_time
        frame.wall_time = self.wall_time
        frame.mpi_nprocs = self.mpi_nprocs
        frame.omp_nthreads = self.omp_nthreads
        frame.mpi_rank = self.mpi_rank
        frame.fname = self.fname

        return frame

    def get_values(self, keys):
        """
        Return a list of values associated to a particular list of keys.
        """
        if is_string(keys):
            return [s.__dict__[keys] for s in self.sections]
        values = []
        for k in keys:
            values.append([s.__dict__[k] for s in self.sections])
        return values

    def names_and_values(self, key, minval=None, minfract=None, sorted=True):
        """
        Select the entries whose value[key] is >= minval or whose fraction[key] is >= minfract
        Return the names of the sections and the corresponding values.
        """
        values = self.get_values(key)
        names = self.get_values("name")

        new_names, new_values = [], []
        other_val = 0.0

        if minval is not None:
            assert minfract is None

            for n, v in zip(names, values):
                if v >= minval:
                    new_names.append(n)
                    new_values.append(v)
                else:
                    other_val += v

            new_names.append("below minval " + str(minval))
            new_values.append(other_val)

        elif minfract is not None:
            assert minval is None

            total = self.sum_sections(key)

            for n, v in zip(names, values):
                if v / total >= minfract:
                    new_names.append(n)
                    new_values.append(v)
                else:
                    other_val += v

            new_names.append("below minfract " + str(minfract))
            new_values.append(other_val)

        else:
            # all values
            new_names, new_values = names, values

        if sorted:
            # Sort new_values and rearrange new_names.
            nandv = list(zip(new_names, new_values))
            nandv.sort(key=lambda t: t[1])
            new_names, new_values = [n[0] for n in nandv], [n[1] for n in nandv]

        return new_names, new_values

    def _reduce_sections(self, keys, operator):
        return operator(self.get_values(keys))

    def sum_sections(self, keys):
        """Sum value of keys."""
        return self._reduce_sections(keys, sum)

    def order_sections(self, key, reverse=True):
        """Sort sections according to the value of key."""
        return sorted(self.sections, key=lambda s: s.__dict__[key], reverse=reverse)

    @add_fig_kwargs
    def cpuwall_histogram(self, ax=None, **kwargs):
        """
        Plot histogram with cpu- and wall-time on axis `ax`.

        Args:
            ax: matplotlib :class:`Axes` or None if a new figure should be created.

        Returns: `matplotlib` figure
        """
        ax, fig, plt = get_ax_fig_plt(ax=ax)

        nk = len(self.sections)
        ind = np.arange(nk)  # the x locations for the groups
        width = 0.35  # the width of the bars

        cpu_times = self.get_values("cpu_time")
        rects1 = plt.bar(ind, cpu_times, width, color="r")

        wall_times = self.get_values("wall_time")
        rects2 = plt.bar(ind + width, wall_times, width, color="y")

        # Add ylable and title
        ax.set_ylabel("Time (s)")

        # plt.title('CPU-time and Wall-time for the different sections of the code')

        ticks = self.get_values("name")
        ax.set_xticks(ind + width, ticks)

        ax.legend((rects1[0], rects2[0]), ("CPU", "Wall"), loc="best")

        return fig

    @add_fig_kwargs
    def pie(self, key="wall_time", minfract=0.05, ax=None, **kwargs):
        """
        Plot pie chart for this timer.

        Args:
            key: Keyword used to extract data from the timer.
            minfract: Don't show sections whose relative weight is less that minfract.
            ax: matplotlib :class:`Axes` or None if a new figure should be created.

        Returns: `matplotlib` figure
        """
        ax, fig, plt = get_ax_fig_plt(ax=ax)
        # Set aspect ratio to be equal so that pie is drawn as a circle.
        ax.axis("equal")
        # Don't show section whose value is less that minfract
        labels, vals = self.names_and_values(key, minfract=minfract)
        ax.pie(vals, explode=None, labels=labels, autopct="%1.1f%%", shadow=True)
        return fig

    @add_fig_kwargs
    def scatter_hist(self, ax=None, **kwargs):
        """
        Scatter plot + histogram.

        Args:
            ax: matplotlib :class:`Axes` or None if a new figure should be created.

        Returns: `matplotlib` figure
        """
        from mpl_toolkits.axes_grid1 import make_axes_locatable

        ax, fig, plt = get_ax_fig_plt(ax=ax)

        x = np.asarray(self.get_values("cpu_time"))
        y = np.asarray(self.get_values("wall_time"))

        # the scatter plot:
        axScatter = plt.subplot(1, 1, 1)
        axScatter.scatter(x, y)
        axScatter.set_aspect("auto")

        # create new axes on the right and on the top of the current axes
        # The first argument of the new_vertical(new_horizontal) method is
        # the height (width) of the axes to be created in inches.
        divider = make_axes_locatable(axScatter)
        axHistx = divider.append_axes("top", 1.2, pad=0.1, sharex=axScatter)
        axHisty = divider.append_axes("right", 1.2, pad=0.1, sharey=axScatter)

        # make some labels invisible
        plt.setp(axHistx.get_xticklabels() + axHisty.get_yticklabels(), visible=False)

        # now determine nice limits by hand:
        binwidth = 0.25
        xymax = np.max([np.max(np.fabs(x)), np.max(np.fabs(y))])
        lim = (int(xymax / binwidth) + 1) * binwidth

        bins = np.arange(-lim, lim + binwidth, binwidth)
        axHistx.hist(x, bins=bins)
        axHisty.hist(y, bins=bins, orientation="horizontal")

        # the xaxis of axHistx and yaxis of axHisty are shared with axScatter,
        # thus there is no need to manually adjust the xlim and ylim of these axis.

        # axHistx.axis["bottom"].major_ticklabels.set_visible(False)
        for tl in axHistx.get_xticklabels():
            tl.set_visible(False)
            axHistx.set_yticks([0, 50, 100])

            # axHisty.axis["left"].major_ticklabels.set_visible(False)
            for tl in axHisty.get_yticklabels():
                tl.set_visible(False)
                axHisty.set_xticks([0, 50, 100])

        # plt.draw()
        return fig