utils/compare.py: Fix support for multiple metrics

[arichardson]

Summary

In addition, this retains the pandas multi-index to improve the
visualization. We also transform the program column in-place now since
it appears that DataFrame.to_string() ignores the formatter argument for
all columns that have dtype=object.

Before: utils/compare.py a.json vs b.json --lhs-name=baseline --rhs-name=new -m compile_time -m link_time --nodiff

Program                                        baseline  new    baseline  new
-C++/Shootout-C++-lists                         0.00      0.00   0.09      0.08
-C++/Shootout-C++-objinst                       0.00      0.00   0.07      0.08
...
         baseline         new    baseline         new
count  315.000000  315.000000  315.000000  315.000000
mean   3.543040    3.531691    0.076500    0.078680
...

As can be seen here, the metric is missing from the column headers, so it's not
clear which one is which if you use multiple metrics.

With this change we now see the metrics, and they are sorted by largest diff:

Program                                       compile_time        link_time
                                              baseline     new    baseline  new
SingleSource/Benchmarks/Stanford/Perm           0.07         0.08   0.08      0.11
SingleSour...Benchmarks/Stanford/Bubblesort     0.07         0.08   0.10      0.10
...
      compile_time               link_time
l/r       baseline         new    baseline         new
count  315.000000   315.000000  315.000000  315.000000
mean   3.543040     3.531691    0.076500    0.078680
...

More importantly (my original motivation for this change), the script currently
fails to run when passing multiple metrics without --nodiff. For example,
utils/compare.py a.json vs b.json -m compile_time -m link_time result in:
pandas.errors.InvalidIndexError: Reindexing only valid with uniquely valued Index objects
I believe this used to work and regressed in 0ee69b891c3b167474212eb996b78c23f82bc13d
(https://reviews.llvm.org/D57828) as the exception is raised in add_geomean_row()

After:

Program                                       compile_time              link_time
                                              baseline     new    diff  baseline  new    diff
SingleSource/Benchmarks/Stanford/Perm           0.07         0.08 26.2%   0.08      0.11  34.4%
SingleSour...Benchmarks/Stanford/Bubblesort     0.07         0.08 16.9%   0.10      0.10   4.0%
...
                           Geomean difference                     -0.4%                    2.9%
      compile_time                           link_time
l/r       baseline         new        diff    baseline         new        diff
count  315.000000   315.000000  283.000000  315.000000  315.000000  315.000000
mean   3.543040     3.531691   -0.002523    0.076500    0.078680    0.041652
...

Listing multiple metrics with a single file was also broken:
utils/compare.py a.json -m compile_time -m link_time results an exception
while executing d['$sortkey'] = d[sortkey].
ValueError: Expected a 1D array, got an array with shape (315, 2).
This problem has most likely existed since the initial commit that added this
script, but with this change we can now handle mutiple metrics using the newer
pandas d.sort_values(by=(metrics[0], sortkey), ascending=False) instead of
adding a new column and sorting on that. We now get a table sorted by the
first metric:

Program                                       compile_time link_time
                                              results      results
ultiSource...Benchmarks/7zip/7zip-benchmark   109.56         0.22
ultiSource/Benchmarks/Bullet/bullet            73.55         0.17
...
      compile_time   link_time
run        results     results
count  315.000000   315.000000
mean   3.543040     0.076500
...

[arichardson]

Ensure the full terminal width is used for the d.describe() call.

[arichardson]

ping?

[MatzeB]

Changes make sense to me, and a quick sanity check of the script seemed fine.

Thanks, LGTM

utils/compare.py

217

Seems like you fixed the TODO and can remove it?

utils/compare.py

 #!/usr/bin/env python
 """Tool to filter, organize, compare and display benchmarking results. Usefull
 for smaller datasets. It works great with a few dozen runs it is not designed to
 deal with hundreds.
 Requires the pandas library to be installed."""
 from __future__ import print_function
 
 import pandas as pd
 from scipy import stats
 import sys
 import os.path
 import re
 import numbers
 import argparse
 
+GEOMEAN_ROW = 'Geomean difference'
+
 def read_lit_json(filename):
     import json
     jsondata = json.load(open(filename))
     columns = []
     columnindexes = {}
     names = set()
     info_columns = ['hash']
     # Pass1: Figure out metrics (= the column index)
def get_values(values):
     if 'diff' in values.columns:
         values = values[[c for c in values.columns if c != 'diff']]
     has_two_runs = len(values.columns) == 2
     if has_two_runs:
         return (values.iloc[:,0], values.iloc[:,1])
     else:
         return (values.min(axis=1), values.max(axis=1))
 
-def add_diff_column(values, absolute_diff=False):
-    values0, values1 = get_values(values)
+def add_diff_column(metric, values, absolute_diff=False):
+    values0, values1 = get_values(values[metric])
     # Quotient or absolute difference?
     if absolute_diff:
-        values['diff'] = values1 - values0
+        values[(metric, 'diff')] = values1 - values0
     else:
-        values['diff'] = values1 / values0
-        values['diff'] -= 1.0
+        values[(metric, 'diff')] = (values1 / values0) - 1.0
     return values
 
-def add_geomean_row(data, dataout):
+def add_geomean_row(metrics, data, dataout):
     """
     Normalize values1 over values0, compute geomean difference and add a
     summary row to dataout.
     """
-    values0, values1 = get_values(data)
+    gm = pd.DataFrame(index=[GEOMEAN_ROW], columns=dataout.columns,
+                      dtype='float64')
+    for metric in metrics:
+        values0, values1 = get_values(data[metric])
+        # Avoid infinite values in the diff and instead use NaN, as otherwise
+        # the computation of the geometric mean will fail.
+        values0 = values0.replace({0: float('NaN')})
         relative = values1 / values0
-    gm_diff = stats.gmean(relative) - 1.0
-
-    gm_row = {c: '' for c in dataout.columns}
-    gm_row['diff'] = gm_diff
-    series = pd.Series(gm_row, name='Geomean difference')
-    return dataout.append(series)
+        gm_diff = stats.gmean(relative.dropna()) - 1.0
+        gm[(metric, 'diff')] = gm_diff
+    gm.Program = GEOMEAN_ROW
+    return pd.concat([dataout, gm])
 
 def filter_failed(data, key='Exec'):
     return data.loc[data[key] == "pass"]
 
 def filter_short(data, key='Exec_Time', threshold=0.6):
     return data.loc[data[key] >= threshold]
 
 def filter_same_hash(data, key='hash'):

 def format_diff(value):
     if not isinstance(value, numbers.Integral):
         return "%4.1f%%" % (value * 100.)
     else:
         return "%-5d" % value
 
 def print_result(d, limit_output=True, shorten_names=True, minimal_names=False,
                  show_diff_column=True, sortkey='diff', sort_by_abs=True):
     # sort (TODO: is there a more elegant way than create+drop a column?)
+    metrics = d.columns.levels[0]
     if sort_by_abs:
-        d['$sortkey'] = d[sortkey].abs()
+        d = d.sort_values(by=(metrics[0], sortkey), key=pd.Series.abs, ascending=False)
     else:
-        d['$sortkey'] = d[sortkey]
+        d = d.sort_values(by=(metrics[0], sortkey), ascending=False)
+
+    # Ensure that the columns are grouped by metric (rather than having the
+    # diffs at the end of the line).
+    d = d.reindex(columns=d.columns.levels[0], level=0)
 
-    d = d.sort_values("$sortkey", ascending=False)
-    del d['$sortkey']
     if not show_diff_column:
-        del d['diff']
+        # Remove all diff columns (using level=1 since level 0 is the metric).
+        d.drop(labels='diff', level=1, axis=1, inplace=True)
     dataout = d
     if limit_output:
         # Take 15 topmost elements
         dataout = dataout.head(15)
 
-    if show_diff_column:
-      dataout = add_geomean_row(d, dataout)
-
-    # Turn index into a column so we can format it...
-    dataout.insert(0, 'Program', dataout.index)
-
     formatters = dict()
-    formatters['diff'] = format_diff
+    for m in metrics:
+        formatters[(m, 'diff')] = format_diff
+    # Turn index into a column so we can format it...
+    formatted_program = dataout.index.to_series()
     if shorten_names:
-        drop_prefix, drop_suffix = determine_common_prefix_suffix(dataout.Program)
         def format_name(name, common_prefix, common_suffix):
             name = name[common_prefix:]
             if common_suffix > 0:
                 name = name[:-common_suffix]
             return "%-45s" % truncate(name, 10, 30)
 
 
         def strip_name_fully(name):
             name = name.split('/')[-1]
             if name.endswith('.test'):
                 name = name[:-5]
             return name
 
+        # The to_string formatters argument appears to be ignored for
+        # dtype=object, so transform the program column manually.
         if minimal_names:
-            formatters['Program'] = strip_name_fully
+            formatted_program = formatted_program.map(strip_name_fully)
         else:
-            formatters['Program'] = lambda name: format_name(name, drop_prefix, drop_suffix)
+            drop_prefix, drop_suffix = determine_common_prefix_suffix(formatted_program)
+            formatted_program = formatted_program.map(lambda name: format_name(name, drop_prefix, drop_suffix))
+    dataout.insert(0, 'Program', formatted_program)
+    # Add the geometric mean row after we have formatted the program names
+    # as it will otherwise interfere with common prefix/suffix computation.
+    if show_diff_column:
+        dataout = add_geomean_row(metrics, d, dataout)
+
     def float_format(x):
         if x == '':
             return ''
         return "%6.2f" % (x,)
 
     pd.set_option("display.max_colwidth", 0)
-    out = dataout.to_string(index=False, justify='left',
+    pd.set_option('display.width', 0)
+    # Print an empty value instead of NaN (for the geomean row).
+    out = dataout.to_string(index=False, justify='left', na_rep='',
                             float_format=float_format, formatters=formatters)
     print(out)
     print(d.describe())
 
 def main():
     parser = argparse.ArgumentParser(prog='compare.py')
     parser.add_argument('-a', '--all', action='store_true')
     parser.add_argument('-f', '--full', action='store_true')
if final_size != initial_size:
         print("Remaining: %d" % (final_size,))
 
     # Reduce / add columns
     print("Metric: %s" % (",".join(metrics),))
     if len(metrics) > 0:
         data = data[metrics]
 
     data = data.unstack(level=0)
-    # unstack() gave us a complicated multiindex for the columns, simplify
-    # things by renaming to a simple index.
-    data.columns = [(c[1] if c[1] else c[0]) for c in data.columns.values]
 
-    data = add_diff_column(data)
+    for metric in data.columns.levels[0]:
+        data = add_diff_column(metric, data)
 
     sortkey = 'diff'
+    # TODO: should we still be sorting by diff even if the diff is hidden?
     if len(config.files) == 1:
-        sortkey = data.columns[0]
+        sortkey = data.columns.levels[1][0]
 
     # Print data
     print("")
     shorten_names = not config.full
     limit_output = (not config.all) and (not config.full)
     print_result(data, limit_output, shorten_names, config.minimal_names, config.show_diff, sortkey, config.no_abs_sort)
 
 
 if __name__ == "__main__":
     main()