blob: 794cb724e0713f30e554d3e1099559f6edeceef5 [file] [log] [blame]
mtklein7ba39cb2014-11-24 12:39:59 -08001#!/usr/bin/env python
2
herbaf4edf92015-07-09 10:50:24 -07003import argparse
mtklein7ba39cb2014-11-24 12:39:59 -08004import sys
mtklein24a592c2015-10-28 09:45:44 -07005
6have_scipy = True
7try:
8 import scipy.stats
9except:
10 have_scipy = False
mtklein7ba39cb2014-11-24 12:39:59 -080011
12SIGNIFICANCE_THRESHOLD = 0.0001
13
herbaf4edf92015-07-09 10:50:24 -070014parser = argparse.ArgumentParser(
15 formatter_class=argparse.RawDescriptionHelpFormatter,
16 description='Compare performance of two runs from nanobench.')
17parser.add_argument('--use_means', action='store_true', default=False,
18 help='Use means to calculate performance ratios.')
19parser.add_argument('baseline', help='Baseline file.')
20parser.add_argument('experiment', help='Experiment file.')
21args = parser.parse_args()
22
mtklein7ba39cb2014-11-24 12:39:59 -080023a,b = {},{}
herbaf4edf92015-07-09 10:50:24 -070024for (path, d) in [(args.baseline, a), (args.experiment, b)]:
mtklein7ba39cb2014-11-24 12:39:59 -080025 for line in open(path):
26 try:
cdalton2c56ba52015-06-26 13:32:53 -070027 tokens = line.split()
28 if tokens[0] != "Samples:":
29 continue
30 samples = tokens[1:-1]
31 label = tokens[-1]
mtklein7ba39cb2014-11-24 12:39:59 -080032 d[label] = map(float, samples)
33 except:
34 pass
35
36common = set(a.keys()).intersection(b.keys())
37
mtklein24a592c2015-10-28 09:45:44 -070038def mean(xs):
39 return sum(xs) / len(xs)
40
mtklein7ba39cb2014-11-24 12:39:59 -080041ps = []
42for key in common:
mtklein24a592c2015-10-28 09:45:44 -070043 p, asem, bsem = 0, 0, 0
44 m = mean if args.use_means else min
45 am, bm = m(a[key]), m(b[key])
46 if have_scipy:
47 _, p = scipy.stats.mannwhitneyu(a[key], b[key])
herbdad837a2015-11-06 10:35:37 -080048 asem, bsem = scipy.stats.sem(a[key]), scipy.stats.sem(b[key])
herbaf4edf92015-07-09 10:50:24 -070049 ps.append((bm/am, p, key, am, bm, asem, bsem))
mtklein7ba39cb2014-11-24 12:39:59 -080050ps.sort(reverse=True)
51
52def humanize(ns):
53 for threshold, suffix in [(1e9, 's'), (1e6, 'ms'), (1e3, 'us'), (1e0, 'ns')]:
54 if ns > threshold:
55 return "%.3g%s" % (ns/threshold, suffix)
56
57maxlen = max(map(len, common))
58
59# We print only signficant changes in benchmark timing distribution.
60bonferroni = SIGNIFICANCE_THRESHOLD / len(ps) # Adjust for the fact we've run multiple tests.
herbaf4edf92015-07-09 10:50:24 -070061for ratio, p, key, am, bm, asem, bsem in ps:
mtklein7ba39cb2014-11-24 12:39:59 -080062 if p < bonferroni:
Mike Klein8a84db92014-11-24 17:44:23 -050063 str_ratio = ('%.2gx' if ratio < 1 else '%.3gx') % ratio
herbaf4edf92015-07-09 10:50:24 -070064 if args.use_means:
65 print '%*s\t%6s(%6s) -> %6s(%6s)\t%s' % (maxlen, key, humanize(am), humanize(asem),
66 humanize(bm), humanize(bsem), str_ratio)
67 else:
68 print '%*s\t%6s -> %6s\t%s' % (maxlen, key, humanize(am), humanize(bm), str_ratio)