12 SIGNIFICANCE_THRESHOLD = 0.0001
14 parser = argparse.ArgumentParser(
15 formatter_class=argparse.RawDescriptionHelpFormatter,
16 description='Compare performance of two runs from nanobench.')
17 parser.add_argument('--use_means', action='store_true', default=False,
18 help='Use means to calculate performance ratios.')
19 parser.add_argument('baseline', help='Baseline file.')
20 parser.add_argument('experiment', help='Experiment file.')
21 args = parser.parse_args()
24 for (path, d) in [(args.baseline, a), (args.experiment, b)]:
25 for line in open(path):
28 if tokens[0] != "Samples:":
30 samples = tokens[1:-1]
32 d[label] = map(float, samples)
36 common = set(a.keys()).intersection(b.keys())
39 return sum(xs) / len(xs)
43 p, asem, bsem = 0, 0, 0
44 m = mean if args.use_means else min
45 am, bm = m(a[key]), m(b[key])
47 _, p = scipy.stats.mannwhitneyu(a[key], b[key])
48 asem, bsem = scipy.stats.sem(a[key]), scipy.stats.sem(b[key])
49 ps.append((bm/am, p, key, am, bm, asem, bsem))
53 for threshold, suffix in [(1e9, 's'), (1e6, 'ms'), (1e3, 'us'), (1e0, 'ns')]:
55 return "%.3g%s" % (ns/threshold, suffix)
57 maxlen = max(map(len, common))
59 # We print only signficant changes in benchmark timing distribution.
60 bonferroni = SIGNIFICANCE_THRESHOLD / len(ps) # Adjust for the fact we've run multiple tests.
61 for ratio, p, key, am, bm, asem, bsem in ps:
63 str_ratio = ('%.2gx' if ratio < 1 else '%.3gx') % ratio
65 print '%*s\t%6s(%6s) -> %6s(%6s)\t%s' % (maxlen, key, humanize(am), humanize(asem),
66 humanize(bm), humanize(bsem), str_ratio)
68 print '%*s\t%6s -> %6s\t%s' % (maxlen, key, humanize(am), humanize(bm), str_ratio)