1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
| #!/usr/bin/env python
|
| import argparse
| import sys
|
| have_scipy = True
| try:
| import scipy.stats
| except:
| have_scipy = False
|
| SIGNIFICANCE_THRESHOLD = 0.0001
|
| parser = argparse.ArgumentParser(
| formatter_class=argparse.RawDescriptionHelpFormatter,
| description='Compare performance of two runs from nanobench.')
| parser.add_argument('--use_means', action='store_true', default=False,
| help='Use means to calculate performance ratios.')
| parser.add_argument('baseline', help='Baseline file.')
| parser.add_argument('experiment', help='Experiment file.')
| args = parser.parse_args()
|
| a,b = {},{}
| for (path, d) in [(args.baseline, a), (args.experiment, b)]:
| for line in open(path):
| try:
| tokens = line.split()
| if tokens[0] != "Samples:":
| continue
| samples = tokens[1:-1]
| label = tokens[-1]
| d[label] = map(float, samples)
| except:
| pass
|
| common = set(a.keys()).intersection(b.keys())
|
| def mean(xs):
| return sum(xs) / len(xs)
|
| ps = []
| for key in common:
| p, asem, bsem = 0, 0, 0
| m = mean if args.use_means else min
| am, bm = m(a[key]), m(b[key])
| if have_scipy:
| _, p = scipy.stats.mannwhitneyu(a[key], b[key])
| asem, bsem = scipy.stats.sem(a[key]), scipy.stats.sem(b[key])
| ps.append((bm/am, p, key, am, bm, asem, bsem))
| ps.sort(reverse=True)
|
| def humanize(ns):
| for threshold, suffix in [(1e9, 's'), (1e6, 'ms'), (1e3, 'us'), (1e0, 'ns')]:
| if ns > threshold:
| return "%.3g%s" % (ns/threshold, suffix)
|
| maxlen = max(map(len, common))
|
| # We print only signficant changes in benchmark timing distribution.
| bonferroni = SIGNIFICANCE_THRESHOLD / len(ps) # Adjust for the fact we've run multiple tests.
| for ratio, p, key, am, bm, asem, bsem in ps:
| if p < bonferroni:
| str_ratio = ('%.2gx' if ratio < 1 else '%.3gx') % ratio
| if args.use_means:
| print '%*s\t%6s(%6s) -> %6s(%6s)\t%s' % (maxlen, key, humanize(am), humanize(asem),
| humanize(bm), humanize(bsem), str_ratio)
| else:
| print '%*s\t%6s -> %6s\t%s' % (maxlen, key, humanize(am), humanize(bm), str_ratio)
|
|
