mtklein | 7ba39cb | 2014-11-24 12:39:59 -0800 | [diff] [blame] | 1 | #!/usr/bin/env python |
| 2 | |
herb | af4edf9 | 2015-07-09 10:50:24 -0700 | [diff] [blame^] | 3 | import argparse |
| 4 | import numpy |
mtklein | 7ba39cb | 2014-11-24 12:39:59 -0800 | [diff] [blame] | 5 | import sys |
| 6 | from scipy.stats import mannwhitneyu |
herb | af4edf9 | 2015-07-09 10:50:24 -0700 | [diff] [blame^] | 7 | from scipy.stats import sem |
mtklein | 7ba39cb | 2014-11-24 12:39:59 -0800 | [diff] [blame] | 8 | |
| 9 | SIGNIFICANCE_THRESHOLD = 0.0001 |
| 10 | |
herb | af4edf9 | 2015-07-09 10:50:24 -0700 | [diff] [blame^] | 11 | parser = argparse.ArgumentParser( |
| 12 | formatter_class=argparse.RawDescriptionHelpFormatter, |
| 13 | description='Compare performance of two runs from nanobench.') |
| 14 | parser.add_argument('--use_means', action='store_true', default=False, |
| 15 | help='Use means to calculate performance ratios.') |
| 16 | parser.add_argument('baseline', help='Baseline file.') |
| 17 | parser.add_argument('experiment', help='Experiment file.') |
| 18 | args = parser.parse_args() |
| 19 | |
mtklein | 7ba39cb | 2014-11-24 12:39:59 -0800 | [diff] [blame] | 20 | a,b = {},{} |
herb | af4edf9 | 2015-07-09 10:50:24 -0700 | [diff] [blame^] | 21 | for (path, d) in [(args.baseline, a), (args.experiment, b)]: |
mtklein | 7ba39cb | 2014-11-24 12:39:59 -0800 | [diff] [blame] | 22 | for line in open(path): |
| 23 | try: |
cdalton | 2c56ba5 | 2015-06-26 13:32:53 -0700 | [diff] [blame] | 24 | tokens = line.split() |
| 25 | if tokens[0] != "Samples:": |
| 26 | continue |
| 27 | samples = tokens[1:-1] |
| 28 | label = tokens[-1] |
mtklein | 7ba39cb | 2014-11-24 12:39:59 -0800 | [diff] [blame] | 29 | d[label] = map(float, samples) |
| 30 | except: |
| 31 | pass |
| 32 | |
| 33 | common = set(a.keys()).intersection(b.keys()) |
| 34 | |
| 35 | ps = [] |
| 36 | for key in common: |
| 37 | _, p = mannwhitneyu(a[key], b[key]) # Non-parametric t-test. Doesn't assume normal dist. |
herb | af4edf9 | 2015-07-09 10:50:24 -0700 | [diff] [blame^] | 38 | if args.use_means: |
| 39 | am, bm = numpy.mean(a[key]), numpy.mean(b[key]) |
| 40 | asem, bsem = sem(a[key]), sem(b[key]) |
| 41 | else: |
| 42 | am, bm = min(a[key]), min(b[key]) |
| 43 | asem, bsem = 0, 0 |
| 44 | ps.append((bm/am, p, key, am, bm, asem, bsem)) |
mtklein | 7ba39cb | 2014-11-24 12:39:59 -0800 | [diff] [blame] | 45 | ps.sort(reverse=True) |
| 46 | |
| 47 | def humanize(ns): |
| 48 | for threshold, suffix in [(1e9, 's'), (1e6, 'ms'), (1e3, 'us'), (1e0, 'ns')]: |
| 49 | if ns > threshold: |
| 50 | return "%.3g%s" % (ns/threshold, suffix) |
| 51 | |
| 52 | maxlen = max(map(len, common)) |
| 53 | |
| 54 | # We print only signficant changes in benchmark timing distribution. |
| 55 | bonferroni = SIGNIFICANCE_THRESHOLD / len(ps) # Adjust for the fact we've run multiple tests. |
herb | af4edf9 | 2015-07-09 10:50:24 -0700 | [diff] [blame^] | 56 | for ratio, p, key, am, bm, asem, bsem in ps: |
mtklein | 7ba39cb | 2014-11-24 12:39:59 -0800 | [diff] [blame] | 57 | if p < bonferroni: |
Mike Klein | 8a84db9 | 2014-11-24 17:44:23 -0500 | [diff] [blame] | 58 | str_ratio = ('%.2gx' if ratio < 1 else '%.3gx') % ratio |
herb | af4edf9 | 2015-07-09 10:50:24 -0700 | [diff] [blame^] | 59 | if args.use_means: |
| 60 | print '%*s\t%6s(%6s) -> %6s(%6s)\t%s' % (maxlen, key, humanize(am), humanize(asem), |
| 61 | humanize(bm), humanize(bsem), str_ratio) |
| 62 | else: |
| 63 | print '%*s\t%6s -> %6s\t%s' % (maxlen, key, humanize(am), humanize(bm), str_ratio) |