blob: 82f85d5fd91ab724e5c3aa6e2f1e65508fa10f79 [file] [log] [blame]
#!/usr/bin/env python
import argparse
import numpy
import sys
from scipy.stats import mannwhitneyu
from scipy.stats import sem
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())
ps = []
for key in common:
_, p = mannwhitneyu(a[key], b[key]) # Non-parametric t-test. Doesn't assume normal dist.
if args.use_means:
am, bm = numpy.mean(a[key]), numpy.mean(b[key])
asem, bsem = sem(a[key]), sem(b[key])
else:
am, bm = min(a[key]), min(b[key])
asem, bsem = 0, 0
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)