Add script comparing v8-perf results

tools/v8-perf-statistics.py

Index: tools/v8-perf-statistics.py
diff --git a/tools/v8-perf-statistics.py b/tools/v8-perf-statistics.py
new file mode 100755
index 0000000000000000000000000000000000000000..5d4bfce4486caa35895bf614d4e136d4a5bd536e
--- /dev/null
+++ b/tools/v8-perf-statistics.py
@@ -0,0 +1,105 @@
+#!/usr/bin/env python
+# Copyright 2016 the V8 project authors. All rights reserved.
+# Use of this source code is governed by a BSD-style license that can be
+# found in the LICENSE file.
+'''
+(You'll need packages python-numpy and python-scipy.)
+
+This is a script for comparing v8-perf statistics for benchmarks run
+on different versions of V8.  Here's an example of how it can be used:
+
+```bash
+cd v8-perf/benchmarks/Octane2.1
+loop -100 ~/vanilla/v8/out/Release/d8 run-some.js -- code-load > vanilla.txt
+loop -100 ~/v8/out/Release/d8 run-some.js -- code-load > modified.txt
+./v8-perf-statistics.py vanilla.txt modified.txt
+```
+
+where I have defined `loop` to be:
+
+```bash
+function loop () {
+  limit=10
+  case "$1" in
+    -*)
+      limit="${1#-}"
+      shift
+  esac
+  for ((i=1; i<=$limit; i++)); do
+    printf "%5d " $i
+    $*
+  done
+}
+```
+
+Multiple files can be supplied to the command line.  The base file is
+taken to be the first, but you can change this, e.g.
+
+```bash
+./v8-perf-statistics.py -2 modified-1.txt vanilla.txt modified-2.txt
+```
+'''
+
+import numpy, scipy, scipy.stats
+from math import sqrt
+
+def statistics(data):
+  N = len(data)
+  average = numpy.average(data)
+  median = numpy.median(data)
+  low = numpy.min(data)
+  high= numpy.max(data)
+  # evaluate sample variance by setting delta degrees of freedom (ddof) to
+  # 1. The degree used in calculations is N - ddof
+  stdev = numpy.std(data, ddof=1)
+  # Get the endpoints of the range that contains 95% of the distribution
+  t_bounds = scipy.stats.t.interval(0.95, N-1)
+  # sum mean to the confidence interval
+  ci = [average + critval * stdev / sqrt(N) for critval in t_bounds]
+  return N, average, median, stdev, low, high, ci
+
+import re, sys
+
+F = []
+T = {}
+base = 0
+
+R = re.compile("\s*(\d+)?\s*([^:]*):\s*(\d+(.\d+)?)\s*")
+
+for arg in sys.argv[1:]:
+  if arg[0] == "-":
+    base = int(arg[1:]) - 1
+    continue
+  F.append(arg)
+  with open(arg, "rt") as f:
+    S = {}
+    for line in f:
+      m = R.match(line)
+      assert m

[Dan Ehrenberg, 2016/03/18 18:40:37]

When I use this script, I replace 'assert m' with 'if not m: continue' so that
it can work on the output of run-perf. A hack for sure, but it works.

[nickie, 2016/03/21 16:09:37]

Done.  I'm fixing this in a better way.

+      key = m.group(2)
+      value = int(m.group(3))
+      if key not in S: S[key] = []
+      S[key].append(value)
+    for key, values in S.items():
+      if key not in T: T[key] = {}
+      T[key][arg] = statistics(values)
+
+base = F[base]
+
+for key in T:
+  print key
+  for arg in F:
+    def compare(i):
+      if arg == base:
+        return T[key][arg][i], "base"
+      diff = T[key][arg][i] - T[key][base][i]
+      perc = 100.0 * diff / T[key][base][i]
+      return T[key][arg][i], "{:+6.2f}".format(perc)
+    print "  {}:".format(arg)
+    print "    samples: {:5}".format(T[key][arg][0])
+    print "    average: {:8.2f} {}".format(*compare(1))
+    print "    median:  {:8.2f} {}".format(*compare(2))
+    print "    stddev:  {:8.2f} {}".format(*compare(3))
+    print "    minimum: {:5}    {}".format(*compare(4))
+    print "    maximum: {:5}    {}".format(*compare(5))
+    print "    CI 95%:  {:8.2f} to {:8.2f}".format(*T[key][arg][6])