contrib/analyze_brprob.py

   1 #!/usr/bin/env python3
   2 #
   3 # Script to analyze results of our branch prediction heuristics
   4 #
   5 # This file is part of GCC.
   6 #
   7 # GCC is free software; you can redistribute it and/or modify it under
   8 # the terms of the GNU General Public License as published by the Free
   9 # Software Foundation; either version 3, or (at your option) any later
  10 # version.
  11 #
  12 # GCC is distributed in the hope that it will be useful, but WITHOUT ANY
  13 # WARRANTY; without even the implied warranty of MERCHANTABILITY or
  14 # FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  15 # for more details.
  16 #
  17 # You should have received a copy of the GNU General Public License
  18 # along with GCC; see the file COPYING3.  If not see
  19 # <http://www.gnu.org/licenses/>.  */
  20 #
  21 #
  22 #
  23 # This script is used to calculate two basic properties of the branch prediction
  24 # heuristics - coverage and hitrate.  Coverage is number of executions
  25 # of a given branch matched by the heuristics and hitrate is probability
  26 # that once branch is predicted as taken it is really taken.
  27 #
  28 # These values are useful to determine the quality of given heuristics.
  29 # Hitrate may be directly used in predict.def.
  30 #
  31 # Usage:
  32 #  Step 1: Compile and profile your program.  You need to use -fprofile-generate
  33 #    flag to get the profiles.
  34 #  Step 2: Make a reference run of the intrumented application.
  35 #  Step 3: Compile the program with collected profile and dump IPA profiles
  36 #          (-fprofile-use -fdump-ipa-profile-details)
  37 #  Step 4: Collect all generated dump files:
  38 #          find . -name '*.profile' | xargs cat > dump_file
  39 #  Step 5: Run the script:
  40 #          ./analyze_brprob.py dump_file
  41 #          and read results.  Basically the following table is printed:
  42 #
  43 # HEURISTICS                           BRANCHES  (REL)  HITRATE                COVERAGE  (REL)
  44 # early return (on trees)                     3   0.2%  35.83% /  93.64%          66360   0.0%
  45 # guess loop iv compare                       8   0.6%  53.35% /  53.73%       11183344   0.0%
  46 # call                                       18   1.4%  31.95% /  69.95%       51880179   0.2%
  47 # loop guard                                 23   1.8%  84.13% /  84.85%    13749065956  42.2%
  48 # opcode values positive (on trees)          42   3.3%  15.71% /  84.81%     6771097902  20.8%
  49 # opcode values nonequal (on trees)         226  17.6%  72.48% /  72.84%      844753864   2.6%
  50 # loop exit                                 231  18.0%  86.97% /  86.98%     8952666897  27.5%
  51 # loop iterations                           239  18.6%  91.10% /  91.10%     3062707264   9.4%
  52 # DS theory                                 281  21.9%  82.08% /  83.39%     7787264075  23.9%
  53 # no prediction                             293  22.9%  46.92% /  70.70%     2293267840   7.0%
  54 # guessed loop iterations                   313  24.4%  76.41% /  76.41%    10782750177  33.1%
  55 # first match                               708  55.2%  82.30% /  82.31%    22489588691  69.0%
  56 # combined                                 1282 100.0%  79.76% /  81.75%    32570120606 100.0%
  57 #
  58 #
  59 #  The heuristics called "first match" is a heuristics used by GCC branch
  60 #  prediction pass and it predicts 55.2% branches correctly. As you can,
  61 #  the heuristics has very good covertage (69.05%).  On the other hand,
  62 #  "opcode values nonequal (on trees)" heuristics has good hirate, but poor
  63 #  coverage.
  64
  65 import sys
  66 import os
  67 import re
  68 import argparse
  69
  70 from math import *
  71
  72 def percentage(a, b):
  73     return 100.0 * a / b
  74
  75 def average(values):
  76     return 1.0 * sum(values) / len(values)
  77
  78 def average_cutoff(values, cut):
  79     l = len(values)
  80     skip = floor(l * cut / 2)
  81     if skip > 0:
  82         values.sort()
  83         values = values[skip:-skip]
  84     return average(values)
  85
  86 def median(values):
  87     values.sort()
  88     return values[int(len(values) / 2)]
  89
  90 class Summary:
  91     def __init__(self, name):
  92         self.name = name
  93         self.branches = 0
  94         self.count = 0
  95         self.hits = 0
  96         self.fits = 0
  97
  98     def get_hitrate(self):
  99         return self.hits / self.count
 100
 101     def count_formatted(self):
 102         v = self.count
 103         for unit in ['','K','M','G','T','P','E','Z']:
 104             if v < 1000:
 105                 return "%3.2f%s" % (v, unit)
 106             v /= 1000.0
 107         return "%.1f%s" % (v, 'Y')
 108
 109 class Profile:
 110     def __init__(self, filename):
 111         self.filename = filename
 112         self.heuristics = {}
 113         self.niter_vector = []
 114
 115     def add(self, name, prediction, count, hits):
 116         if not name in self.heuristics:
 117             self.heuristics[name] = Summary(name)
 118
 119         s = self.heuristics[name]
 120         s.branches += 1
 121         s.count += count
 122         if prediction < 50:
 123             hits = count - hits
 124         s.hits += hits
 125         s.fits += max(hits, count - hits)
 126
 127     def add_loop_niter(self, niter):
 128         if niter > 0:
 129             self.niter_vector.append(niter)
 130
 131     def branches_max(self):
 132         return max([v.branches for k, v in self.heuristics.items()])
 133
 134     def count_max(self):
 135         return max([v.count for k, v in self.heuristics.items()])
 136
 137     def dump(self, sorting):
 138         sorter = lambda x: x[1].branches
 139         if sorting == 'hitrate':
 140             sorter = lambda x: x[1].get_hitrate()
 141         elif sorting == 'coverage':
 142             sorter = lambda x: x[1].count
 143
 144         print('%-40s %8s %6s  %-16s %14s %8s %6s' % ('HEURISTICS', 'BRANCHES', '(REL)',
 145               'HITRATE', 'COVERAGE', 'COVERAGE', '(REL)'))
 146         for (k, v) in sorted(self.heuristics.items(), key = sorter):
 147             print('%-40s %8i %5.1f%% %6.2f%% / %6.2f%% %14i %8s %5.1f%%' %
 148             (k, v.branches, percentage(v.branches, self.branches_max ()),
 149              percentage(v.hits, v.count), percentage(v.fits, v.count),
 150              v.count, v.count_formatted(), percentage(v.count, self.count_max()) ))
 151
 152         if len(self.niter_vector) > 0:
 153             print ('\nLoop count: %d' % len(self.niter_vector)),
 154             print('  avg. # of iter: %.2f' % average(self.niter_vector))
 155             print('  median # of iter: %.2f' % median(self.niter_vector))
 156             for v in [1, 5, 10, 20, 30]:
 157                 cut = 0.01 * v
 158                 print('  avg. (%d%% cutoff) # of iter: %.2f' % (v, average_cutoff(self.niter_vector, cut)))
 159
 160 parser = argparse.ArgumentParser()
 161 parser.add_argument('dump_file', metavar = 'dump_file', help = 'IPA profile dump file')
 162 parser.add_argument('-s', '--sorting', dest = 'sorting', choices = ['branches', 'hitrate', 'coverage'], default = 'branches')
 163
 164 args = parser.parse_args()
 165
 166 profile = Profile(sys.argv[1])
 167 r = re.compile('  (.*) heuristics( of edge [0-9]*->[0-9]*)?( \\(.*\\))?: (.*)%.*exec ([0-9]*) hit ([0-9]*)')
 168 loop_niter_str = ';;  profile-based iteration count: '
 169 for l in open(args.dump_file).readlines():
 170     m = r.match(l)
 171     if m != None and m.group(3) == None:
 172         name = m.group(1)
 173         prediction = float(m.group(4))
 174         count = int(m.group(5))
 175         hits = int(m.group(6))
 176
 177         profile.add(name, prediction, count, hits)
 178     elif l.startswith(loop_niter_str):
 179         v = int(l[len(loop_niter_str):])
 180         profile.add_loop_niter(v)
 181
 182 profile.dump(args.sorting)