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 counter_aggregates = set(['combined', 'first match', 'DS theory',
  73     'no prediction'])
  74
  75 def percentage(a, b):
  76     return 100.0 * a / b
  77
  78 def average(values):
  79     return 1.0 * sum(values) / len(values)
  80
  81 def average_cutoff(values, cut):
  82     l = len(values)
  83     skip = floor(l * cut / 2)
  84     if skip > 0:
  85         values.sort()
  86         values = values[skip:-skip]
  87     return average(values)
  88
  89 def median(values):
  90     values.sort()
  91     return values[int(len(values) / 2)]
  92
  93 class Summary:
  94     def __init__(self, name):
  95         self.name = name
  96         self.branches = 0
  97         self.successfull_branches = 0
  98         self.count = 0
  99         self.hits = 0
 100         self.fits = 0
 101
 102     def get_hitrate(self):
 103         return 100.0 * self.hits / self.count
 104
 105     def get_branch_hitrate(self):
 106         return 100.0 * self.successfull_branches / self.branches
 107
 108     def count_formatted(self):
 109         v = self.count
 110         for unit in ['','K','M','G','T','P','E','Z']:
 111             if v < 1000:
 112                 return "%3.2f%s" % (v, unit)
 113             v /= 1000.0
 114         return "%.1f%s" % (v, 'Y')
 115
 116     def print(self, branches_max, count_max):
 117         print('%-40s %8i %5.1f%% %11.2f%% %7.2f%% / %6.2f%% %14i %8s %5.1f%%' %
 118             (self.name, self.branches,
 119                 percentage(self.branches, branches_max),
 120                 self.get_branch_hitrate(),
 121                 self.get_hitrate(),
 122                 percentage(self.fits, self.count),
 123                 self.count, self.count_formatted(),
 124                 percentage(self.count, count_max)))
 125
 126 class Profile:
 127     def __init__(self, filename):
 128         self.filename = filename
 129         self.heuristics = {}
 130         self.niter_vector = []
 131
 132     def add(self, name, prediction, count, hits):
 133         if not name in self.heuristics:
 134             self.heuristics[name] = Summary(name)
 135
 136         s = self.heuristics[name]
 137         s.branches += 1
 138
 139         s.count += count
 140         if prediction < 50:
 141             hits = count - hits
 142         remaining = count - hits
 143         if hits >= remaining:
 144             s.successfull_branches += 1
 145
 146         s.hits += hits
 147         s.fits += max(hits, remaining)
 148
 149     def add_loop_niter(self, niter):
 150         if niter > 0:
 151             self.niter_vector.append(niter)
 152
 153     def branches_max(self):
 154         return max([v.branches for k, v in self.heuristics.items()])
 155
 156     def count_max(self):
 157         return max([v.count for k, v in self.heuristics.items()])
 158
 159     def print_group(self, sorting, group_name, heuristics):
 160         count_max = self.count_max()
 161         branches_max = self.branches_max()
 162
 163         sorter = lambda x: x.branches
 164         if sorting == 'branch-hitrate':
 165             sorter = lambda x: x.get_branch_hitrate()
 166         elif sorting == 'hitrate':
 167             sorter = lambda x: x.get_hitrate()
 168         elif sorting == 'coverage':
 169             sorter = lambda x: x.count
 170         elif sorting == 'name':
 171             sorter = lambda x: x.name.lower()
 172
 173         print('%-40s %8s %6s %12s %18s %14s %8s %6s' %
 174             ('HEURISTICS', 'BRANCHES', '(REL)',
 175             'BR. HITRATE', 'HITRATE', 'COVERAGE', 'COVERAGE', '(REL)'))
 176         for h in sorted(heuristics, key = sorter):
 177             h.print(branches_max, count_max)
 178
 179     def dump(self, sorting):
 180         heuristics = self.heuristics.values()
 181         if len(heuristics) == 0:
 182             print('No heuristics available')
 183             return
 184
 185         special = list(filter(lambda x: x.name in counter_aggregates,
 186             heuristics))
 187         normal = list(filter(lambda x: x.name not in counter_aggregates,
 188             heuristics))
 189
 190         self.print_group(sorting, 'HEURISTICS', normal)
 191         print()
 192         self.print_group(sorting, 'HEURISTIC AGGREGATES', special)
 193
 194         if len(self.niter_vector) > 0:
 195             print ('\nLoop count: %d' % len(self.niter_vector)),
 196             print('  avg. # of iter: %.2f' % average(self.niter_vector))
 197             print('  median # of iter: %.2f' % median(self.niter_vector))
 198             for v in [1, 5, 10, 20, 30]:
 199                 cut = 0.01 * v
 200                 print('  avg. (%d%% cutoff) # of iter: %.2f'
 201                     % (v, average_cutoff(self.niter_vector, cut)))
 202
 203 parser = argparse.ArgumentParser()
 204 parser.add_argument('dump_file', metavar = 'dump_file',
 205     help = 'IPA profile dump file')
 206 parser.add_argument('-s', '--sorting', dest = 'sorting',
 207     choices = ['branches', 'branch-hitrate', 'hitrate', 'coverage', 'name'],
 208     default = 'branches')
 209
 210 args = parser.parse_args()
 211
 212 profile = Profile(sys.argv[1])
 213 r = re.compile('  (.*) heuristics( of edge [0-9]*->[0-9]*)?( \\(.*\\))?: (.*)%.*exec ([0-9]*) hit ([0-9]*)')
 214 loop_niter_str = ';;  profile-based iteration count: '
 215 for l in open(args.dump_file).readlines():
 216     m = r.match(l)
 217     if m != None and m.group(3) == None:
 218         name = m.group(1)
 219         prediction = float(m.group(4))
 220         count = int(m.group(5))
 221         hits = int(m.group(6))
 222
 223         profile.add(name, prediction, count, hits)
 224     elif l.startswith(loop_niter_str):
 225         v = int(l[len(loop_niter_str):])
 226         profile.add_loop_niter(v)
 227
 228 profile.dump(args.sorting)