skia2/bench/gen_bench_ranges.py

#!/usr/bin/env python
# Copyright (c) 2012 The Chromium Authors. All rights reserved.
# Use of this source code is governed by a BSD-style license that can be found
# in the LICENSE file.

""" Analyze recent SkPicture or Microbench data, and output suggested ranges.

The outputs can be edited and pasted to bench_expectations.txt to trigger
buildbot alerts if the actual benches are out of range. Details are documented
in the .txt file.

Currently the easiest way to batch update bench_expectations.txt is to delete
all bench lines, run this script, and redirect outputs (">>") to be added to the
.txt file.
You can also just manually change a few lines of interest, of course.

Note: since input data are stored in Google Storage, you will need to set up
the corresponding library.
See http://developers.google.com/storage/docs/gspythonlibrary for details.
"""

__author__ = 'bensong@google.com (Ben Chen)'

import bench_util
import boto
import cStringIO
import optparse
import re
import shutil

from oauth2_plugin import oauth2_plugin


# Ratios for calculating suggested picture bench upper and lower bounds.
BENCH_UB = 1.1  # Allow for 10% room for normal variance on the up side.
BENCH_LB = 0.9

# Further allow for a fixed amount of noise. This is especially useful for
# benches of smaller absolute value. Keeping this value small will not affect
# performance tunings.
BENCH_ALLOWED_NOISE = 10

# Name prefix for benchmark builders.
BENCH_BUILDER_PREFIX = 'Perf-'

# List of platforms to track. Feel free to change it to meet your needs.
PLATFORMS = ['Perf-Mac10.8-MacMini4.1-GeForce320M-x86-Release',
             'Perf-Android-Nexus7-Tegra3-Arm7-Release',
             'Perf-Ubuntu12-ShuttleA-ATI5770-x86-Release',
             'Perf-Win7-ShuttleA-HD2000-x86-Release',
            ]

# Filter for configs of no interest. They are old config names replaced by more
# specific ones.
CONFIGS_TO_FILTER = ['gpu', 'raster']

# Template for gsutil uri.
GOOGLE_STORAGE_URI_SCHEME = 'gs'
URI_BUCKET = 'chromium-skia-gm'

# Constants for optparse.
USAGE_STRING = 'USAGE: %s [options]'
HOWTO_STRING = """
Feel free to revise PLATFORMS for your own needs. The default is the most common
combination that we care most about. Platforms that did not run bench_pictures
or benchmain in the given revision range will not have corresponding outputs.
Please check http://go/skpbench to choose a range that fits your needs.
BENCH_UB, BENCH_LB and BENCH_ALLOWED_NOISE can be changed to expand or narrow
the permitted bench ranges without triggering buidbot alerts.
"""
HELP_STRING = """
Outputs expectation picture bench ranges for the latest revisions for the given
revision range. For instance, --rev_range=6000:6000 will return only bench
ranges for the bots that ran benches at rev 6000; --rev-range=6000:7000
may have multiple bench data points for each bench configuration, and the code
returns bench data for the latest revision of all available (closer to 7000).
""" + HOWTO_STRING

OPTION_REVISION_RANGE = '--rev-range'
OPTION_REVISION_RANGE_SHORT = '-r'
# Bench representation algorithm flag.
OPTION_REPRESENTATION_ALG = '--algorithm'
OPTION_REPRESENTATION_ALG_SHORT = '-a'
# Bench type to examine. Either 'micro' or 'skp'.
OPTION_BENCH_TYPE = '--bench-type'
OPTION_BENCH_TYPE_SHORT = '-b'

# List of valid bench types.
BENCH_TYPES = ['micro', 'skp']
# List of valid representation algorithms.
REPRESENTATION_ALGS = ['avg', 'min', 'med', '25th']

def OutputBenchExpectations(bench_type, rev_min, rev_max, representation_alg):
  """Reads bench data from google storage, and outputs expectations.

  Ignores data with revisions outside [rev_min, rev_max] integer range. For
  bench data with multiple revisions, we use higher revisions to calculate
  expected bench values.
  bench_type is either 'micro' or 'skp', according to the flag '-b'.
  Uses the provided representation_alg for calculating bench representations.
  """
  if bench_type not in BENCH_TYPES:
    raise Exception('Not valid bench_type! (%s)' % BENCH_TYPES)
  expectation_dic = {}
  uri = boto.storage_uri(URI_BUCKET, GOOGLE_STORAGE_URI_SCHEME)
  for obj in uri.get_bucket():
    # Filters out non-bench files.
    if ((not obj.name.startswith('perfdata/%s' % BENCH_BUILDER_PREFIX) and
         not obj.name.startswith(
             'playback/perfdata/%s' % BENCH_BUILDER_PREFIX)) or
         obj.name.find('_data') < 0):
      continue
    if ((bench_type == 'micro' and obj.name.find('_data_skp_') > 0) or
        (bench_type == 'skp' and obj.name.find('_skp_') < 0)):
      # Skips wrong bench type.
      continue
    # Ignores uninterested platforms.
    platform = obj.name.split('/')[1]
    if not platform.startswith(BENCH_BUILDER_PREFIX):
      platform = obj.name.split('/')[2]
    if not platform.startswith(BENCH_BUILDER_PREFIX):
      continue  # Ignores non-platform object
    if platform not in PLATFORMS:
      continue
    # Filters by revision.
    to_filter = True
    for rev in range(rev_min, rev_max + 1):
      if '_r%s_' % rev in obj.name:
        to_filter = False
        break
    if to_filter:
      continue
    contents = cStringIO.StringIO()
    obj.get_file(contents)
    for point in bench_util.parse('', contents.getvalue().split('\n'),
                                  representation_alg):
      if point.config in CONFIGS_TO_FILTER:
        continue

      key = '%s_%s_%s,%s-%s' % (point.bench, point.config, point.time_type,
                                platform, representation_alg)
      # It is fine to have later revisions overwrite earlier benches, since we
      # only use the latest bench within revision range to set expectations.
      expectation_dic[key] = point.time
  keys = expectation_dic.keys()
  keys.sort()
  for key in keys:
    bench_val = expectation_dic[key]
    # Prints out expectation lines.
    print '%s,%.3f,%.3f,%.3f' % (key, bench_val,
                                 bench_val * BENCH_LB - BENCH_ALLOWED_NOISE,
                                 bench_val * BENCH_UB + BENCH_ALLOWED_NOISE)

def main():
  """Parses flags and outputs expected Skia bench results."""
  parser = optparse.OptionParser(USAGE_STRING % '%prog' + HELP_STRING)
  parser.add_option(OPTION_REVISION_RANGE_SHORT, OPTION_REVISION_RANGE,
      dest='rev_range',
      help='(Mandatory) revision range separated by ":", e.g., 6000:6005')
  parser.add_option(OPTION_BENCH_TYPE_SHORT, OPTION_BENCH_TYPE,
      dest='bench_type', default='skp',
      help=('Bench type, either "skp" or "micro". Default to "skp".'))
  parser.add_option(OPTION_REPRESENTATION_ALG_SHORT, OPTION_REPRESENTATION_ALG,
      dest='alg', default='25th',
      help=('Bench representation algorithm. One of '
            '%s. Default to "25th".' % str(REPRESENTATION_ALGS)))
  (options, args) = parser.parse_args()
  if options.rev_range:
    range_match = re.search('(\d+)\:(\d+)', options.rev_range)
    if not range_match:
      parser.error('Wrong format for rev-range [%s]' % options.rev_range)
    else:
      rev_min = int(range_match.group(1))
      rev_max = int(range_match.group(2))
      OutputBenchExpectations(options.bench_type, rev_min, rev_max, options.alg)
  else:
    parser.error('Please provide mandatory flag %s' % OPTION_REVISION_RANGE)


if '__main__' == __name__:
  main()