1
0
mirror of https://github.com/ARM-software/workload-automation.git synced 2024-10-06 19:01:15 +01:00
workload-automation/wlauto/instrumentation/fps/__init__.py
2016-11-07 17:39:51 +00:00

410 lines
20 KiB
Python
Executable File

# Copyright 2013-2015 ARM Limited
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# pylint: disable=W0613,E1101
from __future__ import division
import os
import sys
import time
import csv
import shutil
import threading
import errno
import tempfile
import collections
import re
from distutils.version import LooseVersion
try:
import pandas as pd
except ImportError:
pd = None
from wlauto import Instrument, Parameter, IterationResult
from wlauto.instrumentation import instrument_is_installed
from wlauto.exceptions import (InstrumentError, WorkerThreadError, ConfigError,
DeviceNotRespondingError, TimeoutError)
from wlauto.utils.types import boolean, numeric
from wlauto.utils.fps import FpsProcessor, SurfaceFlingerFrame, GfxInfoFrame, GFXINFO_EXEMPT
VSYNC_INTERVAL = 16666667
PAUSE_LATENCY = 20
EPSYLON = 0.0001
class FpsInstrument(Instrument):
name = 'fps'
description = """
Measures Frames Per Second (FPS) and associated metrics for a workload.
.. note:: This instrument depends on pandas Python library (which is not part of standard
WA dependencies), so you will need to install that first, before you can use it.
Android L and below use SurfaceFlinger to calculate the FPS data.
Android M and above use gfxinfo to calculate the FPS data.
SurfaceFlinger:
The view is specified by the workload as ``view`` attribute. This defaults
to ``'SurfaceView'`` for game workloads, and ``None`` for non-game
workloads (as for them FPS mesurement usually doesn't make sense).
Individual workloads may override this.
gfxinfo:
The view is specified by the workload as ``package`` attribute.
This is because gfxinfo already processes for all views in a package.
This instrument adds four metrics to the results:
:FPS: Frames Per Second. This is the frame rate of the workload.
:frame_count: The total number of frames rendered during the execution of
the workload.
:janks: The number of "janks" that occured during execution of the
workload. Janks are sudden shifts in frame rate. They result
in a "stuttery" UI. See http://jankfree.org/jank-busters-io
:not_at_vsync: The number of frames that did not render in a single
vsync cycle.
"""
supported_platforms = ['android']
parameters = [
Parameter('drop_threshold', kind=numeric, default=5,
description='Data points below this FPS will be dropped as they '
'do not constitute "real" gameplay. The assumption '
'being that while actually running, the FPS in the '
'game will not drop below X frames per second, '
'except on loading screens, menus, etc, which '
'should not contribute to FPS calculation. '),
Parameter('keep_raw', kind=boolean, default=False,
description='If set to ``True``, this will keep the raw dumpsys output '
'in the results directory (this is maily used for debugging) '
'Note: frames.csv with collected frames data will always be '
'generated regardless of this setting.'),
Parameter('generate_csv', kind=boolean, default=True,
description='If set to ``True``, this will produce temporal fps data '
'in the results directory, in a file named fps.csv '
'Note: fps data will appear as discrete step-like values '
'in order to produce a more meainingfull representation,'
'a rolling mean can be applied.'),
Parameter('crash_check', kind=boolean, default=True,
description="""
Specifies wither the instrument should check for crashed content by examining
frame data. If this is set, ``execution_time`` instrument must also be installed.
The check is performed by using the measured FPS and exection time to estimate the expected
frames cound and comparing that against the measured frames count. The the ratio of
measured/expected is too low, then it is assumed that the content has crashed part way
during the run. What is "too low" is determined by ``crash_threshold``.
.. note:: This is not 100\% fool-proof. If the crash occurs sufficiently close to
workload's termination, it may not be detected. If this is expected, the
threshold may be adjusted up to compensate.
"""),
Parameter('crash_threshold', kind=float, default=0.7,
description="""
Specifies the threshold used to decided whether a measured/expected frames ration indicates
a content crash. E.g. a value of ``0.75`` means the number of actual frames counted is a
quarter lower than expected, it will treated as a content crash.
"""),
Parameter('dumpsys_period', kind=float, default=2, constraint=lambda x: x > 0,
description="""
Specifies the time period between calls to ``dumpsys SurfaceFlinger --latency`` in
seconds when collecting frame data. Using a lower value improves the granularity
of timings when recording actions that take a short time to complete. Note, this
will produce duplicate frame data in the raw dumpsys output, however, this is
filtered out in frames.csv. It may also affect the overall load on the system.
The default value of 2 seconds corresponds with the NUM_FRAME_RECORDS in
android/services/surfaceflinger/FrameTracker.h (as of the time of writing
currently 128) and a frame rate of 60 fps that is applicable to most devices.
"""),
]
def __init__(self, device, **kwargs):
super(FpsInstrument, self).__init__(device, **kwargs)
self.collector = None
self.outfile = None
self.fps_outfile = None
self.is_enabled = True
self.fps_method = ''
def validate(self):
if not pd or LooseVersion(pd.__version__) < LooseVersion('0.13.1'):
message = ('fps instrument requires pandas Python package (version 0.13.1 or higher) to be installed.\n'
'You can install it with pip, e.g. "sudo pip install pandas"')
raise InstrumentError(message)
if self.crash_check and not instrument_is_installed('execution_time'):
raise ConfigError('execution_time instrument must be installed in order to check for content crash.')
def setup(self, context):
workload = context.workload
if hasattr(workload, 'view'):
self.fps_outfile = os.path.join(context.output_directory, 'fps.csv')
self.outfile = os.path.join(context.output_directory, 'frames.csv')
# Android M brings a new method of collecting FPS data
if self.device.get_sdk_version() >= 23:
# gfxinfo takes in the package name rather than a single view/activity
# so there is no 'list_command' to run and compare against a list of
# views/activities. Additionally, clearing the stats requires the package
# so we need to clear for every package in the workload.
# Usually there is only one package, but some workloads may run multiple
# packages so each one must be reset before continuing
self.fps_method = 'gfxinfo'
runcmd = 'dumpsys gfxinfo {} framestats'
lstcmd = None
params = workload.package
params = [params] if isinstance(params, basestring) else params
for pkg in params:
self.device.execute('dumpsys gfxinfo {} reset'.format(pkg))
else:
self.fps_method = 'surfaceflinger'
runcmd = 'dumpsys SurfaceFlinger --latency {}'
lstcmd = 'dumpsys SurfaceFlinger --list'
params = workload.view
self.device.execute('dumpsys SurfaceFlinger --latency-clear ')
self.collector = LatencyCollector(self.outfile, self.device, params or '',
self.keep_raw, self.logger, self.dumpsys_period,
runcmd, lstcmd, self.fps_method)
else:
self.logger.debug('Workload does not contain a view; disabling...')
self.is_enabled = False
def start(self, context):
if self.is_enabled:
self.logger.debug('Starting Frame Statistics collection...')
self.collector.start()
def stop(self, context):
if self.is_enabled and self.collector.is_alive():
self.logger.debug('Stopping Frame Statistics collection...')
self.collector.stop()
def update_result(self, context):
if self.is_enabled:
fps, frame_count, janks, not_at_vsync = float('nan'), 0, 0, 0
p90, p95, p99 = [float('nan')] * 3
data = pd.read_csv(self.outfile)
if not data.empty: # pylint: disable=maybe-no-member
# gfxinfo method has an additional file generated that contains statistics
stats_file = None
if self.fps_method == 'gfxinfo':
stats_file = os.path.join(os.path.dirname(self.outfile), 'gfxinfo.csv')
fp = FpsProcessor(data, extra_data=stats_file)
per_frame_fps, metrics = fp.process(self.collector.refresh_period, self.drop_threshold)
fps, frame_count, janks, not_at_vsync = metrics
if self.generate_csv:
per_frame_fps.to_csv(self.fps_outfile, index=False, header=True)
context.add_artifact('fps', path='fps.csv', kind='data')
p90, p95, p99 = fp.percentiles()
context.result.add_metric('FPS', fps)
context.result.add_metric('frame_count', frame_count)
context.result.add_metric('janks', janks, lower_is_better=True)
context.result.add_metric('not_at_vsync', not_at_vsync, lower_is_better=True)
context.result.add_metric('frame_time_90percentile', p90, 'ms', lower_is_better=True)
context.result.add_metric('frame_time_95percentile', p95, 'ms', lower_is_better=True)
context.result.add_metric('frame_time_99percentile', p99, 'ms', lower_is_better=True)
def slow_update_result(self, context):
result = context.result
if self.crash_check and result.has_metric('execution_time'):
self.logger.debug('Checking for crashed content.')
exec_time = result['execution_time'].value
fps = result['FPS'].value
frames = result['frame_count'].value
if all([exec_time, fps, frames]):
expected_frames = fps * exec_time
ratio = frames / expected_frames
self.logger.debug('actual/expected frames: {:.2}'.format(ratio))
if ratio < self.crash_threshold:
self.logger.error('Content for {} appears to have crashed.'.format(context.spec.label))
result.status = IterationResult.FAILED
result.add_event('Content crash detected (actual/expected frames: {:.2}).'.format(ratio))
class LatencyCollector(threading.Thread):
# Note: the size of the frames buffer for a particular surface is defined
# by NUM_FRAME_RECORDS inside android/services/surfaceflinger/FrameTracker.h.
# At the time of writing, this was hard-coded to 128. So at 60 fps
# (and there is no reason to go above that, as it matches vsync rate
# on pretty much all phones), there is just over 2 seconds' worth of
# frames in there. Hence the default sleep time of 2 seconds between dumps.
def __init__(self, outfile, device, activities, keep_raw, logger, dumpsys_period,
run_command, list_command, fps_method):
super(LatencyCollector, self).__init__()
self.outfile = outfile
self.device = device
self.keep_raw = keep_raw
self.logger = logger
self.dumpsys_period = dumpsys_period
self.stop_signal = threading.Event()
self.frames = []
self.last_ready_time = 0
self.refresh_period = VSYNC_INTERVAL
self.drop_threshold = self.refresh_period * 1000
self.exc = None
self.unresponsive_count = 0
if isinstance(activities, basestring):
activities = [activities]
self.activities = activities
self.command_template = run_command
self.list_command = list_command
self.fps_method = fps_method
# Based on the fps_method, setup the header for the csv,
# and set the process_trace_line function accordingly
if fps_method == 'surfaceflinger':
self.header = SurfaceFlingerFrame._fields
self.process_trace_line = self._process_surfaceflinger_line
else:
self.header = GfxInfoFrame._fields
self.process_trace_line = self._process_gfxinfo_line
self.re_frame = re.compile('([0-9]+,)+')
self.re_stats = re.compile('.*(percentile|frames|Number).*')
# Create a template summary text block that matches what gfxinfo gives after a reset
# - 133 is the default ms value for percentiles after reset
self.summary = collections.OrderedDict((('Total frames rendered', 0),
('Janky frames', 0),
('90th percentile', 133),
('95th percentile', 133),
('99th percentile', 133),
('Number Missed Vsync', 0),
('Number High input latency', 0),
('Number Slow UI thread', 0),
('Number Slow bitmap uploads', 0),
('Number Slow issue draw commands', 0)))
def run(self):
try:
self.logger.debug('Frame Statistics collection started. Method: ' + self.fps_method)
self.stop_signal.clear()
fd, temp_file = tempfile.mkstemp()
self.logger.debug('temp file: {}'.format(temp_file))
wfh = os.fdopen(fd, 'wb')
try:
view_list = self.activities
while not self.stop_signal.is_set():
# If a list_command is provided, set the view_list to be its output
# Then check for each activity in this list and if there is a match,
# process the output. If no command is provided, then always process.
if self.list_command:
view_list = self.device.execute(self.list_command).split()
for activity in self.activities:
if activity in view_list:
wfh.write(self.device.execute(self.command_template.format(activity)))
time.sleep(self.dumpsys_period)
finally:
wfh.close()
# TODO: this can happen after the run during results processing
with open(temp_file) as fh:
text = fh.read().replace('\r\n', '\n').replace('\r', '\n')
for line in text.split('\n'):
line = line.strip()
if line:
self.process_trace_line(line)
if self.keep_raw:
raw_file = os.path.join(os.path.dirname(self.outfile), self.fps_method + '.raw')
shutil.copy(temp_file, raw_file)
os.unlink(temp_file)
except (DeviceNotRespondingError, TimeoutError): # pylint: disable=W0703
raise
except Exception, e: # pylint: disable=W0703
self.logger.warning('Exception on collector thread: {}({})'.format(e.__class__.__name__, e))
self.exc = WorkerThreadError(self.name, sys.exc_info())
self.logger.debug('Frame Statistics collection stopped.')
with open(self.outfile, 'w') as wfh:
writer = csv.writer(wfh)
writer.writerow(self.header)
writer.writerows(self.frames)
self.logger.debug('Frames data written.')
# gfxinfo outputs its own summary statistics for the run.
# No point calculating those from the raw data, so store in its own file for later use.
if self.fps_method == 'gfxinfo':
stats_file = os.path.join(os.path.dirname(self.outfile), 'gfxinfo.csv')
with open(stats_file, 'w') as wfh:
writer = csv.writer(wfh)
writer.writerows(zip(self.summary.keys(), self.summary.values()))
self.logger.debug('Gfxinfo summary data written.')
def stop(self):
self.stop_signal.set()
self.join()
if self.unresponsive_count:
message = 'LatencyCollector was unrepsonsive {} times.'.format(self.unresponsive_count)
if self.unresponsive_count > 10:
self.logger.warning(message)
else:
self.logger.debug(message)
if self.exc:
raise self.exc # pylint: disable=E0702
self.logger.debug('Frame Statistics complete.')
def _process_surfaceflinger_line(self, line):
parts = line.split()
if len(parts) == 3:
frame = SurfaceFlingerFrame(*map(int, parts))
if frame.frame_ready_time <= self.last_ready_time:
return # duplicate frame
if (frame.frame_ready_time - frame.desired_present_time) > self.drop_threshold:
self.logger.debug('Dropping bogus frame {}.'.format(line))
return # bogus data
self.last_ready_time = frame.frame_ready_time
self.frames.append(frame)
elif len(parts) == 1:
self.refresh_period = int(parts[0])
self.drop_threshold = self.refresh_period * 1000
elif 'SurfaceFlinger appears to be unresponsive, dumping anyways' in line:
self.unresponsive_count += 1
else:
self.logger.warning('Unexpected SurfaceFlinger dump output: {}'.format(line))
def _process_gfxinfo_line(self, line):
if 'No process found for' in line:
self.unresponsive_count += 1
return
# Process lines related to the frame data
match = self.re_frame.match(line)
if match:
data = match.group(0)[:-1]
data = map(int, data.split(','))
frame = GfxInfoFrame(*data)
if frame not in self.frames:
if frame.Flags & GFXINFO_EXEMPT:
self.logger.debug('Dropping exempt frame {}.'.format(line))
else:
self.frames.append(frame)
return
# Process lines related to the summary statistics
match = self.re_stats.match(line)
if match:
data = match.group(0)
title, value = data.split(':', 1)
title = title.strip()
value = value.strip()
if title in self.summary:
if 'ms' in value:
value = value.strip('ms')
if '%' in value:
value = value.split()[0]
self.summary[title] = int(value)