Telemetry Test for WebRTC Rendering.

tools/telemetry/telemetry/web_perf/metrics/webrtc_rendering_stats.py

+# Copyright 2015 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.
+
+import logging
+
+import numpy
+
+DISPLAY_HERTZ = 60.0
+VSYNC_DURATION = 1e6 / DISPLAY_HERTZ
+# When to consider a frame frozen (in VSYNC units): meaning 1 initial
+# frame + 5 repeats of that frame.
+FROZEN_THRESHOLD = 6
+# Severity factor.
+SEVERITY = 3
+
+
+class WebMediaPlayerMsRenderingStats(object):
+  """Analyzes events of WebMediaPlayerMs type."""
+
+  def __init__(self, events):
+    """Save relevant events according to their stream."""
+    self.stream_to_events = self._MapEventsToStream(events)
+
+  def _IsEventValid(self, event):
+    """Check that the needed arguments are present in event.
+
+    Args:
+      event: event to check.
+
+    Returns:
+      True is event is valid, false otherwise."""
+    if not event.args:
+      return False
+    mandatory = ['Actual Render Begin', 'Actual Render End',
+        'Ideal Render Instant', 'Serial']
+    for parameter in mandatory:
+      if not parameter in event.args:
+        return False
+    return True
+
+  def _MapEventsToStream(self, events):
+    """Build a dictionary of events indexed by stream.
+
+    The events of interest have a 'Serial' argument which represents the
+    stream ID. The 'Serial' argument identifies the local or remote nature of
+    the stream with a least significant bit  of 0 or 1 as well as the hash
+    value of the video track's URL. So stream::=hash(0|1} . The method will
+    then list the events of the same stream in a frame_distribution on stream
+    id. Practically speaking remote streams have an odd stream id and local
+    streams have a even stream id.
+    Args:
+      events: Telemetry WebMediaPlayerMs events.
+
+    Returns:
+      A dict of stream IDs mapped to events on that stream.
+    """
+    stream_to_events = {}
+    for event in events:
+      if not self._IsEventValid(event):
+        # This is not a render event, skip it.
+        continue
+      stream = event.args['Serial']
+      events_for_stream = stream_to_events.setdefault(stream, [])
+      events_for_stream.append(event)
+
+    return stream_to_events
+
+  def _GetCadence(self, relevant_events):
+    """Calculate the apparent cadence of the rendering.
+
+    In this paragraph I will be using regex notation. What is intended by the
+    word cadence is a sort of extended instantaneous 'Cadence' (thus not
+    necessarily periodic). Just as an example, a normal 'Cadence' could be
+    something like [2 3] which means possibly an observed frame persistance
+    progression of [{2 3}+] for an ideal 20FPS video source. So what we are
+    calculating here is the list of frame persistance, kind of a
+    'Proto-Cadence', but cadence is shorter so we abuse the word.
+
+    Args:
+      relevant_events: list of Telemetry events.
+
+    Returns:
+      a list of frame persistance values.
+    """
+    cadence = []
+    frame_persistence = 0
+    old_ideal_render = 0
+    ideal_render_instant = 'Ideal Render Instant'

[nednguyen, 2015/09/28 20:56:43]

These constant strings should be declared at the top of the file, s.t like:
IDEAL_RENDER_INSTANT = 'Ideal Render Instant'

[cpaulin (no longer in chrome), 2015/09/29 15:50:25]

Done.

+    for event in relevant_events:
+      if not self._IsEventValid(event):
+        # This event is not a render event so skip it.
+        continue
+      if event.args[ideal_render_instant] == old_ideal_render:
+        frame_persistence += 1
+      else:
+        cadence.append(frame_persistence)
+        frame_persistence = 1
+        old_ideal_render = event.args[ideal_render_instant]
+    cadence.append(frame_persistence)
+    cadence.pop(0)
+    return cadence
+
+  def _GetSourceToOutputDistribution(self, cadence):
+    """Create distribution for the cadence frame display values.
+
+    If the overall display distribution is A1:A2:..:An, this will tell us how
+    many times a frame stays displayed during Ak*VSYNC_DURATION, also known as
+    'source to output' distribution. Or in other terms:
+    a distribution B::= let C be the cadence, B[k]=p with k in Unique(C)
+    and p=Card(k in C).
+
+    Args:
+      cadence: list of frame persistance values.
+
+    Returns:
+      a dictionary containing the distribution
+    """
+    frame_distribution = {}
+    for ticks in cadence:
+      ticks_so_far = frame_distribution.setdefault(ticks, 0)
+      frame_distribution[ticks] = ticks_so_far + 1
+    return frame_distribution
+
+  def _GetFpsFromCadence(self, frame_distribution):
+    """Calculate the apparent FPS from frame distribution.
+
+    Knowing the display frequency and the frame distribution, it is possible to
+    calculate the video apparent frame rate as played by WebMediaPlayerMs
+    module.
+
+    Args:
+      frame_distribution: the source to output distribution.
+
+    Returns:
+      the video apparent frame rate.
+    """
+    number_frames = sum(frame_distribution.values())
+    number_vsyncs = sum([ticks * frame_distribution[ticks]
+       for ticks in frame_distribution])
+    mean_ratio = float(number_vsyncs) / number_frames
+    return DISPLAY_HERTZ / mean_ratio
+
+  def _GetFrozenFramesReports(self, frame_distribution):
+    """Find evidence of frozen frames in distribution.
+
+    For simplicity we count as freezing the frames that appear at least five
+    times in a row counted from 'Ideal Render Instant' perspective. So let's
+    say for 1 source frame, we rendered 6 frames, then we consider 5 of these
+    rendered frames as frozen. But we mitigate this by saying anything under
+    5 frozen frames will not be counted as frozen.
+
+    Args:
+      frame_distribution: the source to output distribution.
+
+    Returns:
+      a list of dicts whose keys are ('frozen_frames', 'occurrences').
+    """
+    frozen_frames = []
+    frozen_frame_vsyncs = [ticks for ticks in frame_distribution if ticks >=
+        FROZEN_THRESHOLD]
+    for frozen_frames_vsync in frozen_frame_vsyncs:
+      logging.debug('%s frames not updated after %s vsyncs',
+          frame_distribution[frozen_frames_vsync], frozen_frames_vsync)
+      frozen_frames.append(
+          {'frozen_frames': frozen_frames_vsync - 1,
+           'occurrences': frame_distribution[frozen_frames_vsync]})
+    return frozen_frames
+
+  def _FrozenPenaltyWeight(self, number_frozen_frames):
+    """Returns the weighted penalty for a number of frozen frames.
+
+    As mentioned earlier, we count for frozen anything above 6 vsync display
+    duration for the same 'Initial Render Instant', which is five frozen
+    frames.
+
+    Args:
+      number_frozen_frames: number of frozen frames.
+
+    Returns:
+      the penalty weight (int) for that number of frozen frames.
+    """
+
+    penalty = {
+      0: 0,
+      1: 0,
+      2: 0,
+      3: 0,
+      4: 0,
+      5: 1,
+      6: 5,
+      7: 15,
+      8: 25
+    }
+    weight = penalty.get(number_frozen_frames, 8 * (number_frozen_frames - 4))
+    return weight
+
+  def _IsRemoteStream(self, stream):
+    """Check if stream is remote."""
+    return stream % 2
+
+  def _GetDrifTimeStats(self, relevant_events, cadence):
+    """Get the drift time statistics.
+
+    This method will calculate drift_time stats, that is to say :
+    drift_time::= list(actual render begin - ideal render).
+    rendering_length error::= the rendering length error.
+
+    Args:
+      relevant_events: events to get drift times stats from.
+      cadence: list of frame persistence values.
+
+    Returns:
+      a tuple of (drift_time, rendering_length_error).
+    """
+    drift_time = []
+    old_ideal_render = 0
+    discrepancy = []
+    index = 0
+    ideal_render_instant = 'Ideal Render Instant'
+    for event in relevant_events:
+      current_ideal_render = event.args[ideal_render_instant]
+      if current_ideal_render == old_ideal_render:
+        # Skip to next event because we're looking for a source frame.
+        continue
+      actual_render_begin = event.args['Actual Render Begin']
+      drift_time.append(actual_render_begin - current_ideal_render)
+      discrepancy.append(abs(current_ideal_render - old_ideal_render
+          - VSYNC_DURATION * cadence[index]))
+      old_ideal_render = current_ideal_render
+      index += 1
+    discrepancy.pop(0)
+    last_ideal_render = relevant_events[-1].args[ideal_render_instant]
+    first_ideal_render = relevant_events[0].args[ideal_render_instant]
+    rendering_length_error = 100.0 * (sum([x for x in discrepancy]) /
+        (last_ideal_render - first_ideal_render))
+
+    return drift_time, rendering_length_error
+
+  def _GetSmoothnessStats(self, norm_drift_time):
+    """Get the smoothness stats from the normalized drift time.
+
+    This method will calculate the smoothness score, along with the percentage
+    of frames badly out of sync and the percentage of frames out of sync. To be
+    considered badly out of sync, a frame has to have missed rendering by at
+    least 2*VSYNC_DURATION. To be considered out of sync, a frame has to have
+    missed rendering by at least one VSYNC_DURATION.
+    The smoothness score is a measure of how out of sync the frames are.
+
+    Args:
+      norm_drift_time: normalized drift time.
+
+    Returns:
+      a tuple of (percent_badly_oos, percent_out_of_sync, smoothness_score)
+    """
+    # How many times is a frame later/earlier than T=2*VSYNC_DURATION. Time is
+    # in microseconds.
+    frames_severely_out_of_sync = len(
+        [x for x in norm_drift_time if abs(x) > 2 * VSYNC_DURATION])
+    percent_badly_oos = (
+        100.0 * frames_severely_out_of_sync / len(norm_drift_time))
+
+    # How many times is a frame later/earlier than VSYNC_DURATION.
+    frames_out_of_sync = len(
+        [x for x in norm_drift_time if abs(x) > VSYNC_DURATION])
+    percent_out_of_sync = (
+        100.0 * frames_out_of_sync / len(norm_drift_time))
+
+    frames_oos_only_once = frames_out_of_sync - frames_severely_out_of_sync
+
+    # Calculate smoothness metric. From the formula, we can see that smoothness
+    # score can be negative.
+    smoothness_score = 100.0 - 100.0 * (frames_oos_only_once +
+        SEVERITY * frames_severely_out_of_sync) / len(norm_drift_time)
+
+    # Minimum smoothness_score value allowed is zero.
+    if smoothness_score < 0:
+      smoothness_score = 0
+
+    return (percent_badly_oos, percent_out_of_sync, smoothness_score)
+
+  def _GetFreezingScore(self, frame_distribution):
+    """Get the freezing score."""
+
+    # The freezing score is based on the source to output distribution.
+    number_vsyncs = sum([n * frame_distribution[n]
+        for n in frame_distribution])
+    frozen_frames = self._GetFrozenFramesReports(frame_distribution)
+
+    # Calculate freezing metric.
+    # Freezing metric can be negative if things are really bad. In that case we
+    # change it to zero as minimum valud.
+    freezing_score = 100.0
+    for frozen_report in frozen_frames:
+      weight = self._FrozenPenaltyWeight(frozen_report['frozen_frames'])
+      freezing_score -= (
+          100.0 * frozen_report['occurrences'] / number_vsyncs * weight)
+    if freezing_score < 0:
+      freezing_score = 0
+
+    return freezing_score
+
+  def GetTimeStats(self):
+    """Calculate time stamp stats for all remote stream events."""
+    stats = {}
+    for stream, relevant_events in self.stream_to_events.iteritems():
+      if len(relevant_events) == 1:
+        logging.debug('Found a stream=%s with just one event', stream)
+        continue
+      if not self._IsRemoteStream(stream):
+        logging.info('Skipping processing of local stream: %s', stream)
+        continue
+
+      cadence = self._GetCadence(relevant_events)
+      frame_distribution = self._GetSourceToOutputDistribution(cadence)
+      fps = self._GetFpsFromCadence(frame_distribution)
+
+      drift_time_stats = self._GetDrifTimeStats(relevant_events, cadence)
+      (drift_time, rendering_length_error) = drift_time_stats
+
+      # Drift time normalization.
+      mean_drift_time = numpy.mean(drift_time)
+      norm_drift_time = [abs(x - mean_drift_time) for x in drift_time]
+
+      smoothness_stats = self._GetSmoothnessStats(norm_drift_time)
+      (percent_badly_oos, percent_out_of_sync,
+          smoothness_score) = smoothness_stats
+
+      freezing_score = self._GetFreezingScore(frame_distribution)
+
+      stats = {
+          'drift_time': drift_time,

[nednguyen, 2015/09/28 20:56:43]

I would prefer you create a light weight "TimeStats" class whose drift_time,
mean_drift_time,..etc are member variable instead of using a dict. This way, if
you make a typo, the pylint tool should help you discover it early.

[cpaulin (no longer in chrome), 2015/09/29 15:50:25]

Sure, opted for this class:
class TimeStats(object):
  """Stats container for webrtc rendering metrics."""

  def __init__(self, **kwargs):
    for name in STATS_GATHERED:
      setattr(self, name, kwargs.get(name, None))

+          'mean_drift_time': mean_drift_time,
+          'std_dev_drift_time': numpy.std(drift_time),
+          'percent_badly_out_of_sync': percent_badly_oos,
+          'percent_out_of_sync': percent_out_of_sync,
+          'smoothness_score': smoothness_score,
+          'freezing_score': freezing_score,
+          'rendering_length_error': rendering_length_error,
+          'fps': fps,
+          'frame_distribution': frame_distribution}
+      print "Stats for remote stream {0}: {1}".format(stream, stats)

[nednguyen, 2015/09/28 20:56:43]

Please remove the debug print

[cpaulin (no longer in chrome), 2015/09/29 15:50:25]

Sure

+    return stats