1 /*
2  *  Copyright (c) 2011 The WebRTC project authors. All Rights Reserved.
3  *
4  *  Use of this source code is governed by a BSD-style license
5  *  that can be found in the LICENSE file in the root of the source
6  *  tree. An additional intellectual property rights grant can be found
7  *  in the file PATENTS.  All contributing project authors may
8  *  be found in the AUTHORS file in the root of the source tree.
9  */
10 
11 #include "webrtc/modules/video_coding/jitter_estimator.h"
12 
13 #include <assert.h>
14 #include <math.h>
15 #include <stdlib.h>
16 #include <string.h>
17 #include <string>
18 
19 #include "webrtc/modules/video_coding/internal_defines.h"
20 #include "webrtc/modules/video_coding/rtt_filter.h"
21 #include "webrtc/system_wrappers/include/clock.h"
22 #include "webrtc/system_wrappers/include/field_trial.h"
23 
24 namespace webrtc {
25 
26 enum { kStartupDelaySamples = 30 };
27 enum { kFsAccuStartupSamples = 5 };
28 enum { kMaxFramerateEstimate = 200 };
29 
VCMJitterEstimator(const Clock * clock,int32_t vcmId,int32_t receiverId)30 VCMJitterEstimator::VCMJitterEstimator(const Clock* clock,
31                                        int32_t vcmId,
32                                        int32_t receiverId)
33     : _vcmId(vcmId),
34       _receiverId(receiverId),
35       _phi(0.97),
36       _psi(0.9999),
37       _alphaCountMax(400),
38       _thetaLow(0.000001),
39       _nackLimit(3),
40       _numStdDevDelayOutlier(15),
41       _numStdDevFrameSizeOutlier(3),
42       _noiseStdDevs(2.33),       // ~Less than 1% chance
43                                  // (look up in normal distribution table)...
44       _noiseStdDevOffset(30.0),  // ...of getting 30 ms freezes
45       _rttFilter(),
46       fps_counter_(30),  // TODO(sprang): Use an estimator with limit based on
47                          // time, rather than number of samples.
48       low_rate_experiment_(kInit),
49       clock_(clock) {
50   Reset();
51 }
52 
~VCMJitterEstimator()53 VCMJitterEstimator::~VCMJitterEstimator() {}
54 
operator =(const VCMJitterEstimator & rhs)55 VCMJitterEstimator& VCMJitterEstimator::operator=(
56     const VCMJitterEstimator& rhs) {
57   if (this != &rhs) {
58     memcpy(_thetaCov, rhs._thetaCov, sizeof(_thetaCov));
59     memcpy(_Qcov, rhs._Qcov, sizeof(_Qcov));
60 
61     _vcmId = rhs._vcmId;
62     _receiverId = rhs._receiverId;
63     _avgFrameSize = rhs._avgFrameSize;
64     _varFrameSize = rhs._varFrameSize;
65     _maxFrameSize = rhs._maxFrameSize;
66     _fsSum = rhs._fsSum;
67     _fsCount = rhs._fsCount;
68     _lastUpdateT = rhs._lastUpdateT;
69     _prevEstimate = rhs._prevEstimate;
70     _prevFrameSize = rhs._prevFrameSize;
71     _avgNoise = rhs._avgNoise;
72     _alphaCount = rhs._alphaCount;
73     _filterJitterEstimate = rhs._filterJitterEstimate;
74     _startupCount = rhs._startupCount;
75     _latestNackTimestamp = rhs._latestNackTimestamp;
76     _nackCount = rhs._nackCount;
77     _rttFilter = rhs._rttFilter;
78   }
79   return *this;
80 }
81 
82 // Resets the JitterEstimate
Reset()83 void VCMJitterEstimator::Reset() {
84   _theta[0] = 1 / (512e3 / 8);
85   _theta[1] = 0;
86   _varNoise = 4.0;
87 
88   _thetaCov[0][0] = 1e-4;
89   _thetaCov[1][1] = 1e2;
90   _thetaCov[0][1] = _thetaCov[1][0] = 0;
91   _Qcov[0][0] = 2.5e-10;
92   _Qcov[1][1] = 1e-10;
93   _Qcov[0][1] = _Qcov[1][0] = 0;
94   _avgFrameSize = 500;
95   _maxFrameSize = 500;
96   _varFrameSize = 100;
97   _lastUpdateT = -1;
98   _prevEstimate = -1.0;
99   _prevFrameSize = 0;
100   _avgNoise = 0.0;
101   _alphaCount = 1;
102   _filterJitterEstimate = 0.0;
103   _latestNackTimestamp = 0;
104   _nackCount = 0;
105   _fsSum = 0;
106   _fsCount = 0;
107   _startupCount = 0;
108   _rttFilter.Reset();
109   fps_counter_.Reset();
110 }
111 
ResetNackCount()112 void VCMJitterEstimator::ResetNackCount() {
113   _nackCount = 0;
114 }
115 
116 // Updates the estimates with the new measurements
UpdateEstimate(int64_t frameDelayMS,uint32_t frameSizeBytes,bool incompleteFrame)117 void VCMJitterEstimator::UpdateEstimate(int64_t frameDelayMS,
118                                         uint32_t frameSizeBytes,
119                                         bool incompleteFrame /* = false */) {
120   if (frameSizeBytes == 0) {
121     return;
122   }
123   int deltaFS = frameSizeBytes - _prevFrameSize;
124   if (_fsCount < kFsAccuStartupSamples) {
125     _fsSum += frameSizeBytes;
126     _fsCount++;
127   } else if (_fsCount == kFsAccuStartupSamples) {
128     // Give the frame size filter
129     _avgFrameSize = static_cast<double>(_fsSum) / static_cast<double>(_fsCount);
130     _fsCount++;
131   }
132   if (!incompleteFrame || frameSizeBytes > _avgFrameSize) {
133     double avgFrameSize = _phi * _avgFrameSize + (1 - _phi) * frameSizeBytes;
134     if (frameSizeBytes < _avgFrameSize + 2 * sqrt(_varFrameSize)) {
135       // Only update the average frame size if this sample wasn't a
136       // key frame
137       _avgFrameSize = avgFrameSize;
138     }
139     // Update the variance anyway since we want to capture cases where we only
140     // get
141     // key frames.
142     _varFrameSize = VCM_MAX(_phi * _varFrameSize +
143                                 (1 - _phi) * (frameSizeBytes - avgFrameSize) *
144                                     (frameSizeBytes - avgFrameSize),
145                             1.0);
146   }
147 
148   // Update max frameSize estimate
149   _maxFrameSize =
150       VCM_MAX(_psi * _maxFrameSize, static_cast<double>(frameSizeBytes));
151 
152   if (_prevFrameSize == 0) {
153     _prevFrameSize = frameSizeBytes;
154     return;
155   }
156   _prevFrameSize = frameSizeBytes;
157 
158   // Only update the Kalman filter if the sample is not considered
159   // an extreme outlier. Even if it is an extreme outlier from a
160   // delay point of view, if the frame size also is large the
161   // deviation is probably due to an incorrect line slope.
162   double deviation = DeviationFromExpectedDelay(frameDelayMS, deltaFS);
163 
164   if (fabs(deviation) < _numStdDevDelayOutlier * sqrt(_varNoise) ||
165       frameSizeBytes >
166           _avgFrameSize + _numStdDevFrameSizeOutlier * sqrt(_varFrameSize)) {
167     // Update the variance of the deviation from the
168     // line given by the Kalman filter
169     EstimateRandomJitter(deviation, incompleteFrame);
170     // Prevent updating with frames which have been congested by a large
171     // frame, and therefore arrives almost at the same time as that frame.
172     // This can occur when we receive a large frame (key frame) which
173     // has been delayed. The next frame is of normal size (delta frame),
174     // and thus deltaFS will be << 0. This removes all frame samples
175     // which arrives after a key frame.
176     if ((!incompleteFrame || deviation >= 0.0) &&
177         static_cast<double>(deltaFS) > -0.25 * _maxFrameSize) {
178       // Update the Kalman filter with the new data
179       KalmanEstimateChannel(frameDelayMS, deltaFS);
180     }
181   } else {
182     int nStdDev =
183         (deviation >= 0) ? _numStdDevDelayOutlier : -_numStdDevDelayOutlier;
184     EstimateRandomJitter(nStdDev * sqrt(_varNoise), incompleteFrame);
185   }
186   // Post process the total estimated jitter
187   if (_startupCount >= kStartupDelaySamples) {
188     PostProcessEstimate();
189   } else {
190     _startupCount++;
191   }
192 }
193 
194 // Updates the nack/packet ratio
FrameNacked()195 void VCMJitterEstimator::FrameNacked() {
196   // Wait until _nackLimit retransmissions has been received,
197   // then always add ~1 RTT delay.
198   // TODO(holmer): Should we ever remove the additional delay if the
199   // the packet losses seem to have stopped? We could for instance scale
200   // the number of RTTs to add with the amount of retransmissions in a given
201   // time interval, or similar.
202   if (_nackCount < _nackLimit) {
203     _nackCount++;
204   }
205 }
206 
207 // Updates Kalman estimate of the channel
208 // The caller is expected to sanity check the inputs.
KalmanEstimateChannel(int64_t frameDelayMS,int32_t deltaFSBytes)209 void VCMJitterEstimator::KalmanEstimateChannel(int64_t frameDelayMS,
210                                                int32_t deltaFSBytes) {
211   double Mh[2];
212   double hMh_sigma;
213   double kalmanGain[2];
214   double measureRes;
215   double t00, t01;
216 
217   // Kalman filtering
218 
219   // Prediction
220   // M = M + Q
221   _thetaCov[0][0] += _Qcov[0][0];
222   _thetaCov[0][1] += _Qcov[0][1];
223   _thetaCov[1][0] += _Qcov[1][0];
224   _thetaCov[1][1] += _Qcov[1][1];
225 
226   // Kalman gain
227   // K = M*h'/(sigma2n + h*M*h') = M*h'/(1 + h*M*h')
228   // h = [dFS 1]
229   // Mh = M*h'
230   // hMh_sigma = h*M*h' + R
231   Mh[0] = _thetaCov[0][0] * deltaFSBytes + _thetaCov[0][1];
232   Mh[1] = _thetaCov[1][0] * deltaFSBytes + _thetaCov[1][1];
233   // sigma weights measurements with a small deltaFS as noisy and
234   // measurements with large deltaFS as good
235   if (_maxFrameSize < 1.0) {
236     return;
237   }
238   double sigma = (300.0 * exp(-fabs(static_cast<double>(deltaFSBytes)) /
239                               (1e0 * _maxFrameSize)) +
240                   1) *
241                  sqrt(_varNoise);
242   if (sigma < 1.0) {
243     sigma = 1.0;
244   }
245   hMh_sigma = deltaFSBytes * Mh[0] + Mh[1] + sigma;
246   if ((hMh_sigma < 1e-9 && hMh_sigma >= 0) ||
247       (hMh_sigma > -1e-9 && hMh_sigma <= 0)) {
248     assert(false);
249     return;
250   }
251   kalmanGain[0] = Mh[0] / hMh_sigma;
252   kalmanGain[1] = Mh[1] / hMh_sigma;
253 
254   // Correction
255   // theta = theta + K*(dT - h*theta)
256   measureRes = frameDelayMS - (deltaFSBytes * _theta[0] + _theta[1]);
257   _theta[0] += kalmanGain[0] * measureRes;
258   _theta[1] += kalmanGain[1] * measureRes;
259 
260   if (_theta[0] < _thetaLow) {
261     _theta[0] = _thetaLow;
262   }
263 
264   // M = (I - K*h)*M
265   t00 = _thetaCov[0][0];
266   t01 = _thetaCov[0][1];
267   _thetaCov[0][0] = (1 - kalmanGain[0] * deltaFSBytes) * t00 -
268                     kalmanGain[0] * _thetaCov[1][0];
269   _thetaCov[0][1] = (1 - kalmanGain[0] * deltaFSBytes) * t01 -
270                     kalmanGain[0] * _thetaCov[1][1];
271   _thetaCov[1][0] = _thetaCov[1][0] * (1 - kalmanGain[1]) -
272                     kalmanGain[1] * deltaFSBytes * t00;
273   _thetaCov[1][1] = _thetaCov[1][1] * (1 - kalmanGain[1]) -
274                     kalmanGain[1] * deltaFSBytes * t01;
275 
276   // Covariance matrix, must be positive semi-definite
277   assert(_thetaCov[0][0] + _thetaCov[1][1] >= 0 &&
278          _thetaCov[0][0] * _thetaCov[1][1] -
279                  _thetaCov[0][1] * _thetaCov[1][0] >=
280              0 &&
281          _thetaCov[0][0] >= 0);
282 }
283 
284 // Calculate difference in delay between a sample and the
285 // expected delay estimated by the Kalman filter
DeviationFromExpectedDelay(int64_t frameDelayMS,int32_t deltaFSBytes) const286 double VCMJitterEstimator::DeviationFromExpectedDelay(
287     int64_t frameDelayMS,
288     int32_t deltaFSBytes) const {
289   return frameDelayMS - (_theta[0] * deltaFSBytes + _theta[1]);
290 }
291 
292 // Estimates the random jitter by calculating the variance of the
293 // sample distance from the line given by theta.
EstimateRandomJitter(double d_dT,bool incompleteFrame)294 void VCMJitterEstimator::EstimateRandomJitter(double d_dT,
295                                               bool incompleteFrame) {
296   uint64_t now = clock_->TimeInMicroseconds();
297   if (_lastUpdateT != -1) {
298     fps_counter_.AddSample(now - _lastUpdateT);
299   }
300   _lastUpdateT = now;
301 
302   if (_alphaCount == 0) {
303     assert(false);
304     return;
305   }
306   double alpha =
307       static_cast<double>(_alphaCount - 1) / static_cast<double>(_alphaCount);
308   _alphaCount++;
309   if (_alphaCount > _alphaCountMax)
310     _alphaCount = _alphaCountMax;
311 
312   if (LowRateExperimentEnabled()) {
313     // In order to avoid a low frame rate stream to react slower to changes,
314     // scale the alpha weight relative a 30 fps stream.
315     double fps = GetFrameRate();
316     if (fps > 0.0) {
317       double rate_scale = 30.0 / fps;
318       // At startup, there can be a lot of noise in the fps estimate.
319       // Interpolate rate_scale linearly, from 1.0 at sample #1, to 30.0 / fps
320       // at sample #kStartupDelaySamples.
321       if (_alphaCount < kStartupDelaySamples) {
322         rate_scale =
323             (_alphaCount * rate_scale + (kStartupDelaySamples - _alphaCount)) /
324             kStartupDelaySamples;
325       }
326       alpha = pow(alpha, rate_scale);
327     }
328   }
329 
330   double avgNoise = alpha * _avgNoise + (1 - alpha) * d_dT;
331   double varNoise =
332       alpha * _varNoise + (1 - alpha) * (d_dT - _avgNoise) * (d_dT - _avgNoise);
333   if (!incompleteFrame || varNoise > _varNoise) {
334     _avgNoise = avgNoise;
335     _varNoise = varNoise;
336   }
337   if (_varNoise < 1.0) {
338     // The variance should never be zero, since we might get
339     // stuck and consider all samples as outliers.
340     _varNoise = 1.0;
341   }
342 }
343 
NoiseThreshold() const344 double VCMJitterEstimator::NoiseThreshold() const {
345   double noiseThreshold = _noiseStdDevs * sqrt(_varNoise) - _noiseStdDevOffset;
346   if (noiseThreshold < 1.0) {
347     noiseThreshold = 1.0;
348   }
349   return noiseThreshold;
350 }
351 
352 // Calculates the current jitter estimate from the filtered estimates
CalculateEstimate()353 double VCMJitterEstimator::CalculateEstimate() {
354   double ret = _theta[0] * (_maxFrameSize - _avgFrameSize) + NoiseThreshold();
355 
356   // A very low estimate (or negative) is neglected
357   if (ret < 1.0) {
358     if (_prevEstimate <= 0.01) {
359       ret = 1.0;
360     } else {
361       ret = _prevEstimate;
362     }
363   }
364   if (ret > 10000.0) {  // Sanity
365     ret = 10000.0;
366   }
367   _prevEstimate = ret;
368   return ret;
369 }
370 
PostProcessEstimate()371 void VCMJitterEstimator::PostProcessEstimate() {
372   _filterJitterEstimate = CalculateEstimate();
373 }
374 
UpdateRtt(int64_t rttMs)375 void VCMJitterEstimator::UpdateRtt(int64_t rttMs) {
376   _rttFilter.Update(rttMs);
377 }
378 
UpdateMaxFrameSize(uint32_t frameSizeBytes)379 void VCMJitterEstimator::UpdateMaxFrameSize(uint32_t frameSizeBytes) {
380   if (_maxFrameSize < frameSizeBytes) {
381     _maxFrameSize = frameSizeBytes;
382   }
383 }
384 
385 // Returns the current filtered estimate if available,
386 // otherwise tries to calculate an estimate.
GetJitterEstimate(double rttMultiplier)387 int VCMJitterEstimator::GetJitterEstimate(double rttMultiplier) {
388   double jitterMS = CalculateEstimate() + OPERATING_SYSTEM_JITTER;
389   if (_filterJitterEstimate > jitterMS)
390     jitterMS = _filterJitterEstimate;
391   if (_nackCount >= _nackLimit)
392     jitterMS += _rttFilter.RttMs() * rttMultiplier;
393 
394   if (LowRateExperimentEnabled()) {
395     static const double kJitterScaleLowThreshold = 5.0;
396     static const double kJitterScaleHighThreshold = 10.0;
397     double fps = GetFrameRate();
398     // Ignore jitter for very low fps streams.
399     if (fps < kJitterScaleLowThreshold) {
400       if (fps == 0.0) {
401         return jitterMS;
402       }
403       return 0;
404     }
405 
406     // Semi-low frame rate; scale by factor linearly interpolated from 0.0 at
407     // kJitterScaleLowThreshold to 1.0 at kJitterScaleHighThreshold.
408     if (fps < kJitterScaleHighThreshold) {
409       jitterMS =
410           (1.0 / (kJitterScaleHighThreshold - kJitterScaleLowThreshold)) *
411           (fps - kJitterScaleLowThreshold) * jitterMS;
412     }
413   }
414 
415   return static_cast<uint32_t>(jitterMS + 0.5);
416 }
417 
LowRateExperimentEnabled()418 bool VCMJitterEstimator::LowRateExperimentEnabled() {
419   if (low_rate_experiment_ == kInit) {
420     std::string group =
421         webrtc::field_trial::FindFullName("WebRTC-ReducedJitterDelay");
422     if (group == "Disabled") {
423       low_rate_experiment_ = kDisabled;
424     } else {
425       low_rate_experiment_ = kEnabled;
426     }
427   }
428   return low_rate_experiment_ == kEnabled ? true : false;
429 }
430 
GetFrameRate() const431 double VCMJitterEstimator::GetFrameRate() const {
432   if (fps_counter_.count() == 0)
433     return 0;
434 
435   double fps = 1000000.0 / fps_counter_.ComputeMean();
436   // Sanity check.
437   assert(fps >= 0.0);
438   if (fps > kMaxFramerateEstimate) {
439     fps = kMaxFramerateEstimate;
440   }
441   return fps;
442 }
443 }  // namespace webrtc
444