modules/audio_coding/neteq/delay_manager.cc

/*
 *  Copyright (c) 2012 The WebRTC project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include "modules/audio_coding/neteq/delay_manager.h"

#include <assert.h>
#include <stdio.h>
#include <stdlib.h>
#include <algorithm>
#include <numeric>
#include <string>

#include "modules/audio_coding/neteq/delay_peak_detector.h"
#include "modules/include/module_common_types_public.h"
#include "rtc_base/checks.h"
#include "rtc_base/logging.h"
#include "rtc_base/numerics/safe_conversions.h"
#include "rtc_base/numerics/safe_minmax.h"
#include "system_wrappers/include/field_trial.h"

namespace {

constexpr int kLimitProbability = 53687091;         // 1/20 in Q30.
constexpr int kLimitProbabilityStreaming = 536871;  // 1/2000 in Q30.
constexpr int kMaxStreamingPeakPeriodMs = 600000;   // 10 minutes in ms.
constexpr int kCumulativeSumDrift = 2;  // Drift term for cumulative sum
                                        // |iat_cumulative_sum_|.
constexpr int kMinBaseMinimumDelayMs = 0;
constexpr int kMaxBaseMinimumDelayMs = 10000;
// Steady-state forgetting factor for |iat_vector_|, 0.9993 in Q15.
constexpr int kIatFactor_ = 32745;
constexpr int kMaxIat = 64;  // Max inter-arrival time to register.
constexpr int kMaxReorderedPackets =
    10;  // Max number of consecutive reordered packets.

absl::optional<int> GetForcedLimitProbability() {
  constexpr char kForceTargetDelayPercentileFieldTrial[] =
      "WebRTC-Audio-NetEqForceTargetDelayPercentile";
  const bool use_forced_target_delay_percentile =
      webrtc::field_trial::IsEnabled(kForceTargetDelayPercentileFieldTrial);
  if (use_forced_target_delay_percentile) {
    const std::string field_trial_string = webrtc::field_trial::FindFullName(
        kForceTargetDelayPercentileFieldTrial);
    double percentile = -1.0;
    if (sscanf(field_trial_string.c_str(), "Enabled-%lf", &percentile) == 1 &&
        percentile >= 0.0 && percentile <= 100.0) {
      return absl::make_optional<int>(static_cast<int>(
          (1 << 30) * (100.0 - percentile) / 100.0 + 0.5));  // in Q30.
    } else {
      RTC_LOG(LS_WARNING) << "Invalid parameter for "
                          << kForceTargetDelayPercentileFieldTrial
                          << ", ignored.";
    }
  }
  return absl::nullopt;
}

}  // namespace

namespace webrtc {

DelayManager::DelayManager(size_t max_packets_in_buffer,
                           int base_minimum_delay_ms,
                           bool enable_rtx_handling,
                           DelayPeakDetector* peak_detector,
                           const TickTimer* tick_timer)
    : first_packet_received_(false),
      max_packets_in_buffer_(max_packets_in_buffer),
      iat_vector_(kMaxIat + 1, 0),
      iat_factor_(0),
      tick_timer_(tick_timer),
      base_minimum_delay_ms_(base_minimum_delay_ms),
      effective_minimum_delay_ms_(base_minimum_delay_ms),
      base_target_level_(4),                   // In Q0 domain.
      target_level_(base_target_level_ << 8),  // In Q8 domain.
      packet_len_ms_(0),
      streaming_mode_(false),
      last_seq_no_(0),
      last_timestamp_(0),
      minimum_delay_ms_(base_minimum_delay_ms_),
      maximum_delay_ms_(0),
      iat_cumulative_sum_(0),
      max_iat_cumulative_sum_(0),
      peak_detector_(*peak_detector),
      last_pack_cng_or_dtmf_(1),
      frame_length_change_experiment_(
          field_trial::IsEnabled("WebRTC-Audio-NetEqFramelengthExperiment")),
      forced_limit_probability_(GetForcedLimitProbability()),
      enable_rtx_handling_(enable_rtx_handling) {
  assert(peak_detector);  // Should never be NULL.
  RTC_DCHECK_GE(base_minimum_delay_ms_, 0);

  Reset();
}

DelayManager::~DelayManager() {}

const DelayManager::IATVector& DelayManager::iat_vector() const {
  return iat_vector_;
}

// Set the histogram vector to an exponentially decaying distribution
// iat_vector_[i] = 0.5^(i+1), i = 0, 1, 2, ...
// iat_vector_ is in Q30.
void DelayManager::ResetHistogram() {
  // Set temp_prob to (slightly more than) 1 in Q14. This ensures that the sum
  // of iat_vector_ is 1.
  uint16_t temp_prob = 0x4002;  // 16384 + 2 = 100000000000010 binary.
  IATVector::iterator it = iat_vector_.begin();
  for (; it < iat_vector_.end(); it++) {
    temp_prob >>= 1;
    (*it) = temp_prob << 16;
  }
  base_target_level_ = 4;
  target_level_ = base_target_level_ << 8;
}

int DelayManager::Update(uint16_t sequence_number,
                         uint32_t timestamp,
                         int sample_rate_hz) {
  if (sample_rate_hz <= 0) {
    return -1;
  }

  if (!first_packet_received_) {
    // Prepare for next packet arrival.
    packet_iat_stopwatch_ = tick_timer_->GetNewStopwatch();
    last_seq_no_ = sequence_number;
    last_timestamp_ = timestamp;
    first_packet_received_ = true;
    return 0;
  }

  // Try calculating packet length from current and previous timestamps.
  int packet_len_ms;
  if (!IsNewerTimestamp(timestamp, last_timestamp_) ||
      !IsNewerSequenceNumber(sequence_number, last_seq_no_)) {
    // Wrong timestamp or sequence order; use stored value.
    packet_len_ms = packet_len_ms_;
  } else {
    // Calculate timestamps per packet and derive packet length in ms.
    int64_t packet_len_samp =
        static_cast<uint32_t>(timestamp - last_timestamp_) /
        static_cast<uint16_t>(sequence_number - last_seq_no_);
    packet_len_ms =
        rtc::saturated_cast<int>(1000 * packet_len_samp / sample_rate_hz);
  }

  bool reordered = false;
  if (packet_len_ms > 0) {
    // Cannot update statistics unless |packet_len_ms| is valid.
    // Calculate inter-arrival time (IAT) in integer "packet times"
    // (rounding down). This is the value used as index to the histogram
    // vector |iat_vector_|.
    int iat_packets = packet_iat_stopwatch_->ElapsedMs() / packet_len_ms;

    if (streaming_mode_) {
      UpdateCumulativeSums(packet_len_ms, sequence_number);
    }

    // Check for discontinuous packet sequence and re-ordering.
    if (IsNewerSequenceNumber(sequence_number, last_seq_no_ + 1)) {
      // Compensate for gap in the sequence numbers. Reduce IAT with the
      // expected extra time due to lost packets, but ensure that the IAT is
      // not negative.
      iat_packets -= static_cast<uint16_t>(sequence_number - last_seq_no_ - 1);
      iat_packets = std::max(iat_packets, 0);
    } else if (!IsNewerSequenceNumber(sequence_number, last_seq_no_)) {
      iat_packets += static_cast<uint16_t>(last_seq_no_ + 1 - sequence_number);
      reordered = true;
    }

    // Saturate IAT at maximum value.
    const int max_iat = kMaxIat;
    iat_packets = std::min(iat_packets, max_iat);
    UpdateHistogram(iat_packets);
    // Calculate new |target_level_| based on updated statistics.
    target_level_ = CalculateTargetLevel(iat_packets, reordered);
    if (streaming_mode_) {
      target_level_ = std::max(target_level_, max_iat_cumulative_sum_);
    }

    LimitTargetLevel();
  }  // End if (packet_len_ms > 0).

  if (enable_rtx_handling_ && reordered &&
      num_reordered_packets_ < kMaxReorderedPackets) {
    ++num_reordered_packets_;
    return 0;
  }
  num_reordered_packets_ = 0;
  // Prepare for next packet arrival.
  packet_iat_stopwatch_ = tick_timer_->GetNewStopwatch();
  last_seq_no_ = sequence_number;
  last_timestamp_ = timestamp;
  return 0;
}

void DelayManager::UpdateCumulativeSums(int packet_len_ms,
                                        uint16_t sequence_number) {
  // Calculate IAT in Q8, including fractions of a packet (i.e., more
  // accurate than |iat_packets|.
  int iat_packets_q8 =
      (packet_iat_stopwatch_->ElapsedMs() << 8) / packet_len_ms;
  // Calculate cumulative sum IAT with sequence number compensation. The sum
  // is zero if there is no clock-drift.
  iat_cumulative_sum_ +=
      (iat_packets_q8 -
       (static_cast<int>(sequence_number - last_seq_no_) << 8));
  // Subtract drift term.
  iat_cumulative_sum_ -= kCumulativeSumDrift;
  // Ensure not negative.
  iat_cumulative_sum_ = std::max(iat_cumulative_sum_, 0);
  if (iat_cumulative_sum_ > max_iat_cumulative_sum_) {
    // Found a new maximum.
    max_iat_cumulative_sum_ = iat_cumulative_sum_;
    max_iat_stopwatch_ = tick_timer_->GetNewStopwatch();
  }
  if (max_iat_stopwatch_->ElapsedMs() > kMaxStreamingPeakPeriodMs) {
    // Too long since the last maximum was observed; decrease max value.
    max_iat_cumulative_sum_ -= kCumulativeSumDrift;
  }
}

// Each element in the vector is first multiplied by the forgetting factor
// |iat_factor_|. Then the vector element indicated by |iat_packets| is then
// increased (additive) by 1 - |iat_factor_|. This way, the probability of
// |iat_packets| is slightly increased, while the sum of the histogram remains
// constant (=1).
// Due to inaccuracies in the fixed-point arithmetic, the histogram may no
// longer sum up to 1 (in Q30) after the update. To correct this, a correction
// term is added or subtracted from the first element (or elements) of the
// vector.
// The forgetting factor |iat_factor_| is also updated. When the DelayManager
// is reset, the factor is set to 0 to facilitate rapid convergence in the
// beginning. With each update of the histogram, the factor is increased towards
// the steady-state value |kIatFactor_|.
void DelayManager::UpdateHistogram(size_t iat_packets) {
  assert(iat_packets < iat_vector_.size());
  int vector_sum = 0;  // Sum up the vector elements as they are processed.
  // Multiply each element in |iat_vector_| with |iat_factor_|.
  for (IATVector::iterator it = iat_vector_.begin(); it != iat_vector_.end();
       ++it) {
    *it = (static_cast<int64_t>(*it) * iat_factor_) >> 15;
    vector_sum += *it;
  }

  // Increase the probability for the currently observed inter-arrival time
  // by 1 - |iat_factor_|. The factor is in Q15, |iat_vector_| in Q30.
  // Thus, left-shift 15 steps to obtain result in Q30.
  iat_vector_[iat_packets] += (32768 - iat_factor_) << 15;
  vector_sum += (32768 - iat_factor_) << 15;  // Add to vector sum.

  // |iat_vector_| should sum up to 1 (in Q30), but it may not due to
  // fixed-point rounding errors.
  vector_sum -= 1 << 30;  // Should be zero. Compensate if not.
  if (vector_sum != 0) {
    // Modify a few values early in |iat_vector_|.
    int flip_sign = vector_sum > 0 ? -1 : 1;
    IATVector::iterator it = iat_vector_.begin();
    while (it != iat_vector_.end() && abs(vector_sum) > 0) {
      // Add/subtract 1/16 of the element, but not more than |vector_sum|.
      int correction = flip_sign * std::min(abs(vector_sum), (*it) >> 4);
      *it += correction;
      vector_sum += correction;
      ++it;
    }
  }
  assert(vector_sum == 0);  // Verify that the above is correct.

  // Update |iat_factor_| (changes only during the first seconds after a reset).
  // The factor converges to |kIatFactor_|.
  iat_factor_ += (kIatFactor_ - iat_factor_ + 3) >> 2;
}

// Enforces upper and lower limits for |target_level_|. The upper limit is
// chosen to be minimum of i) 75% of |max_packets_in_buffer_|, to leave some
// headroom for natural fluctuations around the target, and ii) equivalent of
// |maximum_delay_ms_| in packets. Note that in practice, if no
// |maximum_delay_ms_| is specified, this does not have any impact, since the
// target level is far below the buffer capacity in all reasonable cases.
// The lower limit is equivalent of |effective_minimum_delay_ms_| in packets.
// We update |least_required_level_| while the above limits are applied.
// TODO(hlundin): Move this check to the buffer logistics class.
void DelayManager::LimitTargetLevel() {
  if (packet_len_ms_ > 0 && effective_minimum_delay_ms_ > 0) {
    int minimum_delay_packet_q8 =
        (effective_minimum_delay_ms_ << 8) / packet_len_ms_;
    target_level_ = std::max(target_level_, minimum_delay_packet_q8);
  }

  if (maximum_delay_ms_ > 0 && packet_len_ms_ > 0) {
    int maximum_delay_packet_q8 = (maximum_delay_ms_ << 8) / packet_len_ms_;
    target_level_ = std::min(target_level_, maximum_delay_packet_q8);
  }

  // Shift to Q8, then 75%.;
  int max_buffer_packets_q8 =
      static_cast<int>((3 * (max_packets_in_buffer_ << 8)) / 4);
  target_level_ = std::min(target_level_, max_buffer_packets_q8);

  // Sanity check, at least 1 packet (in Q8).
  target_level_ = std::max(target_level_, 1 << 8);
}

int DelayManager::CalculateTargetLevel(int iat_packets, bool reordered) {
  int limit_probability = forced_limit_probability_.value_or(kLimitProbability);
  if (streaming_mode_) {
    limit_probability = kLimitProbabilityStreaming;
  }

  // Calculate target buffer level from inter-arrival time histogram.
  // Find the |iat_index| for which the probability of observing an
  // inter-arrival time larger than or equal to |iat_index| is less than or
  // equal to |limit_probability|. The sought probability is estimated using
  // the histogram as the reverse cumulant PDF, i.e., the sum of elements from
  // the end up until |iat_index|. Now, since the sum of all elements is 1
  // (in Q30) by definition, and since the solution is often a low value for
  // |iat_index|, it is more efficient to start with |sum| = 1 and subtract
  // elements from the start of the histogram.
  size_t index = 0;           // Start from the beginning of |iat_vector_|.
  int sum = 1 << 30;          // Assign to 1 in Q30.
  sum -= iat_vector_[index];  // Ensure that target level is >= 1.

  do {
    // Subtract the probabilities one by one until the sum is no longer greater
    // than limit_probability.
    ++index;
    sum -= iat_vector_[index];
  } while ((sum > limit_probability) && (index < iat_vector_.size() - 1));

  // This is the base value for the target buffer level.
  int target_level = static_cast<int>(index);
  base_target_level_ = static_cast<int>(index);

  // Update detector for delay peaks.
  bool delay_peak_found =
      peak_detector_.Update(iat_packets, reordered, target_level);
  if (delay_peak_found) {
    target_level = std::max(target_level, peak_detector_.MaxPeakHeight());
  }

  // Sanity check. |target_level| must be strictly positive.
  target_level = std::max(target_level, 1);
  // Scale to Q8 and assign to member variable.
  target_level_ = target_level << 8;
  return target_level_;
}

int DelayManager::SetPacketAudioLength(int length_ms) {
  if (length_ms <= 0) {
    RTC_LOG_F(LS_ERROR) << "length_ms = " << length_ms;
    return -1;
  }
  if (frame_length_change_experiment_ && packet_len_ms_ != length_ms) {
    iat_vector_ = ScaleHistogram(iat_vector_, packet_len_ms_, length_ms);
  }

  packet_len_ms_ = length_ms;
  peak_detector_.SetPacketAudioLength(packet_len_ms_);
  packet_iat_stopwatch_ = tick_timer_->GetNewStopwatch();
  last_pack_cng_or_dtmf_ = 1;  // TODO(hlundin): Legacy. Remove?
  return 0;
}

void DelayManager::Reset() {
  packet_len_ms_ = 0;  // Packet size unknown.
  streaming_mode_ = false;
  peak_detector_.Reset();
  ResetHistogram();  // Resets target levels too.
  iat_factor_ = 0;   // Adapt the histogram faster for the first few packets.
  packet_iat_stopwatch_ = tick_timer_->GetNewStopwatch();
  max_iat_stopwatch_ = tick_timer_->GetNewStopwatch();
  iat_cumulative_sum_ = 0;
  max_iat_cumulative_sum_ = 0;
  last_pack_cng_or_dtmf_ = 1;
}

double DelayManager::EstimatedClockDriftPpm() const {
  double sum = 0.0;
  // Calculate the expected value based on the probabilities in |iat_vector_|.
  for (size_t i = 0; i < iat_vector_.size(); ++i) {
    sum += static_cast<double>(iat_vector_[i]) * i;
  }
  // The probabilities in |iat_vector_| are in Q30. Divide by 1 << 30 to convert
  // to Q0; subtract the nominal inter-arrival time (1) to make a zero
  // clockdrift represent as 0; mulitply by 1000000 to produce parts-per-million
  // (ppm).
  return (sum / (1 << 30) - 1) * 1e6;
}

bool DelayManager::PeakFound() const {
  return peak_detector_.peak_found();
}

void DelayManager::ResetPacketIatCount() {
  packet_iat_stopwatch_ = tick_timer_->GetNewStopwatch();
}

// Note that |low_limit| and |higher_limit| are not assigned to
// |minimum_delay_ms_| and |maximum_delay_ms_| defined by the client of this
// class. They are computed from |target_level_| and used for decision making.
void DelayManager::BufferLimits(int* lower_limit, int* higher_limit) const {
  if (!lower_limit || !higher_limit) {
    RTC_LOG_F(LS_ERROR) << "NULL pointers supplied as input";
    assert(false);
    return;
  }

  int window_20ms = 0x7FFF;  // Default large value for legacy bit-exactness.
  if (packet_len_ms_ > 0) {
    window_20ms = (20 << 8) / packet_len_ms_;
  }

  // |target_level_| is in Q8 already.
  *lower_limit = (target_level_ * 3) / 4;
  // |higher_limit| is equal to |target_level_|, but should at
  // least be 20 ms higher than |lower_limit_|.
  *higher_limit = std::max(target_level_, *lower_limit + window_20ms);
}

int DelayManager::TargetLevel() const {
  return target_level_;
}

void DelayManager::LastDecodedWasCngOrDtmf(bool it_was) {
  if (it_was) {
    last_pack_cng_or_dtmf_ = 1;
  } else if (last_pack_cng_or_dtmf_ != 0) {
    last_pack_cng_or_dtmf_ = -1;
  }
}

void DelayManager::RegisterEmptyPacket() {
  ++last_seq_no_;
}

DelayManager::IATVector DelayManager::ScaleHistogram(const IATVector& histogram,
                                                     int old_packet_length,
                                                     int new_packet_length) {
  if (old_packet_length == 0) {
    // If we don't know the previous frame length, don't make any changes to the
    // histogram.
    return histogram;
  }
  RTC_DCHECK_GT(new_packet_length, 0);
  RTC_DCHECK_EQ(old_packet_length % 10, 0);
  RTC_DCHECK_EQ(new_packet_length % 10, 0);
  IATVector new_histogram(histogram.size(), 0);
  int64_t acc = 0;
  int time_counter = 0;
  size_t new_histogram_idx = 0;
  for (size_t i = 0; i < histogram.size(); i++) {
    acc += histogram[i];
    time_counter += old_packet_length;
    // The bins should be scaled, to ensure the histogram still sums to one.
    const int64_t scaled_acc = acc * new_packet_length / time_counter;
    int64_t actually_used_acc = 0;
    while (time_counter >= new_packet_length) {
      const int64_t old_histogram_val = new_histogram[new_histogram_idx];
      new_histogram[new_histogram_idx] =
          rtc::saturated_cast<int>(old_histogram_val + scaled_acc);
      actually_used_acc += new_histogram[new_histogram_idx] - old_histogram_val;
      new_histogram_idx =
          std::min(new_histogram_idx + 1, new_histogram.size() - 1);
      time_counter -= new_packet_length;
    }
    // Only subtract the part that was succesfully written to the new histogram.
    acc -= actually_used_acc;
  }
  // If there is anything left in acc (due to rounding errors), add it to the
  // last bin. If we cannot add everything to the last bin we need to add as
  // much as possible to the bins after the last bin (this is only possible
  // when compressing a histogram).
  while (acc > 0 && new_histogram_idx < new_histogram.size()) {
    const int64_t old_histogram_val = new_histogram[new_histogram_idx];
    new_histogram[new_histogram_idx] =
        rtc::saturated_cast<int>(old_histogram_val + acc);
    acc -= new_histogram[new_histogram_idx] - old_histogram_val;
    new_histogram_idx++;
  }
  RTC_DCHECK_EQ(histogram.size(), new_histogram.size());
  if (acc == 0) {
    // If acc is non-zero, we were not able to add everything to the new
    // histogram, so this check will not hold.
    RTC_DCHECK_EQ(accumulate(histogram.begin(), histogram.end(), 0ll),
                  accumulate(new_histogram.begin(), new_histogram.end(), 0ll));
  }
  return new_histogram;
}

bool DelayManager::IsValidMinimumDelay(int delay_ms) const {
  return 0 <= delay_ms && delay_ms <= MinimumDelayUpperBound();
}

bool DelayManager::IsValidBaseMinimumDelay(int delay_ms) const {
  return kMinBaseMinimumDelayMs <= delay_ms &&
         delay_ms <= kMaxBaseMinimumDelayMs;
}

bool DelayManager::SetMinimumDelay(int delay_ms) {
  if (!IsValidMinimumDelay(delay_ms)) {
    return false;
  }

  minimum_delay_ms_ = delay_ms;
  UpdateEffectiveMinimumDelay();
  return true;
}

bool DelayManager::SetMaximumDelay(int delay_ms) {
  // If |delay_ms| is zero then it unsets the maximum delay and target level is
  // unconstrained by maximum delay.
  if (delay_ms != 0 &&
      (delay_ms < minimum_delay_ms_ || delay_ms < packet_len_ms_)) {
    // Maximum delay shouldn't be less than minimum delay or less than a packet.
    return false;
  }

  maximum_delay_ms_ = delay_ms;
  UpdateEffectiveMinimumDelay();
  return true;
}

bool DelayManager::SetBaseMinimumDelay(int delay_ms) {
  if (!IsValidBaseMinimumDelay(delay_ms)) {
    return false;
  }

  base_minimum_delay_ms_ = delay_ms;
  UpdateEffectiveMinimumDelay();
  return true;
}

int DelayManager::GetBaseMinimumDelay() const {
  return base_minimum_delay_ms_;
}

int DelayManager::base_target_level() const {
  return base_target_level_;
}
void DelayManager::set_streaming_mode(bool value) {
  streaming_mode_ = value;
}
int DelayManager::last_pack_cng_or_dtmf() const {
  return last_pack_cng_or_dtmf_;
}

void DelayManager::set_last_pack_cng_or_dtmf(int value) {
  last_pack_cng_or_dtmf_ = value;
}

void DelayManager::UpdateEffectiveMinimumDelay() {
  // Clamp |base_minimum_delay_ms_| into the range which can be effectively
  // used.
  const int base_minimum_delay_ms =
      rtc::SafeClamp(base_minimum_delay_ms_, 0, MinimumDelayUpperBound());
  effective_minimum_delay_ms_ =
      std::max(minimum_delay_ms_, base_minimum_delay_ms);
}

int DelayManager::MinimumDelayUpperBound() const {
  // Choose the lowest possible bound discarding 0 cases which mean the value
  // is not set and unconstrained.
  int q75 = MaxBufferTimeQ75();
  q75 = q75 > 0 ? q75 : kMaxBaseMinimumDelayMs;
  const int maximum_delay_ms =
      maximum_delay_ms_ > 0 ? maximum_delay_ms_ : kMaxBaseMinimumDelayMs;
  return std::min(maximum_delay_ms, q75);
}

int DelayManager::MaxBufferTimeQ75() const {
  const int max_buffer_time = max_packets_in_buffer_ * packet_len_ms_;
  return rtc::dchecked_cast<int>(3 * max_buffer_time / 4);
}
}  // namespace webrtc