Home | History | Annotate | Download | only in utility
      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 "delay_estimator.h"
     12 
     13 #include <assert.h>
     14 #include <stdlib.h>
     15 #include <string.h>
     16 
     17 // Number of right shifts for scaling is linearly depending on number of bits in
     18 // the far-end binary spectrum.
     19 static const int kShiftsAtZero = 13;  // Right shifts at zero binary spectrum.
     20 static const int kShiftsLinearSlope = 3;
     21 
     22 static const int32_t kProbabilityOffset = 1024;  // 2 in Q9.
     23 static const int32_t kProbabilityLowerLimit = 8704;  // 17 in Q9.
     24 static const int32_t kProbabilityMinSpread = 2816;  // 5.5 in Q9.
     25 
     26 // Counts and returns number of bits of a 32-bit word.
     27 static int BitCount(uint32_t u32) {
     28   uint32_t tmp = u32 - ((u32 >> 1) & 033333333333) -
     29       ((u32 >> 2) & 011111111111);
     30   tmp = ((tmp + (tmp >> 3)) & 030707070707);
     31   tmp = (tmp + (tmp >> 6));
     32   tmp = (tmp + (tmp >> 12) + (tmp >> 24)) & 077;
     33 
     34   return ((int) tmp);
     35 }
     36 
     37 // Compares the |binary_vector| with all rows of the |binary_matrix| and counts
     38 // per row the number of times they have the same value.
     39 //
     40 // Inputs:
     41 //      - binary_vector     : binary "vector" stored in a long
     42 //      - binary_matrix     : binary "matrix" stored as a vector of long
     43 //      - matrix_size       : size of binary "matrix"
     44 //
     45 // Output:
     46 //      - bit_counts        : "Vector" stored as a long, containing for each
     47 //                            row the number of times the matrix row and the
     48 //                            input vector have the same value
     49 //
     50 static void BitCountComparison(uint32_t binary_vector,
     51                                const uint32_t* binary_matrix,
     52                                int matrix_size,
     53                                int32_t* bit_counts) {
     54   int n = 0;
     55 
     56   // Compare |binary_vector| with all rows of the |binary_matrix|
     57   for (; n < matrix_size; n++) {
     58     bit_counts[n] = (int32_t) BitCount(binary_vector ^ binary_matrix[n]);
     59   }
     60 }
     61 
     62 int WebRtc_FreeBinaryDelayEstimator(BinaryDelayEstimator* handle) {
     63   assert(handle != NULL);
     64 
     65   if (handle->mean_bit_counts != NULL) {
     66     free(handle->mean_bit_counts);
     67     handle->mean_bit_counts = NULL;
     68   }
     69   if (handle->bit_counts != NULL) {
     70     free(handle->bit_counts);
     71     handle->bit_counts = NULL;
     72   }
     73   if (handle->binary_far_history != NULL) {
     74     free(handle->binary_far_history);
     75     handle->binary_far_history = NULL;
     76   }
     77   if (handle->binary_near_history != NULL) {
     78     free(handle->binary_near_history);
     79     handle->binary_near_history = NULL;
     80   }
     81   if (handle->far_bit_counts != NULL) {
     82     free(handle->far_bit_counts);
     83     handle->far_bit_counts = NULL;
     84   }
     85 
     86   free(handle);
     87 
     88   return 0;
     89 }
     90 
     91 int WebRtc_CreateBinaryDelayEstimator(BinaryDelayEstimator** handle,
     92                                       int max_delay,
     93                                       int lookahead) {
     94   BinaryDelayEstimator* self = NULL;
     95   int history_size = max_delay + lookahead;
     96 
     97   if (handle == NULL) {
     98     return -1;
     99   }
    100   if (max_delay < 0) {
    101     return -1;
    102   }
    103   if (lookahead < 0) {
    104     return -1;
    105   }
    106   if (history_size < 2) {
    107     // Must be this large for buffer shifting.
    108     return -1;
    109   }
    110 
    111   self = malloc(sizeof(BinaryDelayEstimator));
    112   *handle = self;
    113   if (self == NULL) {
    114     return -1;
    115   }
    116 
    117   self->mean_bit_counts = NULL;
    118   self->bit_counts = NULL;
    119   self->binary_far_history = NULL;
    120   self->far_bit_counts = NULL;
    121 
    122   self->history_size = history_size;
    123   self->near_history_size = lookahead + 1;
    124 
    125   // Allocate memory for spectrum buffers.
    126   self->mean_bit_counts = malloc(history_size * sizeof(int32_t));
    127   if (self->mean_bit_counts == NULL) {
    128     WebRtc_FreeBinaryDelayEstimator(self);
    129     self = NULL;
    130     return -1;
    131   }
    132   self->bit_counts = malloc(history_size * sizeof(int32_t));
    133   if (self->bit_counts == NULL) {
    134     WebRtc_FreeBinaryDelayEstimator(self);
    135     self = NULL;
    136     return -1;
    137   }
    138   // Allocate memory for history buffers.
    139   self->binary_far_history = malloc(history_size * sizeof(uint32_t));
    140   if (self->binary_far_history == NULL) {
    141     WebRtc_FreeBinaryDelayEstimator(self);
    142     self = NULL;
    143     return -1;
    144   }
    145   self->binary_near_history = malloc(self->near_history_size *
    146       sizeof(uint32_t));
    147   if (self->binary_near_history == NULL) {
    148     WebRtc_FreeBinaryDelayEstimator(self);
    149     self = NULL;
    150     return -1;
    151   }
    152   self->far_bit_counts = malloc(history_size * sizeof(int));
    153   if (self->far_bit_counts == NULL) {
    154     WebRtc_FreeBinaryDelayEstimator(self);
    155     self = NULL;
    156     return -1;
    157   }
    158 
    159   return 0;
    160 }
    161 
    162 int WebRtc_InitBinaryDelayEstimator(BinaryDelayEstimator* handle) {
    163   int i = 0;
    164   assert(handle != NULL);
    165 
    166   memset(handle->bit_counts, 0, sizeof(int32_t) * handle->history_size);
    167   memset(handle->binary_far_history, 0,
    168          sizeof(uint32_t) * handle->history_size);
    169   memset(handle->binary_near_history, 0,
    170          sizeof(uint32_t) * handle->near_history_size);
    171   memset(handle->far_bit_counts, 0, sizeof(int) * handle->history_size);
    172   for (i = 0; i < handle->history_size; ++i) {
    173     handle->mean_bit_counts[i] = (20 << 9);  // 20 in Q9.
    174   }
    175   handle->minimum_probability = (32 << 9);  // 32 in Q9.
    176   handle->last_delay_probability = (32 << 9);  // 32 in Q9.
    177 
    178   // Default return value if we're unable to estimate. -1 is used for errors.
    179   handle->last_delay = -2;
    180 
    181   return 0;
    182 }
    183 
    184 int WebRtc_ProcessBinarySpectrum(BinaryDelayEstimator* handle,
    185                                  uint32_t binary_far_spectrum,
    186                                  uint32_t binary_near_spectrum) {
    187   int i = 0;
    188   int candidate_delay = -1;
    189 
    190   int32_t value_best_candidate = 16384;  // 32 in Q9, (max |mean_bit_counts|).
    191   int32_t value_worst_candidate = 0;
    192 
    193   assert(handle != NULL);
    194   // Shift binary spectrum history and insert current |binary_far_spectrum|.
    195   memmove(&(handle->binary_far_history[1]), &(handle->binary_far_history[0]),
    196           (handle->history_size - 1) * sizeof(uint32_t));
    197   handle->binary_far_history[0] = binary_far_spectrum;
    198 
    199   // Shift history of far-end binary spectrum bit counts and insert bit count
    200   // of current |binary_far_spectrum|.
    201   memmove(&(handle->far_bit_counts[1]), &(handle->far_bit_counts[0]),
    202           (handle->history_size - 1) * sizeof(int));
    203   handle->far_bit_counts[0] = BitCount(binary_far_spectrum);
    204 
    205   if (handle->near_history_size > 1) {
    206     // If we apply lookahead, shift near-end binary spectrum history. Insert
    207     // current |binary_near_spectrum| and pull out the delayed one.
    208     memmove(&(handle->binary_near_history[1]),
    209             &(handle->binary_near_history[0]),
    210             (handle->near_history_size - 1) * sizeof(uint32_t));
    211     handle->binary_near_history[0] = binary_near_spectrum;
    212     binary_near_spectrum =
    213         handle->binary_near_history[handle->near_history_size - 1];
    214   }
    215 
    216   // Compare with delayed spectra and store the |bit_counts| for each delay.
    217   BitCountComparison(binary_near_spectrum,
    218                      handle->binary_far_history,
    219                      handle->history_size,
    220                      handle->bit_counts);
    221 
    222   // Update |mean_bit_counts|, which is the smoothed version of |bit_counts|.
    223   for (i = 0; i < handle->history_size; i++) {
    224     // |bit_counts| is constrained to [0, 32], meaning we can smooth with a
    225     // factor up to 2^26. We use Q9.
    226     int32_t bit_count = (handle->bit_counts[i] << 9);  // Q9.
    227 
    228     // Update |mean_bit_counts| only when far-end signal has something to
    229     // contribute. If |far_bit_counts| is zero the far-end signal is weak and
    230     // we likely have a poor echo condition, hence don't update.
    231     if (handle->far_bit_counts[i] > 0) {
    232       // Make number of right shifts piecewise linear w.r.t. |far_bit_counts|.
    233       int shifts = kShiftsAtZero;
    234       shifts -= (kShiftsLinearSlope * handle->far_bit_counts[i]) >> 4;
    235       WebRtc_MeanEstimatorFix(bit_count, shifts, &(handle->mean_bit_counts[i]));
    236     }
    237   }
    238 
    239   // Find |candidate_delay|, |value_best_candidate| and |value_worst_candidate|
    240   // of |mean_bit_counts|.
    241   for (i = 0; i < handle->history_size; i++) {
    242     if (handle->mean_bit_counts[i] < value_best_candidate) {
    243       value_best_candidate = handle->mean_bit_counts[i];
    244       candidate_delay = i;
    245     }
    246     if (handle->mean_bit_counts[i] > value_worst_candidate) {
    247       value_worst_candidate = handle->mean_bit_counts[i];
    248     }
    249   }
    250 
    251   // The |value_best_candidate| is a good indicator on the probability of
    252   // |candidate_delay| being an accurate delay (a small |value_best_candidate|
    253   // means a good binary match). In the following sections we make a decision
    254   // whether to update |last_delay| or not.
    255   // 1) If the difference bit counts between the best and the worst delay
    256   //    candidates is too small we consider the situation to be unreliable and
    257   //    don't update |last_delay|.
    258   // 2) If the situation is reliable we update |last_delay| if the value of the
    259   //    best candidate delay has a value less than
    260   //     i) an adaptive threshold |minimum_probability|, or
    261   //    ii) this corresponding value |last_delay_probability|, but updated at
    262   //        this time instant.
    263 
    264   // Update |minimum_probability|.
    265   if ((handle->minimum_probability > kProbabilityLowerLimit) &&
    266       (value_worst_candidate - value_best_candidate > kProbabilityMinSpread)) {
    267     // The "hard" threshold can't be lower than 17 (in Q9).
    268     // The valley in the curve also has to be distinct, i.e., the
    269     // difference between |value_worst_candidate| and |value_best_candidate| has
    270     // to be large enough.
    271     int32_t threshold = value_best_candidate + kProbabilityOffset;
    272     if (threshold < kProbabilityLowerLimit) {
    273       threshold = kProbabilityLowerLimit;
    274     }
    275     if (handle->minimum_probability > threshold) {
    276       handle->minimum_probability = threshold;
    277     }
    278   }
    279   // Update |last_delay_probability|.
    280   // We use a Markov type model, i.e., a slowly increasing level over time.
    281   handle->last_delay_probability++;
    282   if (value_worst_candidate > value_best_candidate + kProbabilityOffset) {
    283     // Reliable delay value for usage.
    284     if (value_best_candidate < handle->minimum_probability) {
    285       handle->last_delay = candidate_delay;
    286     }
    287     if (value_best_candidate < handle->last_delay_probability) {
    288       handle->last_delay = candidate_delay;
    289       // Reset |last_delay_probability|.
    290       handle->last_delay_probability = value_best_candidate;
    291     }
    292   }
    293 
    294   return handle->last_delay;
    295 }
    296 
    297 int WebRtc_binary_last_delay(BinaryDelayEstimator* handle) {
    298   assert(handle != NULL);
    299   return handle->last_delay;
    300 }
    301 
    302 int WebRtc_history_size(BinaryDelayEstimator* handle) {
    303   assert(handle != NULL);
    304   return handle->history_size;
    305 }
    306 
    307 void WebRtc_MeanEstimatorFix(int32_t new_value,
    308                              int factor,
    309                              int32_t* mean_value) {
    310   int32_t diff = new_value - *mean_value;
    311 
    312   // mean_new = mean_value + ((new_value - mean_value) >> factor);
    313   if (diff < 0) {
    314     diff = -((-diff) >> factor);
    315   } else {
    316     diff = (diff >> factor);
    317   }
    318   *mean_value += diff;
    319 }
    320