FFmpeg
 All Data Structures Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
g729postfilter.c
Go to the documentation of this file.
1 /*
2  * G.729, G729 Annex D postfilter
3  * Copyright (c) 2008 Vladimir Voroshilov
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21 #include <inttypes.h>
22 #include <limits.h>
23 
24 #include "avcodec.h"
25 #include "g729.h"
26 #include "acelp_pitch_delay.h"
27 #include "g729postfilter.h"
28 #include "celp_math.h"
29 #include "acelp_filters.h"
30 #include "acelp_vectors.h"
31 #include "celp_filters.h"
32 
33 #define FRAC_BITS 15
34 #include "mathops.h"
35 
36 /**
37  * short interpolation filter (of length 33, according to spec)
38  * for computing signal with non-integer delay
39  */
40 static const int16_t ff_g729_interp_filt_short[(ANALYZED_FRAC_DELAYS+1)*SHORT_INT_FILT_LEN] = {
41  0, 31650, 28469, 23705, 18050, 12266, 7041, 2873,
42  0, -1597, -2147, -1992, -1492, -933, -484, -188,
43 };
44 
45 /**
46  * long interpolation filter (of length 129, according to spec)
47  * for computing signal with non-integer delay
48  */
49 static const int16_t ff_g729_interp_filt_long[(ANALYZED_FRAC_DELAYS+1)*LONG_INT_FILT_LEN] = {
50  0, 31915, 29436, 25569, 20676, 15206, 9639, 4439,
51  0, -3390, -5579, -6549, -6414, -5392, -3773, -1874,
52  0, 1595, 2727, 3303, 3319, 2850, 2030, 1023,
53  0, -887, -1527, -1860, -1876, -1614, -1150, -579,
54  0, 501, 859, 1041, 1044, 892, 631, 315,
55  0, -266, -453, -543, -538, -455, -317, -156,
56  0, 130, 218, 258, 253, 212, 147, 72,
57  0, -59, -101, -122, -123, -106, -77, -40,
58 };
59 
60 /**
61  * formant_pp_factor_num_pow[i] = FORMANT_PP_FACTOR_NUM^(i+1)
62  */
63 static const int16_t formant_pp_factor_num_pow[10]= {
64  /* (0.15) */
65  18022, 9912, 5451, 2998, 1649, 907, 499, 274, 151, 83
66 };
67 
68 /**
69  * formant_pp_factor_den_pow[i] = FORMANT_PP_FACTOR_DEN^(i+1)
70  */
71 static const int16_t formant_pp_factor_den_pow[10] = {
72  /* (0.15) */
73  22938, 16057, 11240, 7868, 5508, 3856, 2699, 1889, 1322, 925
74 };
75 
76 /**
77  * \brief Residual signal calculation (4.2.1 if G.729)
78  * \param out [out] output data filtered through A(z/FORMANT_PP_FACTOR_NUM)
79  * \param filter_coeffs (3.12) A(z/FORMANT_PP_FACTOR_NUM) filter coefficients
80  * \param in input speech data to process
81  * \param subframe_size size of one subframe
82  *
83  * \note in buffer must contain 10 items of previous speech data before top of the buffer
84  * \remark It is safe to pass the same buffer for input and output.
85  */
86 static void residual_filter(int16_t* out, const int16_t* filter_coeffs, const int16_t* in,
87  int subframe_size)
88 {
89  int i, n;
90 
91  for (n = subframe_size - 1; n >= 0; n--) {
92  int sum = 0x800;
93  for (i = 0; i < 10; i++)
94  sum += filter_coeffs[i] * in[n - i - 1];
95 
96  out[n] = in[n] + (sum >> 12);
97  }
98 }
99 
100 /**
101  * \brief long-term postfilter (4.2.1)
102  * \param dsp initialized DSP context
103  * \param pitch_delay_int integer part of the pitch delay in the first subframe
104  * \param residual filtering input data
105  * \param residual_filt [out] speech signal with applied A(z/FORMANT_PP_FACTOR_NUM) filter
106  * \param subframe_size size of subframe
107  *
108  * \return 0 if long-term prediction gain is less than 3dB, 1 - otherwise
109  */
110 static int16_t long_term_filter(DSPContext *dsp, int pitch_delay_int,
111  const int16_t* residual, int16_t *residual_filt,
112  int subframe_size)
113 {
114  int i, k, tmp, tmp2;
115  int sum;
116  int L_temp0;
117  int L_temp1;
118  int64_t L64_temp0;
119  int64_t L64_temp1;
120  int16_t shift;
121  int corr_int_num, corr_int_den;
122 
123  int ener;
124  int16_t sh_ener;
125 
126  int16_t gain_num,gain_den; //selected signal's gain numerator and denominator
127  int16_t sh_gain_num, sh_gain_den;
128  int gain_num_square;
129 
130  int16_t gain_long_num,gain_long_den; //filtered through long interpolation filter signal's gain numerator and denominator
131  int16_t sh_gain_long_num, sh_gain_long_den;
132 
133  int16_t best_delay_int, best_delay_frac;
134 
135  int16_t delayed_signal_offset;
136  int lt_filt_factor_a, lt_filt_factor_b;
137 
138  int16_t * selected_signal;
139  const int16_t * selected_signal_const; //Necessary to avoid compiler warning
140 
141  int16_t sig_scaled[SUBFRAME_SIZE + RES_PREV_DATA_SIZE];
142  int16_t delayed_signal[ANALYZED_FRAC_DELAYS][SUBFRAME_SIZE+1];
143  int corr_den[ANALYZED_FRAC_DELAYS][2];
144 
145  tmp = 0;
146  for(i=0; i<subframe_size + RES_PREV_DATA_SIZE; i++)
147  tmp |= FFABS(residual[i]);
148 
149  if(!tmp)
150  shift = 3;
151  else
152  shift = av_log2(tmp) - 11;
153 
154  if (shift > 0)
155  for (i = 0; i < subframe_size + RES_PREV_DATA_SIZE; i++)
156  sig_scaled[i] = residual[i] >> shift;
157  else
158  for (i = 0; i < subframe_size + RES_PREV_DATA_SIZE; i++)
159  sig_scaled[i] = residual[i] << -shift;
160 
161  /* Start of best delay searching code */
162  gain_num = 0;
163 
164  ener = dsp->scalarproduct_int16(sig_scaled + RES_PREV_DATA_SIZE,
165  sig_scaled + RES_PREV_DATA_SIZE,
166  subframe_size);
167  if (ener) {
168  sh_ener = FFMAX(av_log2(ener) - 14, 0);
169  ener >>= sh_ener;
170  /* Search for best pitch delay.
171 
172  sum{ r(n) * r(k,n) ] }^2
173  R'(k)^2 := -------------------------
174  sum{ r(k,n) * r(k,n) }
175 
176 
177  R(T) := sum{ r(n) * r(n-T) ] }
178 
179 
180  where
181  r(n-T) is integer delayed signal with delay T
182  r(k,n) is non-integer delayed signal with integer delay best_delay
183  and fractional delay k */
184 
185  /* Find integer delay best_delay which maximizes correlation R(T).
186 
187  This is also equals to numerator of R'(0),
188  since the fine search (second step) is done with 1/8
189  precision around best_delay. */
190  corr_int_num = 0;
191  best_delay_int = pitch_delay_int - 1;
192  for (i = pitch_delay_int - 1; i <= pitch_delay_int + 1; i++) {
193  sum = dsp->scalarproduct_int16(sig_scaled + RES_PREV_DATA_SIZE,
194  sig_scaled + RES_PREV_DATA_SIZE - i,
195  subframe_size);
196  if (sum > corr_int_num) {
197  corr_int_num = sum;
198  best_delay_int = i;
199  }
200  }
201  if (corr_int_num) {
202  /* Compute denominator of pseudo-normalized correlation R'(0). */
203  corr_int_den = dsp->scalarproduct_int16(sig_scaled - best_delay_int + RES_PREV_DATA_SIZE,
204  sig_scaled - best_delay_int + RES_PREV_DATA_SIZE,
205  subframe_size);
206 
207  /* Compute signals with non-integer delay k (with 1/8 precision),
208  where k is in [0;6] range.
209  Entire delay is qual to best_delay+(k+1)/8
210  This is archieved by applying an interpolation filter of
211  legth 33 to source signal. */
212  for (k = 0; k < ANALYZED_FRAC_DELAYS; k++) {
213  ff_acelp_interpolate(&delayed_signal[k][0],
214  &sig_scaled[RES_PREV_DATA_SIZE - best_delay_int],
215  ff_g729_interp_filt_short,
216  ANALYZED_FRAC_DELAYS+1,
217  8 - k - 1,
218  SHORT_INT_FILT_LEN,
219  subframe_size + 1);
220  }
221 
222  /* Compute denominator of pseudo-normalized correlation R'(k).
223 
224  corr_den[k][0] is square root of R'(k) denominator, for int(T) == int(T0)
225  corr_den[k][1] is square root of R'(k) denominator, for int(T) == int(T0)+1
226 
227  Also compute maximum value of above denominators over all k. */
228  tmp = corr_int_den;
229  for (k = 0; k < ANALYZED_FRAC_DELAYS; k++) {
230  sum = dsp->scalarproduct_int16(&delayed_signal[k][1],
231  &delayed_signal[k][1],
232  subframe_size - 1);
233  corr_den[k][0] = sum + delayed_signal[k][0 ] * delayed_signal[k][0 ];
234  corr_den[k][1] = sum + delayed_signal[k][subframe_size] * delayed_signal[k][subframe_size];
235 
236  tmp = FFMAX3(tmp, corr_den[k][0], corr_den[k][1]);
237  }
238 
239  sh_gain_den = av_log2(tmp) - 14;
240  if (sh_gain_den >= 0) {
241 
242  sh_gain_num = FFMAX(sh_gain_den, sh_ener);
243  /* Loop through all k and find delay that maximizes
244  R'(k) correlation.
245  Search is done in [int(T0)-1; intT(0)+1] range
246  with 1/8 precision. */
247  delayed_signal_offset = 1;
248  best_delay_frac = 0;
249  gain_den = corr_int_den >> sh_gain_den;
250  gain_num = corr_int_num >> sh_gain_num;
251  gain_num_square = gain_num * gain_num;
252  for (k = 0; k < ANALYZED_FRAC_DELAYS; k++) {
253  for (i = 0; i < 2; i++) {
254  int16_t gain_num_short, gain_den_short;
255  int gain_num_short_square;
256  /* Compute numerator of pseudo-normalized
257  correlation R'(k). */
258  sum = dsp->scalarproduct_int16(&delayed_signal[k][i],
259  sig_scaled + RES_PREV_DATA_SIZE,
260  subframe_size);
261  gain_num_short = FFMAX(sum >> sh_gain_num, 0);
262 
263  /*
264  gain_num_short_square gain_num_square
265  R'(T)^2 = -----------------------, max R'(T)^2= --------------
266  den gain_den
267  */
268  gain_num_short_square = gain_num_short * gain_num_short;
269  gain_den_short = corr_den[k][i] >> sh_gain_den;
270 
271  tmp = MULL(gain_num_short_square, gain_den, FRAC_BITS);
272  tmp2 = MULL(gain_num_square, gain_den_short, FRAC_BITS);
273 
274  // R'(T)^2 > max R'(T)^2
275  if (tmp > tmp2) {
276  gain_num = gain_num_short;
277  gain_den = gain_den_short;
278  gain_num_square = gain_num_short_square;
279  delayed_signal_offset = i;
280  best_delay_frac = k + 1;
281  }
282  }
283  }
284 
285  /*
286  R'(T)^2
287  2 * --------- < 1
288  R(0)
289  */
290  L64_temp0 = (int64_t)gain_num_square << ((sh_gain_num << 1) + 1);
291  L64_temp1 = ((int64_t)gain_den * ener) << (sh_gain_den + sh_ener);
292  if (L64_temp0 < L64_temp1)
293  gain_num = 0;
294  } // if(sh_gain_den >= 0)
295  } // if(corr_int_num)
296  } // if(ener)
297  /* End of best delay searching code */
298 
299  if (!gain_num) {
300  memcpy(residual_filt, residual + RES_PREV_DATA_SIZE, subframe_size * sizeof(int16_t));
301 
302  /* Long-term prediction gain is less than 3dB. Long-term postfilter is disabled. */
303  return 0;
304  }
305  if (best_delay_frac) {
306  /* Recompute delayed signal with an interpolation filter of length 129. */
307  ff_acelp_interpolate(residual_filt,
308  &sig_scaled[RES_PREV_DATA_SIZE - best_delay_int + delayed_signal_offset],
309  ff_g729_interp_filt_long,
310  ANALYZED_FRAC_DELAYS + 1,
311  8 - best_delay_frac,
312  LONG_INT_FILT_LEN,
313  subframe_size + 1);
314  /* Compute R'(k) correlation's numerator. */
315  sum = dsp->scalarproduct_int16(residual_filt,
316  sig_scaled + RES_PREV_DATA_SIZE,
317  subframe_size);
318 
319  if (sum < 0) {
320  gain_long_num = 0;
321  sh_gain_long_num = 0;
322  } else {
323  tmp = FFMAX(av_log2(sum) - 14, 0);
324  sum >>= tmp;
325  gain_long_num = sum;
326  sh_gain_long_num = tmp;
327  }
328 
329  /* Compute R'(k) correlation's denominator. */
330  sum = dsp->scalarproduct_int16(residual_filt, residual_filt, subframe_size);
331 
332  tmp = FFMAX(av_log2(sum) - 14, 0);
333  sum >>= tmp;
334  gain_long_den = sum;
335  sh_gain_long_den = tmp;
336 
337  /* Select between original and delayed signal.
338  Delayed signal will be selected if it increases R'(k)
339  correlation. */
340  L_temp0 = gain_num * gain_num;
341  L_temp0 = MULL(L_temp0, gain_long_den, FRAC_BITS);
342 
343  L_temp1 = gain_long_num * gain_long_num;
344  L_temp1 = MULL(L_temp1, gain_den, FRAC_BITS);
345 
346  tmp = ((sh_gain_long_num - sh_gain_num) << 1) - (sh_gain_long_den - sh_gain_den);
347  if (tmp > 0)
348  L_temp0 >>= tmp;
349  else
350  L_temp1 >>= -tmp;
351 
352  /* Check if longer filter increases the values of R'(k). */
353  if (L_temp1 > L_temp0) {
354  /* Select long filter. */
355  selected_signal = residual_filt;
356  gain_num = gain_long_num;
357  gain_den = gain_long_den;
358  sh_gain_num = sh_gain_long_num;
359  sh_gain_den = sh_gain_long_den;
360  } else
361  /* Select short filter. */
362  selected_signal = &delayed_signal[best_delay_frac-1][delayed_signal_offset];
363 
364  /* Rescale selected signal to original value. */
365  if (shift > 0)
366  for (i = 0; i < subframe_size; i++)
367  selected_signal[i] <<= shift;
368  else
369  for (i = 0; i < subframe_size; i++)
370  selected_signal[i] >>= -shift;
371 
372  /* necessary to avoid compiler warning */
373  selected_signal_const = selected_signal;
374  } // if(best_delay_frac)
375  else
376  selected_signal_const = residual + RES_PREV_DATA_SIZE - (best_delay_int + 1 - delayed_signal_offset);
377 #ifdef G729_BITEXACT
378  tmp = sh_gain_num - sh_gain_den;
379  if (tmp > 0)
380  gain_den >>= tmp;
381  else
382  gain_num >>= -tmp;
383 
384  if (gain_num > gain_den)
385  lt_filt_factor_a = MIN_LT_FILT_FACTOR_A;
386  else {
387  gain_num >>= 2;
388  gain_den >>= 1;
389  lt_filt_factor_a = (gain_den << 15) / (gain_den + gain_num);
390  }
391 #else
392  L64_temp0 = ((int64_t)gain_num) << (sh_gain_num - 1);
393  L64_temp1 = ((int64_t)gain_den) << sh_gain_den;
394  lt_filt_factor_a = FFMAX((L64_temp1 << 15) / (L64_temp1 + L64_temp0), MIN_LT_FILT_FACTOR_A);
395 #endif
396 
397  /* Filter through selected filter. */
398  lt_filt_factor_b = 32767 - lt_filt_factor_a + 1;
399 
400  ff_acelp_weighted_vector_sum(residual_filt, residual + RES_PREV_DATA_SIZE,
401  selected_signal_const,
402  lt_filt_factor_a, lt_filt_factor_b,
403  1<<14, 15, subframe_size);
404 
405  // Long-term prediction gain is larger than 3dB.
406  return 1;
407 }
408 
409 /**
410  * \brief Calculate reflection coefficient for tilt compensation filter (4.2.3).
411  * \param dsp initialized DSP context
412  * \param lp_gn (3.12) coefficients of A(z/FORMANT_PP_FACTOR_NUM) filter
413  * \param lp_gd (3.12) coefficients of A(z/FORMANT_PP_FACTOR_DEN) filter
414  * \param speech speech to update
415  * \param subframe_size size of subframe
416  *
417  * \return (3.12) reflection coefficient
418  *
419  * \remark The routine also calculates the gain term for the short-term
420  * filter (gf) and multiplies the speech data by 1/gf.
421  *
422  * \note All members of lp_gn, except 10-19 must be equal to zero.
423  */
424 static int16_t get_tilt_comp(DSPContext *dsp, int16_t *lp_gn,
425  const int16_t *lp_gd, int16_t* speech,
426  int subframe_size)
427 {
428  int rh1,rh0; // (3.12)
429  int temp;
430  int i;
431  int gain_term;
432 
433  lp_gn[10] = 4096; //1.0 in (3.12)
434 
435  /* Apply 1/A(z/FORMANT_PP_FACTOR_DEN) filter to hf. */
436  ff_celp_lp_synthesis_filter(lp_gn + 11, lp_gd + 1, lp_gn + 11, 22, 10, 0, 0, 0x800);
437  /* Now lp_gn (starting with 10) contains impulse response
438  of A(z/FORMANT_PP_FACTOR_NUM)/A(z/FORMANT_PP_FACTOR_DEN) filter. */
439 
440  rh0 = dsp->scalarproduct_int16(lp_gn + 10, lp_gn + 10, 20);
441  rh1 = dsp->scalarproduct_int16(lp_gn + 10, lp_gn + 11, 20);
442 
443  /* downscale to avoid overflow */
444  temp = av_log2(rh0) - 14;
445  if (temp > 0) {
446  rh0 >>= temp;
447  rh1 >>= temp;
448  }
449 
450  if (FFABS(rh1) > rh0 || !rh0)
451  return 0;
452 
453  gain_term = 0;
454  for (i = 0; i < 20; i++)
455  gain_term += FFABS(lp_gn[i + 10]);
456  gain_term >>= 2; // (3.12) -> (5.10)
457 
458  if (gain_term > 0x400) { // 1.0 in (5.10)
459  temp = 0x2000000 / gain_term; // 1.0/gain_term in (0.15)
460  for (i = 0; i < subframe_size; i++)
461  speech[i] = (speech[i] * temp + 0x4000) >> 15;
462  }
463 
464  return -(rh1 << 15) / rh0;
465 }
466 
467 /**
468  * \brief Apply tilt compensation filter (4.2.3).
469  * \param res_pst [in/out] residual signal (partially filtered)
470  * \param k1 (3.12) reflection coefficient
471  * \param subframe_size size of subframe
472  * \param ht_prev_data previous data for 4.2.3, equation 86
473  *
474  * \return new value for ht_prev_data
475 */
476 static int16_t apply_tilt_comp(int16_t* out, int16_t* res_pst, int refl_coeff,
477  int subframe_size, int16_t ht_prev_data)
478 {
479  int tmp, tmp2;
480  int i;
481  int gt, ga;
482  int fact, sh_fact;
483 
484  if (refl_coeff > 0) {
485  gt = (refl_coeff * G729_TILT_FACTOR_PLUS + 0x4000) >> 15;
486  fact = 0x4000; // 0.5 in (0.15)
487  sh_fact = 15;
488  } else {
489  gt = (refl_coeff * G729_TILT_FACTOR_MINUS + 0x4000) >> 15;
490  fact = 0x800; // 0.5 in (3.12)
491  sh_fact = 12;
492  }
493  ga = (fact << 15) / av_clip_int16(32768 - FFABS(gt));
494  gt >>= 1;
495 
496  /* Apply tilt compensation filter to signal. */
497  tmp = res_pst[subframe_size - 1];
498 
499  for (i = subframe_size - 1; i >= 1; i--) {
500  tmp2 = (res_pst[i] << 15) + ((gt * res_pst[i-1]) << 1);
501  tmp2 = (tmp2 + 0x4000) >> 15;
502 
503  tmp2 = (tmp2 * ga * 2 + fact) >> sh_fact;
504  out[i] = tmp2;
505  }
506  tmp2 = (res_pst[0] << 15) + ((gt * ht_prev_data) << 1);
507  tmp2 = (tmp2 + 0x4000) >> 15;
508  tmp2 = (tmp2 * ga * 2 + fact) >> sh_fact;
509  out[0] = tmp2;
510 
511  return tmp;
512 }
513 
514 void ff_g729_postfilter(DSPContext *dsp, int16_t* ht_prev_data, int* voicing,
515  const int16_t *lp_filter_coeffs, int pitch_delay_int,
516  int16_t* residual, int16_t* res_filter_data,
517  int16_t* pos_filter_data, int16_t *speech, int subframe_size)
518 {
519  int16_t residual_filt_buf[SUBFRAME_SIZE+11];
520  int16_t lp_gn[33]; // (3.12)
521  int16_t lp_gd[11]; // (3.12)
522  int tilt_comp_coeff;
523  int i;
524 
525  /* Zero-filling is necessary for tilt-compensation filter. */
526  memset(lp_gn, 0, 33 * sizeof(int16_t));
527 
528  /* Calculate A(z/FORMANT_PP_FACTOR_NUM) filter coefficients. */
529  for (i = 0; i < 10; i++)
530  lp_gn[i + 11] = (lp_filter_coeffs[i + 1] * formant_pp_factor_num_pow[i] + 0x4000) >> 15;
531 
532  /* Calculate A(z/FORMANT_PP_FACTOR_DEN) filter coefficients. */
533  for (i = 0; i < 10; i++)
534  lp_gd[i + 1] = (lp_filter_coeffs[i + 1] * formant_pp_factor_den_pow[i] + 0x4000) >> 15;
535 
536  /* residual signal calculation (one-half of short-term postfilter) */
537  memcpy(speech - 10, res_filter_data, 10 * sizeof(int16_t));
538  residual_filter(residual + RES_PREV_DATA_SIZE, lp_gn + 11, speech, subframe_size);
539  /* Save data to use it in the next subframe. */
540  memcpy(res_filter_data, speech + subframe_size - 10, 10 * sizeof(int16_t));
541 
542  /* long-term filter. If long-term prediction gain is larger than 3dB (returned value is
543  nonzero) then declare current subframe as periodic. */
544  *voicing = FFMAX(*voicing, long_term_filter(dsp, pitch_delay_int,
545  residual, residual_filt_buf + 10,
546  subframe_size));
547 
548  /* shift residual for using in next subframe */
549  memmove(residual, residual + subframe_size, RES_PREV_DATA_SIZE * sizeof(int16_t));
550 
551  /* short-term filter tilt compensation */
552  tilt_comp_coeff = get_tilt_comp(dsp, lp_gn, lp_gd, residual_filt_buf + 10, subframe_size);
553 
554  /* Apply second half of short-term postfilter: 1/A(z/FORMANT_PP_FACTOR_DEN) */
555  ff_celp_lp_synthesis_filter(pos_filter_data + 10, lp_gd + 1,
556  residual_filt_buf + 10,
557  subframe_size, 10, 0, 0, 0x800);
558  memcpy(pos_filter_data, pos_filter_data + subframe_size, 10 * sizeof(int16_t));
559 
560  *ht_prev_data = apply_tilt_comp(speech, pos_filter_data + 10, tilt_comp_coeff,
561  subframe_size, *ht_prev_data);
562 }
563 
564 /**
565  * \brief Adaptive gain control (4.2.4)
566  * \param gain_before gain of speech before applying postfilters
567  * \param gain_after gain of speech after applying postfilters
568  * \param speech [in/out] signal buffer
569  * \param subframe_size length of subframe
570  * \param gain_prev (3.12) previous value of gain coefficient
571  *
572  * \return (3.12) last value of gain coefficient
573  */
574 int16_t ff_g729_adaptive_gain_control(int gain_before, int gain_after, int16_t *speech,
575  int subframe_size, int16_t gain_prev)
576 {
577  int gain; // (3.12)
578  int n;
579  int exp_before, exp_after;
580 
581  if(!gain_after && gain_before)
582  return 0;
583 
584  if (gain_before) {
585 
586  exp_before = 14 - av_log2(gain_before);
587  gain_before = bidir_sal(gain_before, exp_before);
588 
589  exp_after = 14 - av_log2(gain_after);
590  gain_after = bidir_sal(gain_after, exp_after);
591 
592  if (gain_before < gain_after) {
593  gain = (gain_before << 15) / gain_after;
594  gain = bidir_sal(gain, exp_after - exp_before - 1);
595  } else {
596  gain = ((gain_before - gain_after) << 14) / gain_after + 0x4000;
597  gain = bidir_sal(gain, exp_after - exp_before);
598  }
599  gain = (gain * G729_AGC_FAC1 + 0x4000) >> 15; // gain * (1-0.9875)
600  } else
601  gain = 0;
602 
603  for (n = 0; n < subframe_size; n++) {
604  // gain_prev = gain + 0.9875 * gain_prev
605  gain_prev = (G729_AGC_FACTOR * gain_prev + 0x4000) >> 15;
606  gain_prev = av_clip_int16(gain + gain_prev);
607  speech[n] = av_clip_int16((speech[n] * gain_prev + 0x2000) >> 14);
608  }
609  return gain_prev;
610 }
Generated on Sat May 25 2013 04:01:04 for FFmpeg by   doxygen 1.8.2