amrnbdec.c
Go to the documentation of this file.
1 /*
2  * AMR narrowband decoder
3  * Copyright (c) 2006-2007 Robert Swain
4  * Copyright (c) 2009 Colin McQuillan
5  *
6  * This file is part of Libav.
7  *
8  * Libav is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU Lesser General Public
10  * License as published by the Free Software Foundation; either
11  * version 2.1 of the License, or (at your option) any later version.
12  *
13  * Libav is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16  * Lesser General Public License for more details.
17  *
18  * You should have received a copy of the GNU Lesser General Public
19  * License along with Libav; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  */
22 
23 
43 #include <string.h>
44 #include <math.h>
45 
47 #include "avcodec.h"
48 #include "dsputil.h"
49 #include "libavutil/common.h"
50 #include "celp_filters.h"
51 #include "acelp_filters.h"
52 #include "acelp_vectors.h"
53 #include "acelp_pitch_delay.h"
54 #include "lsp.h"
55 #include "amr.h"
56 #include "internal.h"
57 
58 #include "amrnbdata.h"
59 
60 #define AMR_BLOCK_SIZE 160
61 #define AMR_SAMPLE_BOUND 32768.0
62 
63 
72 #define AMR_SAMPLE_SCALE (2.0 / 32768.0)
73 
75 #define PRED_FAC_MODE_12k2 0.65
76 
77 #define LSF_R_FAC (8000.0 / 32768.0)
78 #define MIN_LSF_SPACING (50.0488 / 8000.0)
79 #define PITCH_LAG_MIN_MODE_12k2 18
80 
81 
82 #define MIN_ENERGY -14.0
83 
89 #define SHARP_MAX 0.79449462890625
90 
92 #define AMR_TILT_RESPONSE 22
93 
94 #define AMR_TILT_GAMMA_T 0.8
95 
96 #define AMR_AGC_ALPHA 0.9
97 
98 typedef struct AMRContext {
103 
105  double lsp[4][LP_FILTER_ORDER];
107 
108  float lsf_q[4][LP_FILTER_ORDER];
110 
111  float lpc[4][LP_FILTER_ORDER];
112 
114 
116  float *excitation;
117 
120 
121  float prediction_error[4];
122  float pitch_gain[5];
123  float fixed_gain[5];
124 
125  float beta;
128 
132 
133  float postfilter_mem[10];
134  float tilt_mem;
136  float high_pass_mem[2];
137 
139 
140 } AMRContext;
141 
143 static void weighted_vector_sumd(double *out, const double *in_a,
144  const double *in_b, double weight_coeff_a,
145  double weight_coeff_b, int length)
146 {
147  int i;
148 
149  for (i = 0; i < length; i++)
150  out[i] = weight_coeff_a * in_a[i]
151  + weight_coeff_b * in_b[i];
152 }
153 
155 {
156  AMRContext *p = avctx->priv_data;
157  int i;
158 
159  if (avctx->channels > 1) {
160  av_log_missing_feature(avctx, "multi-channel AMR", 0);
161  return AVERROR_PATCHWELCOME;
162  }
163 
164  avctx->channels = 1;
166  avctx->sample_rate = 8000;
167  avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
168 
169  // p->excitation always points to the same position in p->excitation_buf
171 
172  for (i = 0; i < LP_FILTER_ORDER; i++) {
173  p->prev_lsp_sub4[i] = lsp_sub4_init[i] * 1000 / (float)(1 << 15);
174  p->lsf_avg[i] = p->lsf_q[3][i] = lsp_avg_init[i] / (float)(1 << 15);
175  }
176 
177  for (i = 0; i < 4; i++)
179 
181  avctx->coded_frame = &p->avframe;
182 
183  return 0;
184 }
185 
186 
198 static enum Mode unpack_bitstream(AMRContext *p, const uint8_t *buf,
199  int buf_size)
200 {
201  enum Mode mode;
202 
203  // Decode the first octet.
204  mode = buf[0] >> 3 & 0x0F; // frame type
205  p->bad_frame_indicator = (buf[0] & 0x4) != 0x4; // quality bit
206 
207  if (mode >= N_MODES || buf_size < frame_sizes_nb[mode] + 1) {
208  return NO_DATA;
209  }
210 
211  if (mode < MODE_DTX)
212  ff_amr_bit_reorder((uint16_t *) &p->frame, sizeof(AMRNBFrame), buf + 1,
214 
215  return mode;
216 }
217 
218 
221 
229 static void interpolate_lsf(float lsf_q[4][LP_FILTER_ORDER], float *lsf_new)
230 {
231  int i;
232 
233  for (i = 0; i < 4; i++)
234  ff_weighted_vector_sumf(lsf_q[i], lsf_q[3], lsf_new,
235  0.25 * (3 - i), 0.25 * (i + 1),
236  LP_FILTER_ORDER);
237 }
238 
250 static void lsf2lsp_for_mode12k2(AMRContext *p, double lsp[LP_FILTER_ORDER],
251  const float lsf_no_r[LP_FILTER_ORDER],
252  const int16_t *lsf_quantizer[5],
253  const int quantizer_offset,
254  const int sign, const int update)
255 {
256  int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
257  float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
258  int i;
259 
260  for (i = 0; i < LP_FILTER_ORDER >> 1; i++)
261  memcpy(&lsf_r[i << 1], &lsf_quantizer[i][quantizer_offset],
262  2 * sizeof(*lsf_r));
263 
264  if (sign) {
265  lsf_r[4] *= -1;
266  lsf_r[5] *= -1;
267  }
268 
269  if (update)
270  memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
271 
272  for (i = 0; i < LP_FILTER_ORDER; i++)
273  lsf_q[i] = lsf_r[i] * (LSF_R_FAC / 8000.0) + lsf_no_r[i] * (1.0 / 8000.0);
274 
275  ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
276 
277  if (update)
278  interpolate_lsf(p->lsf_q, lsf_q);
279 
280  ff_acelp_lsf2lspd(lsp, lsf_q, LP_FILTER_ORDER);
281 }
282 
288 static void lsf2lsp_5(AMRContext *p)
289 {
290  const uint16_t *lsf_param = p->frame.lsf;
291  float lsf_no_r[LP_FILTER_ORDER]; // LSFs without the residual vector
292  const int16_t *lsf_quantizer[5];
293  int i;
294 
295  lsf_quantizer[0] = lsf_5_1[lsf_param[0]];
296  lsf_quantizer[1] = lsf_5_2[lsf_param[1]];
297  lsf_quantizer[2] = lsf_5_3[lsf_param[2] >> 1];
298  lsf_quantizer[3] = lsf_5_4[lsf_param[3]];
299  lsf_quantizer[4] = lsf_5_5[lsf_param[4]];
300 
301  for (i = 0; i < LP_FILTER_ORDER; i++)
302  lsf_no_r[i] = p->prev_lsf_r[i] * LSF_R_FAC * PRED_FAC_MODE_12k2 + lsf_5_mean[i];
303 
304  lsf2lsp_for_mode12k2(p, p->lsp[1], lsf_no_r, lsf_quantizer, 0, lsf_param[2] & 1, 0);
305  lsf2lsp_for_mode12k2(p, p->lsp[3], lsf_no_r, lsf_quantizer, 2, lsf_param[2] & 1, 1);
306 
307  // interpolate LSP vectors at subframes 1 and 3
308  weighted_vector_sumd(p->lsp[0], p->prev_lsp_sub4, p->lsp[1], 0.5, 0.5, LP_FILTER_ORDER);
309  weighted_vector_sumd(p->lsp[2], p->lsp[1] , p->lsp[3], 0.5, 0.5, LP_FILTER_ORDER);
310 }
311 
317 static void lsf2lsp_3(AMRContext *p)
318 {
319  const uint16_t *lsf_param = p->frame.lsf;
320  int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
321  float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
322  const int16_t *lsf_quantizer;
323  int i, j;
324 
325  lsf_quantizer = (p->cur_frame_mode == MODE_7k95 ? lsf_3_1_MODE_7k95 : lsf_3_1)[lsf_param[0]];
326  memcpy(lsf_r, lsf_quantizer, 3 * sizeof(*lsf_r));
327 
328  lsf_quantizer = lsf_3_2[lsf_param[1] << (p->cur_frame_mode <= MODE_5k15)];
329  memcpy(lsf_r + 3, lsf_quantizer, 3 * sizeof(*lsf_r));
330 
331  lsf_quantizer = (p->cur_frame_mode <= MODE_5k15 ? lsf_3_3_MODE_5k15 : lsf_3_3)[lsf_param[2]];
332  memcpy(lsf_r + 6, lsf_quantizer, 4 * sizeof(*lsf_r));
333 
334  // calculate mean-removed LSF vector and add mean
335  for (i = 0; i < LP_FILTER_ORDER; i++)
336  lsf_q[i] = (lsf_r[i] + p->prev_lsf_r[i] * pred_fac[i]) * (LSF_R_FAC / 8000.0) + lsf_3_mean[i] * (1.0 / 8000.0);
337 
338  ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
339 
340  // store data for computing the next frame's LSFs
341  interpolate_lsf(p->lsf_q, lsf_q);
342  memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
343 
344  ff_acelp_lsf2lspd(p->lsp[3], lsf_q, LP_FILTER_ORDER);
345 
346  // interpolate LSP vectors at subframes 1, 2 and 3
347  for (i = 1; i <= 3; i++)
348  for(j = 0; j < LP_FILTER_ORDER; j++)
349  p->lsp[i-1][j] = p->prev_lsp_sub4[j] +
350  (p->lsp[3][j] - p->prev_lsp_sub4[j]) * 0.25 * i;
351 }
352 
354 
355 
358 
362 static void decode_pitch_lag_1_6(int *lag_int, int *lag_frac, int pitch_index,
363  const int prev_lag_int, const int subframe)
364 {
365  if (subframe == 0 || subframe == 2) {
366  if (pitch_index < 463) {
367  *lag_int = (pitch_index + 107) * 10923 >> 16;
368  *lag_frac = pitch_index - *lag_int * 6 + 105;
369  } else {
370  *lag_int = pitch_index - 368;
371  *lag_frac = 0;
372  }
373  } else {
374  *lag_int = ((pitch_index + 5) * 10923 >> 16) - 1;
375  *lag_frac = pitch_index - *lag_int * 6 - 3;
376  *lag_int += av_clip(prev_lag_int - 5, PITCH_LAG_MIN_MODE_12k2,
377  PITCH_DELAY_MAX - 9);
378  }
379 }
380 
382  const AMRNBSubframe *amr_subframe,
383  const int subframe)
384 {
385  int pitch_lag_int, pitch_lag_frac;
386  enum Mode mode = p->cur_frame_mode;
387 
388  if (p->cur_frame_mode == MODE_12k2) {
389  decode_pitch_lag_1_6(&pitch_lag_int, &pitch_lag_frac,
390  amr_subframe->p_lag, p->pitch_lag_int,
391  subframe);
392  } else
393  ff_decode_pitch_lag(&pitch_lag_int, &pitch_lag_frac,
394  amr_subframe->p_lag,
395  p->pitch_lag_int, subframe,
396  mode != MODE_4k75 && mode != MODE_5k15,
397  mode <= MODE_6k7 ? 4 : (mode == MODE_7k95 ? 5 : 6));
398 
399  p->pitch_lag_int = pitch_lag_int; // store previous lag in a uint8_t
400 
401  pitch_lag_frac <<= (p->cur_frame_mode != MODE_12k2);
402 
403  pitch_lag_int += pitch_lag_frac > 0;
404 
405  /* Calculate the pitch vector by interpolating the past excitation at the
406  pitch lag using a b60 hamming windowed sinc function. */
407  ff_acelp_interpolatef(p->excitation, p->excitation + 1 - pitch_lag_int,
408  ff_b60_sinc, 6,
409  pitch_lag_frac + 6 - 6*(pitch_lag_frac > 0),
410  10, AMR_SUBFRAME_SIZE);
411 
412  memcpy(p->pitch_vector, p->excitation, AMR_SUBFRAME_SIZE * sizeof(float));
413 }
414 
416 
417 
420 
424 static void decode_10bit_pulse(int code, int pulse_position[8],
425  int i1, int i2, int i3)
426 {
427  // coded using 7+3 bits with the 3 LSBs being, individually, the LSB of 1 of
428  // the 3 pulses and the upper 7 bits being coded in base 5
429  const uint8_t *positions = base_five_table[code >> 3];
430  pulse_position[i1] = (positions[2] << 1) + ( code & 1);
431  pulse_position[i2] = (positions[1] << 1) + ((code >> 1) & 1);
432  pulse_position[i3] = (positions[0] << 1) + ((code >> 2) & 1);
433 }
434 
442 static void decode_8_pulses_31bits(const int16_t *fixed_index,
443  AMRFixed *fixed_sparse)
444 {
445  int pulse_position[8];
446  int i, temp;
447 
448  decode_10bit_pulse(fixed_index[4], pulse_position, 0, 4, 1);
449  decode_10bit_pulse(fixed_index[5], pulse_position, 2, 6, 5);
450 
451  // coded using 5+2 bits with the 2 LSBs being, individually, the LSB of 1 of
452  // the 2 pulses and the upper 5 bits being coded in base 5
453  temp = ((fixed_index[6] >> 2) * 25 + 12) >> 5;
454  pulse_position[3] = temp % 5;
455  pulse_position[7] = temp / 5;
456  if (pulse_position[7] & 1)
457  pulse_position[3] = 4 - pulse_position[3];
458  pulse_position[3] = (pulse_position[3] << 1) + ( fixed_index[6] & 1);
459  pulse_position[7] = (pulse_position[7] << 1) + ((fixed_index[6] >> 1) & 1);
460 
461  fixed_sparse->n = 8;
462  for (i = 0; i < 4; i++) {
463  const int pos1 = (pulse_position[i] << 2) + i;
464  const int pos2 = (pulse_position[i + 4] << 2) + i;
465  const float sign = fixed_index[i] ? -1.0 : 1.0;
466  fixed_sparse->x[i ] = pos1;
467  fixed_sparse->x[i + 4] = pos2;
468  fixed_sparse->y[i ] = sign;
469  fixed_sparse->y[i + 4] = pos2 < pos1 ? -sign : sign;
470  }
471 }
472 
488 static void decode_fixed_sparse(AMRFixed *fixed_sparse, const uint16_t *pulses,
489  const enum Mode mode, const int subframe)
490 {
491  assert(MODE_4k75 <= mode && mode <= MODE_12k2);
492 
493  if (mode == MODE_12k2) {
494  ff_decode_10_pulses_35bits(pulses, fixed_sparse, gray_decode, 5, 3);
495  } else if (mode == MODE_10k2) {
496  decode_8_pulses_31bits(pulses, fixed_sparse);
497  } else {
498  int *pulse_position = fixed_sparse->x;
499  int i, pulse_subset;
500  const int fixed_index = pulses[0];
501 
502  if (mode <= MODE_5k15) {
503  pulse_subset = ((fixed_index >> 3) & 8) + (subframe << 1);
504  pulse_position[0] = ( fixed_index & 7) * 5 + track_position[pulse_subset];
505  pulse_position[1] = ((fixed_index >> 3) & 7) * 5 + track_position[pulse_subset + 1];
506  fixed_sparse->n = 2;
507  } else if (mode == MODE_5k9) {
508  pulse_subset = ((fixed_index & 1) << 1) + 1;
509  pulse_position[0] = ((fixed_index >> 1) & 7) * 5 + pulse_subset;
510  pulse_subset = (fixed_index >> 4) & 3;
511  pulse_position[1] = ((fixed_index >> 6) & 7) * 5 + pulse_subset + (pulse_subset == 3 ? 1 : 0);
512  fixed_sparse->n = pulse_position[0] == pulse_position[1] ? 1 : 2;
513  } else if (mode == MODE_6k7) {
514  pulse_position[0] = (fixed_index & 7) * 5;
515  pulse_subset = (fixed_index >> 2) & 2;
516  pulse_position[1] = ((fixed_index >> 4) & 7) * 5 + pulse_subset + 1;
517  pulse_subset = (fixed_index >> 6) & 2;
518  pulse_position[2] = ((fixed_index >> 8) & 7) * 5 + pulse_subset + 2;
519  fixed_sparse->n = 3;
520  } else { // mode <= MODE_7k95
521  pulse_position[0] = gray_decode[ fixed_index & 7];
522  pulse_position[1] = gray_decode[(fixed_index >> 3) & 7] + 1;
523  pulse_position[2] = gray_decode[(fixed_index >> 6) & 7] + 2;
524  pulse_subset = (fixed_index >> 9) & 1;
525  pulse_position[3] = gray_decode[(fixed_index >> 10) & 7] + pulse_subset + 3;
526  fixed_sparse->n = 4;
527  }
528  for (i = 0; i < fixed_sparse->n; i++)
529  fixed_sparse->y[i] = (pulses[1] >> i) & 1 ? 1.0 : -1.0;
530  }
531 }
532 
541 static void pitch_sharpening(AMRContext *p, int subframe, enum Mode mode,
542  AMRFixed *fixed_sparse)
543 {
544  // The spec suggests the current pitch gain is always used, but in other
545  // modes the pitch and codebook gains are joinly quantized (sec 5.8.2)
546  // so the codebook gain cannot depend on the quantized pitch gain.
547  if (mode == MODE_12k2)
548  p->beta = FFMIN(p->pitch_gain[4], 1.0);
549 
550  fixed_sparse->pitch_lag = p->pitch_lag_int;
551  fixed_sparse->pitch_fac = p->beta;
552 
553  // Save pitch sharpening factor for the next subframe
554  // MODE_4k75 only updates on the 2nd and 4th subframes - this follows from
555  // the fact that the gains for two subframes are jointly quantized.
556  if (mode != MODE_4k75 || subframe & 1)
557  p->beta = av_clipf(p->pitch_gain[4], 0.0, SHARP_MAX);
558 }
560 
561 
564 
577 static float fixed_gain_smooth(AMRContext *p , const float *lsf,
578  const float *lsf_avg, const enum Mode mode)
579 {
580  float diff = 0.0;
581  int i;
582 
583  for (i = 0; i < LP_FILTER_ORDER; i++)
584  diff += fabs(lsf_avg[i] - lsf[i]) / lsf_avg[i];
585 
586  // If diff is large for ten subframes, disable smoothing for a 40-subframe
587  // hangover period.
588  p->diff_count++;
589  if (diff <= 0.65)
590  p->diff_count = 0;
591 
592  if (p->diff_count > 10) {
593  p->hang_count = 0;
594  p->diff_count--; // don't let diff_count overflow
595  }
596 
597  if (p->hang_count < 40) {
598  p->hang_count++;
599  } else if (mode < MODE_7k4 || mode == MODE_10k2) {
600  const float smoothing_factor = av_clipf(4.0 * diff - 1.6, 0.0, 1.0);
601  const float fixed_gain_mean = (p->fixed_gain[0] + p->fixed_gain[1] +
602  p->fixed_gain[2] + p->fixed_gain[3] +
603  p->fixed_gain[4]) * 0.2;
604  return smoothing_factor * p->fixed_gain[4] +
605  (1.0 - smoothing_factor) * fixed_gain_mean;
606  }
607  return p->fixed_gain[4];
608 }
609 
619 static void decode_gains(AMRContext *p, const AMRNBSubframe *amr_subframe,
620  const enum Mode mode, const int subframe,
621  float *fixed_gain_factor)
622 {
623  if (mode == MODE_12k2 || mode == MODE_7k95) {
624  p->pitch_gain[4] = qua_gain_pit [amr_subframe->p_gain ]
625  * (1.0 / 16384.0);
626  *fixed_gain_factor = qua_gain_code[amr_subframe->fixed_gain]
627  * (1.0 / 2048.0);
628  } else {
629  const uint16_t *gains;
630 
631  if (mode >= MODE_6k7) {
632  gains = gains_high[amr_subframe->p_gain];
633  } else if (mode >= MODE_5k15) {
634  gains = gains_low [amr_subframe->p_gain];
635  } else {
636  // gain index is only coded in subframes 0,2 for MODE_4k75
637  gains = gains_MODE_4k75[(p->frame.subframe[subframe & 2].p_gain << 1) + (subframe & 1)];
638  }
639 
640  p->pitch_gain[4] = gains[0] * (1.0 / 16384.0);
641  *fixed_gain_factor = gains[1] * (1.0 / 4096.0);
642  }
643 }
644 
646 
647 
650 
661 static void apply_ir_filter(float *out, const AMRFixed *in,
662  const float *filter)
663 {
664  float filter1[AMR_SUBFRAME_SIZE],
665  filter2[AMR_SUBFRAME_SIZE];
666  int lag = in->pitch_lag;
667  float fac = in->pitch_fac;
668  int i;
669 
670  if (lag < AMR_SUBFRAME_SIZE) {
671  ff_celp_circ_addf(filter1, filter, filter, lag, fac,
673 
674  if (lag < AMR_SUBFRAME_SIZE >> 1)
675  ff_celp_circ_addf(filter2, filter, filter1, lag, fac,
677  }
678 
679  memset(out, 0, sizeof(float) * AMR_SUBFRAME_SIZE);
680  for (i = 0; i < in->n; i++) {
681  int x = in->x[i];
682  float y = in->y[i];
683  const float *filterp;
684 
685  if (x >= AMR_SUBFRAME_SIZE - lag) {
686  filterp = filter;
687  } else if (x >= AMR_SUBFRAME_SIZE - (lag << 1)) {
688  filterp = filter1;
689  } else
690  filterp = filter2;
691 
692  ff_celp_circ_addf(out, out, filterp, x, y, AMR_SUBFRAME_SIZE);
693  }
694 }
695 
708 static const float *anti_sparseness(AMRContext *p, AMRFixed *fixed_sparse,
709  const float *fixed_vector,
710  float fixed_gain, float *out)
711 {
712  int ir_filter_nr;
713 
714  if (p->pitch_gain[4] < 0.6) {
715  ir_filter_nr = 0; // strong filtering
716  } else if (p->pitch_gain[4] < 0.9) {
717  ir_filter_nr = 1; // medium filtering
718  } else
719  ir_filter_nr = 2; // no filtering
720 
721  // detect 'onset'
722  if (fixed_gain > 2.0 * p->prev_sparse_fixed_gain) {
723  p->ir_filter_onset = 2;
724  } else if (p->ir_filter_onset)
725  p->ir_filter_onset--;
726 
727  if (!p->ir_filter_onset) {
728  int i, count = 0;
729 
730  for (i = 0; i < 5; i++)
731  if (p->pitch_gain[i] < 0.6)
732  count++;
733  if (count > 2)
734  ir_filter_nr = 0;
735 
736  if (ir_filter_nr > p->prev_ir_filter_nr + 1)
737  ir_filter_nr--;
738  } else if (ir_filter_nr < 2)
739  ir_filter_nr++;
740 
741  // Disable filtering for very low level of fixed_gain.
742  // Note this step is not specified in the technical description but is in
743  // the reference source in the function Ph_disp.
744  if (fixed_gain < 5.0)
745  ir_filter_nr = 2;
746 
748  && ir_filter_nr < 2) {
749  apply_ir_filter(out, fixed_sparse,
750  (p->cur_frame_mode == MODE_7k95 ?
752  ir_filters_lookup)[ir_filter_nr]);
753  fixed_vector = out;
754  }
755 
756  // update ir filter strength history
757  p->prev_ir_filter_nr = ir_filter_nr;
758  p->prev_sparse_fixed_gain = fixed_gain;
759 
760  return fixed_vector;
761 }
762 
764 
765 
768 
779 static int synthesis(AMRContext *p, float *lpc,
780  float fixed_gain, const float *fixed_vector,
781  float *samples, uint8_t overflow)
782 {
783  int i;
784  float excitation[AMR_SUBFRAME_SIZE];
785 
786  // if an overflow has been detected, the pitch vector is scaled down by a
787  // factor of 4
788  if (overflow)
789  for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
790  p->pitch_vector[i] *= 0.25;
791 
792  ff_weighted_vector_sumf(excitation, p->pitch_vector, fixed_vector,
793  p->pitch_gain[4], fixed_gain, AMR_SUBFRAME_SIZE);
794 
795  // emphasize pitch vector contribution
796  if (p->pitch_gain[4] > 0.5 && !overflow) {
797  float energy = ff_scalarproduct_float_c(excitation, excitation,
798  AMR_SUBFRAME_SIZE);
799  float pitch_factor =
800  p->pitch_gain[4] *
801  (p->cur_frame_mode == MODE_12k2 ?
802  0.25 * FFMIN(p->pitch_gain[4], 1.0) :
803  0.5 * FFMIN(p->pitch_gain[4], SHARP_MAX));
804 
805  for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
806  excitation[i] += pitch_factor * p->pitch_vector[i];
807 
808  ff_scale_vector_to_given_sum_of_squares(excitation, excitation, energy,
809  AMR_SUBFRAME_SIZE);
810  }
811 
812  ff_celp_lp_synthesis_filterf(samples, lpc, excitation, AMR_SUBFRAME_SIZE,
814 
815  // detect overflow
816  for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
817  if (fabsf(samples[i]) > AMR_SAMPLE_BOUND) {
818  return 1;
819  }
820 
821  return 0;
822 }
823 
825 
826 
829 
835 static void update_state(AMRContext *p)
836 {
837  memcpy(p->prev_lsp_sub4, p->lsp[3], LP_FILTER_ORDER * sizeof(p->lsp[3][0]));
838 
839  memmove(&p->excitation_buf[0], &p->excitation_buf[AMR_SUBFRAME_SIZE],
840  (PITCH_DELAY_MAX + LP_FILTER_ORDER + 1) * sizeof(float));
841 
842  memmove(&p->pitch_gain[0], &p->pitch_gain[1], 4 * sizeof(float));
843  memmove(&p->fixed_gain[0], &p->fixed_gain[1], 4 * sizeof(float));
844 
845  memmove(&p->samples_in[0], &p->samples_in[AMR_SUBFRAME_SIZE],
846  LP_FILTER_ORDER * sizeof(float));
847 }
848 
850 
851 
854 
861 static float tilt_factor(float *lpc_n, float *lpc_d)
862 {
863  float rh0, rh1; // autocorrelation at lag 0 and 1
864 
865  // LP_FILTER_ORDER prior zeros are needed for ff_celp_lp_synthesis_filterf
866  float impulse_buffer[LP_FILTER_ORDER + AMR_TILT_RESPONSE] = { 0 };
867  float *hf = impulse_buffer + LP_FILTER_ORDER; // start of impulse response
868 
869  hf[0] = 1.0;
870  memcpy(hf + 1, lpc_n, sizeof(float) * LP_FILTER_ORDER);
872  LP_FILTER_ORDER);
873 
875  rh1 = ff_scalarproduct_float_c(hf, hf + 1, AMR_TILT_RESPONSE - 1);
876 
877  // The spec only specifies this check for 12.2 and 10.2 kbit/s
878  // modes. But in the ref source the tilt is always non-negative.
879  return rh1 >= 0.0 ? rh1 / rh0 * AMR_TILT_GAMMA_T : 0.0;
880 }
881 
890 static void postfilter(AMRContext *p, float *lpc, float *buf_out)
891 {
892  int i;
893  float *samples = p->samples_in + LP_FILTER_ORDER; // Start of input
894 
895  float speech_gain = ff_scalarproduct_float_c(samples, samples,
897 
898  float pole_out[AMR_SUBFRAME_SIZE + LP_FILTER_ORDER]; // Output of pole filter
899  const float *gamma_n, *gamma_d; // Formant filter factor table
900  float lpc_n[LP_FILTER_ORDER], lpc_d[LP_FILTER_ORDER]; // Transfer function coefficients
901 
902  if (p->cur_frame_mode == MODE_12k2 || p->cur_frame_mode == MODE_10k2) {
903  gamma_n = ff_pow_0_7;
904  gamma_d = ff_pow_0_75;
905  } else {
906  gamma_n = ff_pow_0_55;
907  gamma_d = ff_pow_0_7;
908  }
909 
910  for (i = 0; i < LP_FILTER_ORDER; i++) {
911  lpc_n[i] = lpc[i] * gamma_n[i];
912  lpc_d[i] = lpc[i] * gamma_d[i];
913  }
914 
915  memcpy(pole_out, p->postfilter_mem, sizeof(float) * LP_FILTER_ORDER);
916  ff_celp_lp_synthesis_filterf(pole_out + LP_FILTER_ORDER, lpc_d, samples,
917  AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
918  memcpy(p->postfilter_mem, pole_out + AMR_SUBFRAME_SIZE,
919  sizeof(float) * LP_FILTER_ORDER);
920 
921  ff_celp_lp_zero_synthesis_filterf(buf_out, lpc_n,
922  pole_out + LP_FILTER_ORDER,
923  AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
924 
925  ff_tilt_compensation(&p->tilt_mem, tilt_factor(lpc_n, lpc_d), buf_out,
927 
928  ff_adaptive_gain_control(buf_out, buf_out, speech_gain, AMR_SUBFRAME_SIZE,
930 }
931 
933 
934 static int amrnb_decode_frame(AVCodecContext *avctx, void *data,
935  int *got_frame_ptr, AVPacket *avpkt)
936 {
937 
938  AMRContext *p = avctx->priv_data; // pointer to private data
939  const uint8_t *buf = avpkt->data;
940  int buf_size = avpkt->size;
941  float *buf_out; // pointer to the output data buffer
942  int i, subframe, ret;
943  float fixed_gain_factor;
944  AMRFixed fixed_sparse = {0}; // fixed vector up to anti-sparseness processing
945  float spare_vector[AMR_SUBFRAME_SIZE]; // extra stack space to hold result from anti-sparseness processing
946  float synth_fixed_gain; // the fixed gain that synthesis should use
947  const float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
948 
949  /* get output buffer */
951  if ((ret = ff_get_buffer(avctx, &p->avframe)) < 0) {
952  av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
953  return ret;
954  }
955  buf_out = (float *)p->avframe.data[0];
956 
957  p->cur_frame_mode = unpack_bitstream(p, buf, buf_size);
958  if (p->cur_frame_mode == NO_DATA) {
959  av_log(avctx, AV_LOG_ERROR, "Corrupt bitstream\n");
960  return AVERROR_INVALIDDATA;
961  }
962  if (p->cur_frame_mode == MODE_DTX) {
963  av_log_missing_feature(avctx, "dtx mode", 1);
964  return AVERROR_PATCHWELCOME;
965  }
966 
967  if (p->cur_frame_mode == MODE_12k2) {
968  lsf2lsp_5(p);
969  } else
970  lsf2lsp_3(p);
971 
972  for (i = 0; i < 4; i++)
973  ff_acelp_lspd2lpc(p->lsp[i], p->lpc[i], 5);
974 
975  for (subframe = 0; subframe < 4; subframe++) {
976  const AMRNBSubframe *amr_subframe = &p->frame.subframe[subframe];
977 
978  decode_pitch_vector(p, amr_subframe, subframe);
979 
980  decode_fixed_sparse(&fixed_sparse, amr_subframe->pulses,
981  p->cur_frame_mode, subframe);
982 
983  // The fixed gain (section 6.1.3) depends on the fixed vector
984  // (section 6.1.2), but the fixed vector calculation uses
985  // pitch sharpening based on the on the pitch gain (section 6.1.3).
986  // So the correct order is: pitch gain, pitch sharpening, fixed gain.
987  decode_gains(p, amr_subframe, p->cur_frame_mode, subframe,
988  &fixed_gain_factor);
989 
990  pitch_sharpening(p, subframe, p->cur_frame_mode, &fixed_sparse);
991 
992  if (fixed_sparse.pitch_lag == 0) {
993  av_log(avctx, AV_LOG_ERROR, "The file is corrupted, pitch_lag = 0 is not allowed\n");
994  return AVERROR_INVALIDDATA;
995  }
996  ff_set_fixed_vector(p->fixed_vector, &fixed_sparse, 1.0,
998 
999  p->fixed_gain[4] =
1000  ff_amr_set_fixed_gain(fixed_gain_factor,
1002  p->fixed_vector,
1005  p->prediction_error,
1007 
1008  // The excitation feedback is calculated without any processing such
1009  // as fixed gain smoothing. This isn't mentioned in the specification.
1010  for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1011  p->excitation[i] *= p->pitch_gain[4];
1012  ff_set_fixed_vector(p->excitation, &fixed_sparse, p->fixed_gain[4],
1013  AMR_SUBFRAME_SIZE);
1014 
1015  // In the ref decoder, excitation is stored with no fractional bits.
1016  // This step prevents buzz in silent periods. The ref encoder can
1017  // emit long sequences with pitch factor greater than one. This
1018  // creates unwanted feedback if the excitation vector is nonzero.
1019  // (e.g. test sequence T19_795.COD in 3GPP TS 26.074)
1020  for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1021  p->excitation[i] = truncf(p->excitation[i]);
1022 
1023  // Smooth fixed gain.
1024  // The specification is ambiguous, but in the reference source, the
1025  // smoothed value is NOT fed back into later fixed gain smoothing.
1026  synth_fixed_gain = fixed_gain_smooth(p, p->lsf_q[subframe],
1027  p->lsf_avg, p->cur_frame_mode);
1028 
1029  synth_fixed_vector = anti_sparseness(p, &fixed_sparse, p->fixed_vector,
1030  synth_fixed_gain, spare_vector);
1031 
1032  if (synthesis(p, p->lpc[subframe], synth_fixed_gain,
1033  synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 0))
1034  // overflow detected -> rerun synthesis scaling pitch vector down
1035  // by a factor of 4, skipping pitch vector contribution emphasis
1036  // and adaptive gain control
1037  synthesis(p, p->lpc[subframe], synth_fixed_gain,
1038  synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 1);
1039 
1040  postfilter(p, p->lpc[subframe], buf_out + subframe * AMR_SUBFRAME_SIZE);
1041 
1042  // update buffers and history
1043  ff_clear_fixed_vector(p->fixed_vector, &fixed_sparse, AMR_SUBFRAME_SIZE);
1044  update_state(p);
1045  }
1046 
1051 
1052  /* Update averaged lsf vector (used for fixed gain smoothing).
1053  *
1054  * Note that lsf_avg should not incorporate the current frame's LSFs
1055  * for fixed_gain_smooth.
1056  * The specification has an incorrect formula: the reference decoder uses
1057  * qbar(n-1) rather than qbar(n) in section 6.1(4) equation 71. */
1059  0.84, 0.16, LP_FILTER_ORDER);
1060 
1061  *got_frame_ptr = 1;
1062  *(AVFrame *)data = p->avframe;
1063 
1064  /* return the amount of bytes consumed if everything was OK */
1065  return frame_sizes_nb[p->cur_frame_mode] + 1; // +7 for rounding and +8 for TOC
1066 }
1067 
1068 
1070  .name = "amrnb",
1071  .type = AVMEDIA_TYPE_AUDIO,
1072  .id = AV_CODEC_ID_AMR_NB,
1073  .priv_data_size = sizeof(AMRContext),
1076  .capabilities = CODEC_CAP_DR1,
1077  .long_name = NULL_IF_CONFIG_SMALL("AMR-NB (Adaptive Multi-Rate NarrowBand)"),
1078  .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,
1080 };