amrwbdec.c
Go to the documentation of this file.
1 /*
2  * AMR wideband decoder
3  * Copyright (c) 2010 Marcelo Galvao Povoa
4  *
5  * This file is part of Libav.
6  *
7  * Libav 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  * Libav 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 Libav; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21 
27 #include "libavutil/lfg.h"
28 
29 #include "avcodec.h"
30 #include "get_bits.h"
31 #include "lsp.h"
32 #include "celp_math.h"
33 #include "celp_filters.h"
34 #include "acelp_filters.h"
35 #include "acelp_vectors.h"
36 #include "acelp_pitch_delay.h"
37 
38 #define AMR_USE_16BIT_TABLES
39 #include "amr.h"
40 
41 #include "amrwbdata.h"
42 
43 typedef struct {
46  enum Mode fr_cur_mode;
47  uint8_t fr_quality;
48  float isf_cur[LP_ORDER];
49  float isf_q_past[LP_ORDER];
50  float isf_past_final[LP_ORDER];
51  double isp[4][LP_ORDER];
52  double isp_sub4_past[LP_ORDER];
53 
54  float lp_coef[4][LP_ORDER];
55 
56  uint8_t base_pitch_lag;
57  uint8_t pitch_lag_int;
58 
59  float excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 2 + AMRWB_SFR_SIZE];
60  float *excitation;
61 
62  float pitch_vector[AMRWB_SFR_SIZE];
63  float fixed_vector[AMRWB_SFR_SIZE];
64 
65  float prediction_error[4];
66  float pitch_gain[6];
67  float fixed_gain[2];
68 
69  float tilt_coef;
70 
73  float prev_tr_gain;
74 
75  float samples_az[LP_ORDER + AMRWB_SFR_SIZE];
76  float samples_up[UPS_MEM_SIZE + AMRWB_SFR_SIZE];
77  float samples_hb[LP_ORDER_16k + AMRWB_SFR_SIZE_16k];
78 
79  float hpf_31_mem[2], hpf_400_mem[2];
80  float demph_mem[1];
81  float bpf_6_7_mem[HB_FIR_SIZE];
82  float lpf_7_mem[HB_FIR_SIZE];
83 
85  uint8_t first_frame;
86 } AMRWBContext;
87 
89 {
90  AMRWBContext *ctx = avctx->priv_data;
91  int i;
92 
94 
95  av_lfg_init(&ctx->prng, 1);
96 
98  ctx->first_frame = 1;
99 
100  for (i = 0; i < LP_ORDER; i++)
101  ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15));
102 
103  for (i = 0; i < 4; i++)
104  ctx->prediction_error[i] = MIN_ENERGY;
105 
107  avctx->coded_frame = &ctx->avframe;
108 
109  return 0;
110 }
111 
121 static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf)
122 {
123  GetBitContext gb;
124  init_get_bits(&gb, buf, 8);
125 
126  /* Decode frame header (1st octet) */
127  skip_bits(&gb, 1); // padding bit
128  ctx->fr_cur_mode = get_bits(&gb, 4);
129  ctx->fr_quality = get_bits1(&gb);
130  skip_bits(&gb, 2); // padding bits
131 
132  return 1;
133 }
134 
142 static void decode_isf_indices_36b(uint16_t *ind, float *isf_q)
143 {
144  int i;
145 
146  for (i = 0; i < 9; i++)
147  isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
148 
149  for (i = 0; i < 7; i++)
150  isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
151 
152  for (i = 0; i < 5; i++)
153  isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15));
154 
155  for (i = 0; i < 4; i++)
156  isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15));
157 
158  for (i = 0; i < 7; i++)
159  isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15));
160 }
161 
169 static void decode_isf_indices_46b(uint16_t *ind, float *isf_q)
170 {
171  int i;
172 
173  for (i = 0; i < 9; i++)
174  isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
175 
176  for (i = 0; i < 7; i++)
177  isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
178 
179  for (i = 0; i < 3; i++)
180  isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15));
181 
182  for (i = 0; i < 3; i++)
183  isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15));
184 
185  for (i = 0; i < 3; i++)
186  isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15));
187 
188  for (i = 0; i < 3; i++)
189  isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15));
190 
191  for (i = 0; i < 4; i++)
192  isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15));
193 }
194 
203 static void isf_add_mean_and_past(float *isf_q, float *isf_past)
204 {
205  int i;
206  float tmp;
207 
208  for (i = 0; i < LP_ORDER; i++) {
209  tmp = isf_q[i];
210  isf_q[i] += isf_mean[i] * (1.0f / (1 << 15));
211  isf_q[i] += PRED_FACTOR * isf_past[i];
212  isf_past[i] = tmp;
213  }
214 }
215 
223 static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past)
224 {
225  int i, k;
226 
227  for (k = 0; k < 3; k++) {
228  float c = isfp_inter[k];
229  for (i = 0; i < LP_ORDER; i++)
230  isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i];
231  }
232 }
233 
245 static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index,
246  uint8_t *base_lag_int, int subframe)
247 {
248  if (subframe == 0 || subframe == 2) {
249  if (pitch_index < 376) {
250  *lag_int = (pitch_index + 137) >> 2;
251  *lag_frac = pitch_index - (*lag_int << 2) + 136;
252  } else if (pitch_index < 440) {
253  *lag_int = (pitch_index + 257 - 376) >> 1;
254  *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) << 1;
255  /* the actual resolution is 1/2 but expressed as 1/4 */
256  } else {
257  *lag_int = pitch_index - 280;
258  *lag_frac = 0;
259  }
260  /* minimum lag for next subframe */
261  *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
263  // XXX: the spec states clearly that *base_lag_int should be
264  // the nearest integer to *lag_int (minus 8), but the ref code
265  // actually always uses its floor, I'm following the latter
266  } else {
267  *lag_int = (pitch_index + 1) >> 2;
268  *lag_frac = pitch_index - (*lag_int << 2);
269  *lag_int += *base_lag_int;
270  }
271 }
272 
278 static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index,
279  uint8_t *base_lag_int, int subframe, enum Mode mode)
280 {
281  if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) {
282  if (pitch_index < 116) {
283  *lag_int = (pitch_index + 69) >> 1;
284  *lag_frac = (pitch_index - (*lag_int << 1) + 68) << 1;
285  } else {
286  *lag_int = pitch_index - 24;
287  *lag_frac = 0;
288  }
289  // XXX: same problem as before
290  *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
292  } else {
293  *lag_int = (pitch_index + 1) >> 1;
294  *lag_frac = (pitch_index - (*lag_int << 1)) << 1;
295  *lag_int += *base_lag_int;
296  }
297 }
298 
308  const AMRWBSubFrame *amr_subframe,
309  const int subframe)
310 {
311  int pitch_lag_int, pitch_lag_frac;
312  int i;
313  float *exc = ctx->excitation;
314  enum Mode mode = ctx->fr_cur_mode;
315 
316  if (mode <= MODE_8k85) {
317  decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
318  &ctx->base_pitch_lag, subframe, mode);
319  } else
320  decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
321  &ctx->base_pitch_lag, subframe);
322 
323  ctx->pitch_lag_int = pitch_lag_int;
324  pitch_lag_int += pitch_lag_frac > 0;
325 
326  /* Calculate the pitch vector by interpolating the past excitation at the
327  pitch lag using a hamming windowed sinc function */
328  ff_acelp_interpolatef(exc, exc + 1 - pitch_lag_int,
329  ac_inter, 4,
330  pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4),
331  LP_ORDER, AMRWB_SFR_SIZE + 1);
332 
333  /* Check which pitch signal path should be used
334  * 6k60 and 8k85 modes have the ltp flag set to 0 */
335  if (amr_subframe->ltp) {
336  memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float));
337  } else {
338  for (i = 0; i < AMRWB_SFR_SIZE; i++)
339  ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] +
340  0.18 * exc[i + 1];
341  memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float));
342  }
343 }
344 
346 #define BIT_STR(x,lsb,len) (((x) >> (lsb)) & ((1 << (len)) - 1))
347 
349 #define BIT_POS(x, p) (((x) >> (p)) & 1)
350 
364 static inline void decode_1p_track(int *out, int code, int m, int off)
365 {
366  int pos = BIT_STR(code, 0, m) + off;
367 
368  out[0] = BIT_POS(code, m) ? -pos : pos;
369 }
370 
371 static inline void decode_2p_track(int *out, int code, int m, int off)
372 {
373  int pos0 = BIT_STR(code, m, m) + off;
374  int pos1 = BIT_STR(code, 0, m) + off;
375 
376  out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0;
377  out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1;
378  out[1] = pos0 > pos1 ? -out[1] : out[1];
379 }
380 
381 static void decode_3p_track(int *out, int code, int m, int off)
382 {
383  int half_2p = BIT_POS(code, 2*m - 1) << (m - 1);
384 
385  decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
386  m - 1, off + half_2p);
387  decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off);
388 }
389 
390 static void decode_4p_track(int *out, int code, int m, int off)
391 {
392  int half_4p, subhalf_2p;
393  int b_offset = 1 << (m - 1);
394 
395  switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */
396  case 0: /* 0 pulses in A, 4 pulses in B or vice versa */
397  half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses
398  subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2);
399 
400  decode_2p_track(out, BIT_STR(code, 0, 2*m - 3),
401  m - 2, off + half_4p + subhalf_2p);
402  decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1),
403  m - 1, off + half_4p);
404  break;
405  case 1: /* 1 pulse in A, 3 pulses in B */
406  decode_1p_track(out, BIT_STR(code, 3*m - 2, m),
407  m - 1, off);
408  decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2),
409  m - 1, off + b_offset);
410  break;
411  case 2: /* 2 pulses in each half */
412  decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1),
413  m - 1, off);
414  decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1),
415  m - 1, off + b_offset);
416  break;
417  case 3: /* 3 pulses in A, 1 pulse in B */
418  decode_3p_track(out, BIT_STR(code, m, 3*m - 2),
419  m - 1, off);
420  decode_1p_track(out + 3, BIT_STR(code, 0, m),
421  m - 1, off + b_offset);
422  break;
423  }
424 }
425 
426 static void decode_5p_track(int *out, int code, int m, int off)
427 {
428  int half_3p = BIT_POS(code, 5*m - 1) << (m - 1);
429 
430  decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2),
431  m - 1, off + half_3p);
432 
433  decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off);
434 }
435 
436 static void decode_6p_track(int *out, int code, int m, int off)
437 {
438  int b_offset = 1 << (m - 1);
439  /* which half has more pulses in cases 0 to 2 */
440  int half_more = BIT_POS(code, 6*m - 5) << (m - 1);
441  int half_other = b_offset - half_more;
442 
443  switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */
444  case 0: /* 0 pulses in A, 6 pulses in B or vice versa */
445  decode_1p_track(out, BIT_STR(code, 0, m),
446  m - 1, off + half_more);
447  decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
448  m - 1, off + half_more);
449  break;
450  case 1: /* 1 pulse in A, 5 pulses in B or vice versa */
451  decode_1p_track(out, BIT_STR(code, 0, m),
452  m - 1, off + half_other);
453  decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
454  m - 1, off + half_more);
455  break;
456  case 2: /* 2 pulses in A, 4 pulses in B or vice versa */
457  decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
458  m - 1, off + half_other);
459  decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4),
460  m - 1, off + half_more);
461  break;
462  case 3: /* 3 pulses in A, 3 pulses in B */
463  decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2),
464  m - 1, off);
465  decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2),
466  m - 1, off + b_offset);
467  break;
468  }
469 }
470 
480 static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi,
481  const uint16_t *pulse_lo, const enum Mode mode)
482 {
483  /* sig_pos stores for each track the decoded pulse position indexes
484  * (1-based) multiplied by its corresponding amplitude (+1 or -1) */
485  int sig_pos[4][6];
486  int spacing = (mode == MODE_6k60) ? 2 : 4;
487  int i, j;
488 
489  switch (mode) {
490  case MODE_6k60:
491  for (i = 0; i < 2; i++)
492  decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1);
493  break;
494  case MODE_8k85:
495  for (i = 0; i < 4; i++)
496  decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1);
497  break;
498  case MODE_12k65:
499  for (i = 0; i < 4; i++)
500  decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
501  break;
502  case MODE_14k25:
503  for (i = 0; i < 2; i++)
504  decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
505  for (i = 2; i < 4; i++)
506  decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
507  break;
508  case MODE_15k85:
509  for (i = 0; i < 4; i++)
510  decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
511  break;
512  case MODE_18k25:
513  for (i = 0; i < 4; i++)
514  decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
515  ((int) pulse_hi[i] << 14), 4, 1);
516  break;
517  case MODE_19k85:
518  for (i = 0; i < 2; i++)
519  decode_5p_track(sig_pos[i], (int) pulse_lo[i] +
520  ((int) pulse_hi[i] << 10), 4, 1);
521  for (i = 2; i < 4; i++)
522  decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
523  ((int) pulse_hi[i] << 14), 4, 1);
524  break;
525  case MODE_23k05:
526  case MODE_23k85:
527  for (i = 0; i < 4; i++)
528  decode_6p_track(sig_pos[i], (int) pulse_lo[i] +
529  ((int) pulse_hi[i] << 11), 4, 1);
530  break;
531  }
532 
533  memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE);
534 
535  for (i = 0; i < 4; i++)
536  for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) {
537  int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i;
538 
539  fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0;
540  }
541 }
542 
551 static void decode_gains(const uint8_t vq_gain, const enum Mode mode,
552  float *fixed_gain_factor, float *pitch_gain)
553 {
554  const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] :
555  qua_gain_7b[vq_gain]);
556 
557  *pitch_gain = gains[0] * (1.0f / (1 << 14));
558  *fixed_gain_factor = gains[1] * (1.0f / (1 << 11));
559 }
560 
567 // XXX: Spec states this procedure should be applied when the pitch
568 // lag is less than 64, but this checking seems absent in reference and AMR-NB
569 static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector)
570 {
571  int i;
572 
573  /* Tilt part */
574  for (i = AMRWB_SFR_SIZE - 1; i != 0; i--)
575  fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef;
576 
577  /* Periodicity enhancement part */
578  for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++)
579  fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85;
580 }
581 
588 // XXX: There is something wrong with the precision here! The magnitudes
589 // of the energies are not correct. Please check the reference code carefully
590 static float voice_factor(float *p_vector, float p_gain,
591  float *f_vector, float f_gain)
592 {
593  double p_ener = (double) ff_dot_productf(p_vector, p_vector,
594  AMRWB_SFR_SIZE) * p_gain * p_gain;
595  double f_ener = (double) ff_dot_productf(f_vector, f_vector,
596  AMRWB_SFR_SIZE) * f_gain * f_gain;
597 
598  return (p_ener - f_ener) / (p_ener + f_ener);
599 }
600 
611 static float *anti_sparseness(AMRWBContext *ctx,
612  float *fixed_vector, float *buf)
613 {
614  int ir_filter_nr;
615 
616  if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes
617  return fixed_vector;
618 
619  if (ctx->pitch_gain[0] < 0.6) {
620  ir_filter_nr = 0; // strong filtering
621  } else if (ctx->pitch_gain[0] < 0.9) {
622  ir_filter_nr = 1; // medium filtering
623  } else
624  ir_filter_nr = 2; // no filtering
625 
626  /* detect 'onset' */
627  if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) {
628  if (ir_filter_nr < 2)
629  ir_filter_nr++;
630  } else {
631  int i, count = 0;
632 
633  for (i = 0; i < 6; i++)
634  if (ctx->pitch_gain[i] < 0.6)
635  count++;
636 
637  if (count > 2)
638  ir_filter_nr = 0;
639 
640  if (ir_filter_nr > ctx->prev_ir_filter_nr + 1)
641  ir_filter_nr--;
642  }
643 
644  /* update ir filter strength history */
645  ctx->prev_ir_filter_nr = ir_filter_nr;
646 
647  ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85);
648 
649  if (ir_filter_nr < 2) {
650  int i;
651  const float *coef = ir_filters_lookup[ir_filter_nr];
652 
653  /* Circular convolution code in the reference
654  * decoder was modified to avoid using one
655  * extra array. The filtered vector is given by:
656  *
657  * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) }
658  */
659 
660  memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE);
661  for (i = 0; i < AMRWB_SFR_SIZE; i++)
662  if (fixed_vector[i])
663  ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i],
664  AMRWB_SFR_SIZE);
665  fixed_vector = buf;
666  }
667 
668  return fixed_vector;
669 }
670 
675 static float stability_factor(const float *isf, const float *isf_past)
676 {
677  int i;
678  float acc = 0.0;
679 
680  for (i = 0; i < LP_ORDER - 1; i++)
681  acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]);
682 
683  // XXX: This part is not so clear from the reference code
684  // the result is more accurate changing the "/ 256" to "* 512"
685  return FFMAX(0.0, 1.25 - acc * 0.8 * 512);
686 }
687 
699 static float noise_enhancer(float fixed_gain, float *prev_tr_gain,
700  float voice_fac, float stab_fac)
701 {
702  float sm_fac = 0.5 * (1 - voice_fac) * stab_fac;
703  float g0;
704 
705  // XXX: the following fixed-point constants used to in(de)crement
706  // gain by 1.5dB were taken from the reference code, maybe it could
707  // be simpler
708  if (fixed_gain < *prev_tr_gain) {
709  g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain *
710  (6226 * (1.0f / (1 << 15)))); // +1.5 dB
711  } else
712  g0 = FFMAX(*prev_tr_gain, fixed_gain *
713  (27536 * (1.0f / (1 << 15)))); // -1.5 dB
714 
715  *prev_tr_gain = g0; // update next frame threshold
716 
717  return sm_fac * g0 + (1 - sm_fac) * fixed_gain;
718 }
719 
726 static void pitch_enhancer(float *fixed_vector, float voice_fac)
727 {
728  int i;
729  float cpe = 0.125 * (1 + voice_fac);
730  float last = fixed_vector[0]; // holds c(i - 1)
731 
732  fixed_vector[0] -= cpe * fixed_vector[1];
733 
734  for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) {
735  float cur = fixed_vector[i];
736 
737  fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]);
738  last = cur;
739  }
740 
741  fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last;
742 }
743 
754 static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
755  float fixed_gain, const float *fixed_vector,
756  float *samples)
757 {
758  ff_weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
759  ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);
760 
761  /* emphasize pitch vector contribution in low bitrate modes */
762  if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {
763  int i;
764  float energy = ff_dot_productf(excitation, excitation,
766 
767  // XXX: Weird part in both ref code and spec. A unknown parameter
768  // {beta} seems to be identical to the current pitch gain
769  float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0];
770 
771  for (i = 0; i < AMRWB_SFR_SIZE; i++)
772  excitation[i] += pitch_factor * ctx->pitch_vector[i];
773 
774  ff_scale_vector_to_given_sum_of_squares(excitation, excitation,
775  energy, AMRWB_SFR_SIZE);
776  }
777 
778  ff_celp_lp_synthesis_filterf(samples, lpc, excitation,
780 }
781 
791 static void de_emphasis(float *out, float *in, float m, float mem[1])
792 {
793  int i;
794 
795  out[0] = in[0] + m * mem[0];
796 
797  for (i = 1; i < AMRWB_SFR_SIZE; i++)
798  out[i] = in[i] + out[i - 1] * m;
799 
800  mem[0] = out[AMRWB_SFR_SIZE - 1];
801 }
802 
811 static void upsample_5_4(float *out, const float *in, int o_size)
812 {
813  const float *in0 = in - UPS_FIR_SIZE + 1;
814  int i, j, k;
815  int int_part = 0, frac_part;
816 
817  i = 0;
818  for (j = 0; j < o_size / 5; j++) {
819  out[i] = in[int_part];
820  frac_part = 4;
821  i++;
822 
823  for (k = 1; k < 5; k++) {
824  out[i] = ff_dot_productf(in0 + int_part, upsample_fir[4 - frac_part],
825  UPS_MEM_SIZE);
826  int_part++;
827  frac_part--;
828  i++;
829  }
830  }
831 }
832 
842 static float find_hb_gain(AMRWBContext *ctx, const float *synth,
843  uint16_t hb_idx, uint8_t vad)
844 {
845  int wsp = (vad > 0);
846  float tilt;
847 
848  if (ctx->fr_cur_mode == MODE_23k85)
849  return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));
850 
851  tilt = ff_dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1) /
852  ff_dot_productf(synth, synth, AMRWB_SFR_SIZE);
853 
854  /* return gain bounded by [0.1, 1.0] */
855  return av_clipf((1.0 - FFMAX(0.0, tilt)) * (1.25 - 0.25 * wsp), 0.1, 1.0);
856 }
857 
867 static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,
868  const float *synth_exc, float hb_gain)
869 {
870  int i;
871  float energy = ff_dot_productf(synth_exc, synth_exc, AMRWB_SFR_SIZE);
872 
873  /* Generate a white-noise excitation */
874  for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
875  hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng);
876 
878  energy * hb_gain * hb_gain,
879  AMRWB_SFR_SIZE_16k);
880 }
881 
885 static float auto_correlation(float *diff_isf, float mean, int lag)
886 {
887  int i;
888  float sum = 0.0;
889 
890  for (i = 7; i < LP_ORDER - 2; i++) {
891  float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean);
892  sum += prod * prod;
893  }
894  return sum;
895 }
896 
904 static void extrapolate_isf(float isf[LP_ORDER_16k])
905 {
906  float diff_isf[LP_ORDER - 2], diff_mean;
907  float *diff_hi = diff_isf - LP_ORDER + 1; // diff array for extrapolated indexes
908  float corr_lag[3];
909  float est, scale;
910  int i, i_max_corr;
911 
912  isf[LP_ORDER_16k - 1] = isf[LP_ORDER - 1];
913 
914  /* Calculate the difference vector */
915  for (i = 0; i < LP_ORDER - 2; i++)
916  diff_isf[i] = isf[i + 1] - isf[i];
917 
918  diff_mean = 0.0;
919  for (i = 2; i < LP_ORDER - 2; i++)
920  diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4));
921 
922  /* Find which is the maximum autocorrelation */
923  i_max_corr = 0;
924  for (i = 0; i < 3; i++) {
925  corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2);
926 
927  if (corr_lag[i] > corr_lag[i_max_corr])
928  i_max_corr = i;
929  }
930  i_max_corr++;
931 
932  for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
933  isf[i] = isf[i - 1] + isf[i - 1 - i_max_corr]
934  - isf[i - 2 - i_max_corr];
935 
936  /* Calculate an estimate for ISF(18) and scale ISF based on the error */
937  est = 7965 + (isf[2] - isf[3] - isf[4]) / 6.0;
938  scale = 0.5 * (FFMIN(est, 7600) - isf[LP_ORDER - 2]) /
939  (isf[LP_ORDER_16k - 2] - isf[LP_ORDER - 2]);
940 
941  for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
942  diff_hi[i] = scale * (isf[i] - isf[i - 1]);
943 
944  /* Stability insurance */
945  for (i = LP_ORDER; i < LP_ORDER_16k - 1; i++)
946  if (diff_hi[i] + diff_hi[i - 1] < 5.0) {
947  if (diff_hi[i] > diff_hi[i - 1]) {
948  diff_hi[i - 1] = 5.0 - diff_hi[i];
949  } else
950  diff_hi[i] = 5.0 - diff_hi[i - 1];
951  }
952 
953  for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
954  isf[i] = isf[i - 1] + diff_hi[i] * (1.0f / (1 << 15));
955 
956  /* Scale the ISF vector for 16000 Hz */
957  for (i = 0; i < LP_ORDER_16k - 1; i++)
958  isf[i] *= 0.8;
959 }
960 
970 static void lpc_weighting(float *out, const float *lpc, float gamma, int size)
971 {
972  int i;
973  float fac = gamma;
974 
975  for (i = 0; i < size; i++) {
976  out[i] = lpc[i] * fac;
977  fac *= gamma;
978  }
979 }
980 
992 static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
993  const float *exc, const float *isf, const float *isf_past)
994 {
995  float hb_lpc[LP_ORDER_16k];
996  enum Mode mode = ctx->fr_cur_mode;
997 
998  if (mode == MODE_6k60) {
999  float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation
1000  double e_isp[LP_ORDER_16k];
1001 
1002  ff_weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
1003  1.0 - isfp_inter[subframe], LP_ORDER);
1004 
1005  extrapolate_isf(e_isf);
1006 
1007  e_isf[LP_ORDER_16k - 1] *= 2.0;
1008  ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k);
1009  ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k);
1010 
1011  lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k);
1012  } else {
1013  lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);
1014  }
1015 
1016  ff_celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
1017  (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);
1018 }
1019 
1031 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
1032  float mem[HB_FIR_SIZE], const float *in)
1033 {
1034  int i, j;
1035  float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples
1036 
1037  memcpy(data, mem, HB_FIR_SIZE * sizeof(float));
1038  memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float));
1039 
1040  for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) {
1041  out[i] = 0.0;
1042  for (j = 0; j <= HB_FIR_SIZE; j++)
1043  out[i] += data[i + j] * fir_coef[j];
1044  }
1045 
1046  memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));
1047 }
1048 
1053 {
1054  memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE],
1055  (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float));
1056 
1057  memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float));
1058  memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float));
1059 
1060  memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE],
1061  LP_ORDER * sizeof(float));
1062  memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE],
1063  UPS_MEM_SIZE * sizeof(float));
1064  memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k],
1065  LP_ORDER_16k * sizeof(float));
1066 }
1067 
1068 static int amrwb_decode_frame(AVCodecContext *avctx, void *data,
1069  int *got_frame_ptr, AVPacket *avpkt)
1070 {
1071  AMRWBContext *ctx = avctx->priv_data;
1072  AMRWBFrame *cf = &ctx->frame;
1073  const uint8_t *buf = avpkt->data;
1074  int buf_size = avpkt->size;
1075  int expected_fr_size, header_size;
1076  float *buf_out;
1077  float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing
1078  float fixed_gain_factor; // fixed gain correction factor (gamma)
1079  float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
1080  float synth_fixed_gain; // the fixed gain that synthesis should use
1081  float voice_fac, stab_fac; // parameters used for gain smoothing
1082  float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis
1083  float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band
1084  float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis
1085  float hb_gain;
1086  int sub, i, ret;
1087 
1088  /* get output buffer */
1090  if ((ret = avctx->get_buffer(avctx, &ctx->avframe)) < 0) {
1091  av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1092  return ret;
1093  }
1094  buf_out = (float *)ctx->avframe.data[0];
1095 
1096  header_size = decode_mime_header(ctx, buf);
1097  if (ctx->fr_cur_mode > MODE_SID) {
1098  av_log(avctx, AV_LOG_ERROR,
1099  "Invalid mode %d\n", ctx->fr_cur_mode);
1100  return AVERROR_INVALIDDATA;
1101  }
1102  expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1;
1103 
1104  if (buf_size < expected_fr_size) {
1105  av_log(avctx, AV_LOG_ERROR,
1106  "Frame too small (%d bytes). Truncated file?\n", buf_size);
1107  *got_frame_ptr = 0;
1108  return AVERROR_INVALIDDATA;
1109  }
1110 
1111  if (!ctx->fr_quality || ctx->fr_cur_mode > MODE_SID)
1112  av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n");
1113 
1114  if (ctx->fr_cur_mode == MODE_SID) { /* Comfort noise frame */
1115  av_log_missing_feature(avctx, "SID mode", 1);
1116  return -1;
1117  }
1118 
1119  ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame),
1120  buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]);
1121 
1122  /* Decode the quantized ISF vector */
1123  if (ctx->fr_cur_mode == MODE_6k60) {
1125  } else {
1127  }
1128 
1131 
1132  stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final);
1133 
1134  ctx->isf_cur[LP_ORDER - 1] *= 2.0;
1135  ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER);
1136 
1137  /* Generate a ISP vector for each subframe */
1138  if (ctx->first_frame) {
1139  ctx->first_frame = 0;
1140  memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double));
1141  }
1142  interpolate_isp(ctx->isp, ctx->isp_sub4_past);
1143 
1144  for (sub = 0; sub < 4; sub++)
1145  ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER);
1146 
1147  for (sub = 0; sub < 4; sub++) {
1148  const AMRWBSubFrame *cur_subframe = &cf->subframe[sub];
1149  float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k;
1150 
1151  /* Decode adaptive codebook (pitch vector) */
1152  decode_pitch_vector(ctx, cur_subframe, sub);
1153  /* Decode innovative codebook (fixed vector) */
1154  decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih,
1155  cur_subframe->pul_il, ctx->fr_cur_mode);
1156 
1157  pitch_sharpening(ctx, ctx->fixed_vector);
1158 
1159  decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode,
1160  &fixed_gain_factor, &ctx->pitch_gain[0]);
1161 
1162  ctx->fixed_gain[0] =
1163  ff_amr_set_fixed_gain(fixed_gain_factor,
1166  ctx->prediction_error,
1168 
1169  /* Calculate voice factor and store tilt for next subframe */
1170  voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],
1171  ctx->fixed_vector, ctx->fixed_gain[0]);
1172  ctx->tilt_coef = voice_fac * 0.25 + 0.25;
1173 
1174  /* Construct current excitation */
1175  for (i = 0; i < AMRWB_SFR_SIZE; i++) {
1176  ctx->excitation[i] *= ctx->pitch_gain[0];
1177  ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i];
1178  ctx->excitation[i] = truncf(ctx->excitation[i]);
1179  }
1180 
1181  /* Post-processing of excitation elements */
1182  synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain,
1183  voice_fac, stab_fac);
1184 
1185  synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector,
1186  spare_vector);
1187 
1188  pitch_enhancer(synth_fixed_vector, voice_fac);
1189 
1190  synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain,
1191  synth_fixed_vector, &ctx->samples_az[LP_ORDER]);
1192 
1193  /* Synthesis speech post-processing */
1195  &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);
1196 
1199  hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);
1200 
1201  upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],
1202  AMRWB_SFR_SIZE_16k);
1203 
1204  /* High frequency band (6.4 - 7.0 kHz) generation part */
1207  hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);
1208 
1209  hb_gain = find_hb_gain(ctx, hb_samples,
1210  cur_subframe->hb_gain, cf->vad);
1211 
1212  scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain);
1213 
1214  hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k],
1215  hb_exc, ctx->isf_cur, ctx->isf_past_final);
1216 
1217  /* High-band post-processing filters */
1218  hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem,
1219  &ctx->samples_hb[LP_ORDER_16k]);
1220 
1221  if (ctx->fr_cur_mode == MODE_23k85)
1222  hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem,
1223  hb_samples);
1224 
1225  /* Add the low and high frequency bands */
1226  for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
1227  sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15));
1228 
1229  /* Update buffers and history */
1230  update_sub_state(ctx);
1231  }
1232 
1233  /* update state for next frame */
1234  memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0]));
1235  memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float));
1236 
1237  *got_frame_ptr = 1;
1238  *(AVFrame *)data = ctx->avframe;
1239 
1240  return expected_fr_size;
1241 }
1242 
1244  .name = "amrwb",
1245  .type = AVMEDIA_TYPE_AUDIO,
1246  .id = CODEC_ID_AMR_WB,
1247  .priv_data_size = sizeof(AMRWBContext),
1250  .capabilities = CODEC_CAP_DR1,
1251  .long_name = NULL_IF_CONFIG_SMALL("Adaptive Multi-Rate WideBand"),
1252  .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_FLT,AV_SAMPLE_FMT_NONE},
1253 };