Libav
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 
28 #include "libavutil/common.h"
29 #include "libavutil/float_dsp.h"
30 #include "libavutil/lfg.h"
31 
32 #include "avcodec.h"
33 #include "lsp.h"
34 #include "celp_filters.h"
35 #include "acelp_filters.h"
36 #include "acelp_vectors.h"
37 #include "acelp_pitch_delay.h"
38 #include "internal.h"
39 
40 #define AMR_USE_16BIT_TABLES
41 #include "amr.h"
42 
43 #include "amrwbdata.h"
44 
45 typedef struct {
47  enum Mode fr_cur_mode;
49  float isf_cur[LP_ORDER];
50  float isf_q_past[LP_ORDER];
51  float isf_past_final[LP_ORDER];
52  double isp[4][LP_ORDER];
53  double isp_sub4_past[LP_ORDER];
54 
55  float lp_coef[4][LP_ORDER];
56 
59 
60  float excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 2 + AMRWB_SFR_SIZE];
61  float *excitation;
62 
63  float pitch_vector[AMRWB_SFR_SIZE];
64  float fixed_vector[AMRWB_SFR_SIZE];
65 
66  float prediction_error[4];
67  float pitch_gain[6];
68  float fixed_gain[2];
69 
70  float tilt_coef;
71 
74  float prev_tr_gain;
75 
76  float samples_az[LP_ORDER + AMRWB_SFR_SIZE];
77  float samples_up[UPS_MEM_SIZE + AMRWB_SFR_SIZE];
78  float samples_hb[LP_ORDER_16k + AMRWB_SFR_SIZE_16k];
79 
80  float hpf_31_mem[2], hpf_400_mem[2];
81  float demph_mem[1];
82  float bpf_6_7_mem[HB_FIR_SIZE];
83  float lpf_7_mem[HB_FIR_SIZE];
84 
87 } AMRWBContext;
88 
90 {
91  AMRWBContext *ctx = avctx->priv_data;
92  int i;
93 
94  if (avctx->channels > 1) {
95  avpriv_report_missing_feature(avctx, "multi-channel AMR");
96  return AVERROR_PATCHWELCOME;
97  }
98 
99  avctx->channels = 1;
101  avctx->sample_rate = 16000;
102  avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
103 
104  av_lfg_init(&ctx->prng, 1);
105 
107  ctx->first_frame = 1;
108 
109  for (i = 0; i < LP_ORDER; i++)
110  ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15));
111 
112  for (i = 0; i < 4; i++)
113  ctx->prediction_error[i] = MIN_ENERGY;
114 
115  return 0;
116 }
117 
127 static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf)
128 {
129  /* Decode frame header (1st octet) */
130  ctx->fr_cur_mode = buf[0] >> 3 & 0x0F;
131  ctx->fr_quality = (buf[0] & 0x4) == 0x4;
132 
133  return 1;
134 }
135 
143 static void decode_isf_indices_36b(uint16_t *ind, float *isf_q)
144 {
145  int i;
146 
147  for (i = 0; i < 9; i++)
148  isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
149 
150  for (i = 0; i < 7; i++)
151  isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
152 
153  for (i = 0; i < 5; i++)
154  isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15));
155 
156  for (i = 0; i < 4; i++)
157  isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15));
158 
159  for (i = 0; i < 7; i++)
160  isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15));
161 }
162 
170 static void decode_isf_indices_46b(uint16_t *ind, float *isf_q)
171 {
172  int i;
173 
174  for (i = 0; i < 9; i++)
175  isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
176 
177  for (i = 0; i < 7; i++)
178  isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
179 
180  for (i = 0; i < 3; i++)
181  isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15));
182 
183  for (i = 0; i < 3; i++)
184  isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15));
185 
186  for (i = 0; i < 3; i++)
187  isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15));
188 
189  for (i = 0; i < 3; i++)
190  isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15));
191 
192  for (i = 0; i < 4; i++)
193  isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15));
194 }
195 
204 static void isf_add_mean_and_past(float *isf_q, float *isf_past)
205 {
206  int i;
207  float tmp;
208 
209  for (i = 0; i < LP_ORDER; i++) {
210  tmp = isf_q[i];
211  isf_q[i] += isf_mean[i] * (1.0f / (1 << 15));
212  isf_q[i] += PRED_FACTOR * isf_past[i];
213  isf_past[i] = tmp;
214  }
215 }
216 
224 static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past)
225 {
226  int i, k;
227 
228  for (k = 0; k < 3; k++) {
229  float c = isfp_inter[k];
230  for (i = 0; i < LP_ORDER; i++)
231  isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i];
232  }
233 }
234 
246 static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index,
247  uint8_t *base_lag_int, int subframe)
248 {
249  if (subframe == 0 || subframe == 2) {
250  if (pitch_index < 376) {
251  *lag_int = (pitch_index + 137) >> 2;
252  *lag_frac = pitch_index - (*lag_int << 2) + 136;
253  } else if (pitch_index < 440) {
254  *lag_int = (pitch_index + 257 - 376) >> 1;
255  *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) << 1;
256  /* the actual resolution is 1/2 but expressed as 1/4 */
257  } else {
258  *lag_int = pitch_index - 280;
259  *lag_frac = 0;
260  }
261  /* minimum lag for next subframe */
262  *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
264  // XXX: the spec states clearly that *base_lag_int should be
265  // the nearest integer to *lag_int (minus 8), but the ref code
266  // actually always uses its floor, I'm following the latter
267  } else {
268  *lag_int = (pitch_index + 1) >> 2;
269  *lag_frac = pitch_index - (*lag_int << 2);
270  *lag_int += *base_lag_int;
271  }
272 }
273 
279 static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index,
280  uint8_t *base_lag_int, int subframe, enum Mode mode)
281 {
282  if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) {
283  if (pitch_index < 116) {
284  *lag_int = (pitch_index + 69) >> 1;
285  *lag_frac = (pitch_index - (*lag_int << 1) + 68) << 1;
286  } else {
287  *lag_int = pitch_index - 24;
288  *lag_frac = 0;
289  }
290  // XXX: same problem as before
291  *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
293  } else {
294  *lag_int = (pitch_index + 1) >> 1;
295  *lag_frac = (pitch_index - (*lag_int << 1)) << 1;
296  *lag_int += *base_lag_int;
297  }
298 }
299 
309  const AMRWBSubFrame *amr_subframe,
310  const int subframe)
311 {
312  int pitch_lag_int, pitch_lag_frac;
313  int i;
314  float *exc = ctx->excitation;
315  enum Mode mode = ctx->fr_cur_mode;
316 
317  if (mode <= MODE_8k85) {
318  decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
319  &ctx->base_pitch_lag, subframe, mode);
320  } else
321  decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
322  &ctx->base_pitch_lag, subframe);
323 
324  ctx->pitch_lag_int = pitch_lag_int;
325  pitch_lag_int += pitch_lag_frac > 0;
326 
327  /* Calculate the pitch vector by interpolating the past excitation at the
328  pitch lag using a hamming windowed sinc function */
329  ff_acelp_interpolatef(exc, exc + 1 - pitch_lag_int,
330  ac_inter, 4,
331  pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4),
332  LP_ORDER, AMRWB_SFR_SIZE + 1);
333 
334  /* Check which pitch signal path should be used
335  * 6k60 and 8k85 modes have the ltp flag set to 0 */
336  if (amr_subframe->ltp) {
337  memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float));
338  } else {
339  for (i = 0; i < AMRWB_SFR_SIZE; i++)
340  ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] +
341  0.18 * exc[i + 1];
342  memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float));
343  }
344 }
345 
347 #define BIT_STR(x,lsb,len) (((x) >> (lsb)) & ((1 << (len)) - 1))
348 
350 #define BIT_POS(x, p) (((x) >> (p)) & 1)
351 
365 static inline void decode_1p_track(int *out, int code, int m, int off)
366 {
367  int pos = BIT_STR(code, 0, m) + off;
368 
369  out[0] = BIT_POS(code, m) ? -pos : pos;
370 }
371 
372 static inline void decode_2p_track(int *out, int code, int m, int off)
373 {
374  int pos0 = BIT_STR(code, m, m) + off;
375  int pos1 = BIT_STR(code, 0, m) + off;
376 
377  out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0;
378  out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1;
379  out[1] = pos0 > pos1 ? -out[1] : out[1];
380 }
381 
382 static void decode_3p_track(int *out, int code, int m, int off)
383 {
384  int half_2p = BIT_POS(code, 2*m - 1) << (m - 1);
385 
386  decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
387  m - 1, off + half_2p);
388  decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off);
389 }
390 
391 static void decode_4p_track(int *out, int code, int m, int off)
392 {
393  int half_4p, subhalf_2p;
394  int b_offset = 1 << (m - 1);
395 
396  switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */
397  case 0: /* 0 pulses in A, 4 pulses in B or vice versa */
398  half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses
399  subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2);
400 
401  decode_2p_track(out, BIT_STR(code, 0, 2*m - 3),
402  m - 2, off + half_4p + subhalf_2p);
403  decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1),
404  m - 1, off + half_4p);
405  break;
406  case 1: /* 1 pulse in A, 3 pulses in B */
407  decode_1p_track(out, BIT_STR(code, 3*m - 2, m),
408  m - 1, off);
409  decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2),
410  m - 1, off + b_offset);
411  break;
412  case 2: /* 2 pulses in each half */
413  decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1),
414  m - 1, off);
415  decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1),
416  m - 1, off + b_offset);
417  break;
418  case 3: /* 3 pulses in A, 1 pulse in B */
419  decode_3p_track(out, BIT_STR(code, m, 3*m - 2),
420  m - 1, off);
421  decode_1p_track(out + 3, BIT_STR(code, 0, m),
422  m - 1, off + b_offset);
423  break;
424  }
425 }
426 
427 static void decode_5p_track(int *out, int code, int m, int off)
428 {
429  int half_3p = BIT_POS(code, 5*m - 1) << (m - 1);
430 
431  decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2),
432  m - 1, off + half_3p);
433 
434  decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off);
435 }
436 
437 static void decode_6p_track(int *out, int code, int m, int off)
438 {
439  int b_offset = 1 << (m - 1);
440  /* which half has more pulses in cases 0 to 2 */
441  int half_more = BIT_POS(code, 6*m - 5) << (m - 1);
442  int half_other = b_offset - half_more;
443 
444  switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */
445  case 0: /* 0 pulses in A, 6 pulses in B or vice versa */
446  decode_1p_track(out, BIT_STR(code, 0, m),
447  m - 1, off + half_more);
448  decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
449  m - 1, off + half_more);
450  break;
451  case 1: /* 1 pulse in A, 5 pulses in B or vice versa */
452  decode_1p_track(out, BIT_STR(code, 0, m),
453  m - 1, off + half_other);
454  decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
455  m - 1, off + half_more);
456  break;
457  case 2: /* 2 pulses in A, 4 pulses in B or vice versa */
458  decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
459  m - 1, off + half_other);
460  decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4),
461  m - 1, off + half_more);
462  break;
463  case 3: /* 3 pulses in A, 3 pulses in B */
464  decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2),
465  m - 1, off);
466  decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2),
467  m - 1, off + b_offset);
468  break;
469  }
470 }
471 
481 static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi,
482  const uint16_t *pulse_lo, const enum Mode mode)
483 {
484  /* sig_pos stores for each track the decoded pulse position indexes
485  * (1-based) multiplied by its corresponding amplitude (+1 or -1) */
486  int sig_pos[4][6];
487  int spacing = (mode == MODE_6k60) ? 2 : 4;
488  int i, j;
489 
490  switch (mode) {
491  case MODE_6k60:
492  for (i = 0; i < 2; i++)
493  decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1);
494  break;
495  case MODE_8k85:
496  for (i = 0; i < 4; i++)
497  decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1);
498  break;
499  case MODE_12k65:
500  for (i = 0; i < 4; i++)
501  decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
502  break;
503  case MODE_14k25:
504  for (i = 0; i < 2; i++)
505  decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
506  for (i = 2; i < 4; i++)
507  decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
508  break;
509  case MODE_15k85:
510  for (i = 0; i < 4; i++)
511  decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
512  break;
513  case MODE_18k25:
514  for (i = 0; i < 4; i++)
515  decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
516  ((int) pulse_hi[i] << 14), 4, 1);
517  break;
518  case MODE_19k85:
519  for (i = 0; i < 2; i++)
520  decode_5p_track(sig_pos[i], (int) pulse_lo[i] +
521  ((int) pulse_hi[i] << 10), 4, 1);
522  for (i = 2; i < 4; i++)
523  decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
524  ((int) pulse_hi[i] << 14), 4, 1);
525  break;
526  case MODE_23k05:
527  case MODE_23k85:
528  for (i = 0; i < 4; i++)
529  decode_6p_track(sig_pos[i], (int) pulse_lo[i] +
530  ((int) pulse_hi[i] << 11), 4, 1);
531  break;
532  }
533 
534  memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE);
535 
536  for (i = 0; i < 4; i++)
537  for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) {
538  int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i;
539 
540  fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0;
541  }
542 }
543 
552 static void decode_gains(const uint8_t vq_gain, const enum Mode mode,
553  float *fixed_gain_factor, float *pitch_gain)
554 {
555  const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] :
556  qua_gain_7b[vq_gain]);
557 
558  *pitch_gain = gains[0] * (1.0f / (1 << 14));
559  *fixed_gain_factor = gains[1] * (1.0f / (1 << 11));
560 }
561 
568 // XXX: Spec states this procedure should be applied when the pitch
569 // lag is less than 64, but this checking seems absent in reference and AMR-NB
570 static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector)
571 {
572  int i;
573 
574  /* Tilt part */
575  for (i = AMRWB_SFR_SIZE - 1; i != 0; i--)
576  fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef;
577 
578  /* Periodicity enhancement part */
579  for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++)
580  fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85;
581 }
582 
589 // XXX: There is something wrong with the precision here! The magnitudes
590 // of the energies are not correct. Please check the reference code carefully
591 static float voice_factor(float *p_vector, float p_gain,
592  float *f_vector, float f_gain)
593 {
594  double p_ener = (double) avpriv_scalarproduct_float_c(p_vector, p_vector,
595  AMRWB_SFR_SIZE) *
596  p_gain * p_gain;
597  double f_ener = (double) avpriv_scalarproduct_float_c(f_vector, f_vector,
598  AMRWB_SFR_SIZE) *
599  f_gain * f_gain;
600 
601  return (p_ener - f_ener) / (p_ener + f_ener);
602 }
603 
614 static float *anti_sparseness(AMRWBContext *ctx,
615  float *fixed_vector, float *buf)
616 {
617  int ir_filter_nr;
618 
619  if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes
620  return fixed_vector;
621 
622  if (ctx->pitch_gain[0] < 0.6) {
623  ir_filter_nr = 0; // strong filtering
624  } else if (ctx->pitch_gain[0] < 0.9) {
625  ir_filter_nr = 1; // medium filtering
626  } else
627  ir_filter_nr = 2; // no filtering
628 
629  /* detect 'onset' */
630  if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) {
631  if (ir_filter_nr < 2)
632  ir_filter_nr++;
633  } else {
634  int i, count = 0;
635 
636  for (i = 0; i < 6; i++)
637  if (ctx->pitch_gain[i] < 0.6)
638  count++;
639 
640  if (count > 2)
641  ir_filter_nr = 0;
642 
643  if (ir_filter_nr > ctx->prev_ir_filter_nr + 1)
644  ir_filter_nr--;
645  }
646 
647  /* update ir filter strength history */
648  ctx->prev_ir_filter_nr = ir_filter_nr;
649 
650  ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85);
651 
652  if (ir_filter_nr < 2) {
653  int i;
654  const float *coef = ir_filters_lookup[ir_filter_nr];
655 
656  /* Circular convolution code in the reference
657  * decoder was modified to avoid using one
658  * extra array. The filtered vector is given by:
659  *
660  * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) }
661  */
662 
663  memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE);
664  for (i = 0; i < AMRWB_SFR_SIZE; i++)
665  if (fixed_vector[i])
666  ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i],
667  AMRWB_SFR_SIZE);
668  fixed_vector = buf;
669  }
670 
671  return fixed_vector;
672 }
673 
678 static float stability_factor(const float *isf, const float *isf_past)
679 {
680  int i;
681  float acc = 0.0;
682 
683  for (i = 0; i < LP_ORDER - 1; i++)
684  acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]);
685 
686  // XXX: This part is not so clear from the reference code
687  // the result is more accurate changing the "/ 256" to "* 512"
688  return FFMAX(0.0, 1.25 - acc * 0.8 * 512);
689 }
690 
702 static float noise_enhancer(float fixed_gain, float *prev_tr_gain,
703  float voice_fac, float stab_fac)
704 {
705  float sm_fac = 0.5 * (1 - voice_fac) * stab_fac;
706  float g0;
707 
708  // XXX: the following fixed-point constants used to in(de)crement
709  // gain by 1.5dB were taken from the reference code, maybe it could
710  // be simpler
711  if (fixed_gain < *prev_tr_gain) {
712  g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain *
713  (6226 * (1.0f / (1 << 15)))); // +1.5 dB
714  } else
715  g0 = FFMAX(*prev_tr_gain, fixed_gain *
716  (27536 * (1.0f / (1 << 15)))); // -1.5 dB
717 
718  *prev_tr_gain = g0; // update next frame threshold
719 
720  return sm_fac * g0 + (1 - sm_fac) * fixed_gain;
721 }
722 
729 static void pitch_enhancer(float *fixed_vector, float voice_fac)
730 {
731  int i;
732  float cpe = 0.125 * (1 + voice_fac);
733  float last = fixed_vector[0]; // holds c(i - 1)
734 
735  fixed_vector[0] -= cpe * fixed_vector[1];
736 
737  for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) {
738  float cur = fixed_vector[i];
739 
740  fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]);
741  last = cur;
742  }
743 
744  fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last;
745 }
746 
757 static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
758  float fixed_gain, const float *fixed_vector,
759  float *samples)
760 {
761  ff_weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
762  ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);
763 
764  /* emphasize pitch vector contribution in low bitrate modes */
765  if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {
766  int i;
767  float energy = avpriv_scalarproduct_float_c(excitation, excitation,
769 
770  // XXX: Weird part in both ref code and spec. A unknown parameter
771  // {beta} seems to be identical to the current pitch gain
772  float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0];
773 
774  for (i = 0; i < AMRWB_SFR_SIZE; i++)
775  excitation[i] += pitch_factor * ctx->pitch_vector[i];
776 
777  ff_scale_vector_to_given_sum_of_squares(excitation, excitation,
778  energy, AMRWB_SFR_SIZE);
779  }
780 
781  ff_celp_lp_synthesis_filterf(samples, lpc, excitation,
783 }
784 
794 static void de_emphasis(float *out, float *in, float m, float mem[1])
795 {
796  int i;
797 
798  out[0] = in[0] + m * mem[0];
799 
800  for (i = 1; i < AMRWB_SFR_SIZE; i++)
801  out[i] = in[i] + out[i - 1] * m;
802 
803  mem[0] = out[AMRWB_SFR_SIZE - 1];
804 }
805 
814 static void upsample_5_4(float *out, const float *in, int o_size)
815 {
816  const float *in0 = in - UPS_FIR_SIZE + 1;
817  int i, j, k;
818  int int_part = 0, frac_part;
819 
820  i = 0;
821  for (j = 0; j < o_size / 5; j++) {
822  out[i] = in[int_part];
823  frac_part = 4;
824  i++;
825 
826  for (k = 1; k < 5; k++) {
827  out[i] = avpriv_scalarproduct_float_c(in0 + int_part,
828  upsample_fir[4 - frac_part],
829  UPS_MEM_SIZE);
830  int_part++;
831  frac_part--;
832  i++;
833  }
834  }
835 }
836 
846 static float find_hb_gain(AMRWBContext *ctx, const float *synth,
847  uint16_t hb_idx, uint8_t vad)
848 {
849  int wsp = (vad > 0);
850  float tilt;
851 
852  if (ctx->fr_cur_mode == MODE_23k85)
853  return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));
854 
855  tilt = avpriv_scalarproduct_float_c(synth, synth + 1, AMRWB_SFR_SIZE - 1) /
857 
858  /* return gain bounded by [0.1, 1.0] */
859  return av_clipf((1.0 - FFMAX(0.0, tilt)) * (1.25 - 0.25 * wsp), 0.1, 1.0);
860 }
861 
871 static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,
872  const float *synth_exc, float hb_gain)
873 {
874  int i;
875  float energy = avpriv_scalarproduct_float_c(synth_exc, synth_exc,
877 
878  /* Generate a white-noise excitation */
879  for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
880  hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng);
881 
883  energy * hb_gain * hb_gain,
884  AMRWB_SFR_SIZE_16k);
885 }
886 
890 static float auto_correlation(float *diff_isf, float mean, int lag)
891 {
892  int i;
893  float sum = 0.0;
894 
895  for (i = 7; i < LP_ORDER - 2; i++) {
896  float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean);
897  sum += prod * prod;
898  }
899  return sum;
900 }
901 
909 static void extrapolate_isf(float isf[LP_ORDER_16k])
910 {
911  float diff_isf[LP_ORDER - 2], diff_mean;
912  float corr_lag[3];
913  float est, scale;
914  int i, j, i_max_corr;
915 
916  isf[LP_ORDER_16k - 1] = isf[LP_ORDER - 1];
917 
918  /* Calculate the difference vector */
919  for (i = 0; i < LP_ORDER - 2; i++)
920  diff_isf[i] = isf[i + 1] - isf[i];
921 
922  diff_mean = 0.0;
923  for (i = 2; i < LP_ORDER - 2; i++)
924  diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4));
925 
926  /* Find which is the maximum autocorrelation */
927  i_max_corr = 0;
928  for (i = 0; i < 3; i++) {
929  corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2);
930 
931  if (corr_lag[i] > corr_lag[i_max_corr])
932  i_max_corr = i;
933  }
934  i_max_corr++;
935 
936  for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
937  isf[i] = isf[i - 1] + isf[i - 1 - i_max_corr]
938  - isf[i - 2 - i_max_corr];
939 
940  /* Calculate an estimate for ISF(18) and scale ISF based on the error */
941  est = 7965 + (isf[2] - isf[3] - isf[4]) / 6.0;
942  scale = 0.5 * (FFMIN(est, 7600) - isf[LP_ORDER - 2]) /
943  (isf[LP_ORDER_16k - 2] - isf[LP_ORDER - 2]);
944 
945  for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
946  diff_isf[j] = scale * (isf[i] - isf[i - 1]);
947 
948  /* Stability insurance */
949  for (i = 1; i < LP_ORDER_16k - LP_ORDER; i++)
950  if (diff_isf[i] + diff_isf[i - 1] < 5.0) {
951  if (diff_isf[i] > diff_isf[i - 1]) {
952  diff_isf[i - 1] = 5.0 - diff_isf[i];
953  } else
954  diff_isf[i] = 5.0 - diff_isf[i - 1];
955  }
956 
957  for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
958  isf[i] = isf[i - 1] + diff_isf[j] * (1.0f / (1 << 15));
959 
960  /* Scale the ISF vector for 16000 Hz */
961  for (i = 0; i < LP_ORDER_16k - 1; i++)
962  isf[i] *= 0.8;
963 }
964 
974 static void lpc_weighting(float *out, const float *lpc, float gamma, int size)
975 {
976  int i;
977  float fac = gamma;
978 
979  for (i = 0; i < size; i++) {
980  out[i] = lpc[i] * fac;
981  fac *= gamma;
982  }
983 }
984 
996 static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
997  const float *exc, const float *isf, const float *isf_past)
998 {
999  float hb_lpc[LP_ORDER_16k];
1000  enum Mode mode = ctx->fr_cur_mode;
1001 
1002  if (mode == MODE_6k60) {
1003  float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation
1004  double e_isp[LP_ORDER_16k];
1005 
1006  ff_weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
1007  1.0 - isfp_inter[subframe], LP_ORDER);
1008 
1009  extrapolate_isf(e_isf);
1010 
1011  e_isf[LP_ORDER_16k - 1] *= 2.0;
1012  ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k);
1013  ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k);
1014 
1015  lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k);
1016  } else {
1017  lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);
1018  }
1019 
1020  ff_celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
1021  (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);
1022 }
1023 
1035 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
1036  float mem[HB_FIR_SIZE], const float *in)
1037 {
1038  int i, j;
1039  float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples
1040 
1041  memcpy(data, mem, HB_FIR_SIZE * sizeof(float));
1042  memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float));
1043 
1044  for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) {
1045  out[i] = 0.0;
1046  for (j = 0; j <= HB_FIR_SIZE; j++)
1047  out[i] += data[i + j] * fir_coef[j];
1048  }
1049 
1050  memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));
1051 }
1052 
1057 {
1058  memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE],
1059  (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float));
1060 
1061  memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float));
1062  memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float));
1063 
1064  memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE],
1065  LP_ORDER * sizeof(float));
1066  memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE],
1067  UPS_MEM_SIZE * sizeof(float));
1068  memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k],
1069  LP_ORDER_16k * sizeof(float));
1070 }
1071 
1072 static int amrwb_decode_frame(AVCodecContext *avctx, void *data,
1073  int *got_frame_ptr, AVPacket *avpkt)
1074 {
1075  AMRWBContext *ctx = avctx->priv_data;
1076  AVFrame *frame = data;
1077  AMRWBFrame *cf = &ctx->frame;
1078  const uint8_t *buf = avpkt->data;
1079  int buf_size = avpkt->size;
1080  int expected_fr_size, header_size;
1081  float *buf_out;
1082  float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing
1083  float fixed_gain_factor; // fixed gain correction factor (gamma)
1084  float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
1085  float synth_fixed_gain; // the fixed gain that synthesis should use
1086  float voice_fac, stab_fac; // parameters used for gain smoothing
1087  float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis
1088  float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band
1089  float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis
1090  float hb_gain;
1091  int sub, i, ret;
1092 
1093  /* get output buffer */
1094  frame->nb_samples = 4 * AMRWB_SFR_SIZE_16k;
1095  if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) {
1096  av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1097  return ret;
1098  }
1099  buf_out = (float *)frame->data[0];
1100 
1101  header_size = decode_mime_header(ctx, buf);
1102  if (ctx->fr_cur_mode > MODE_SID) {
1103  av_log(avctx, AV_LOG_ERROR,
1104  "Invalid mode %d\n", ctx->fr_cur_mode);
1105  return AVERROR_INVALIDDATA;
1106  }
1107  expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1;
1108 
1109  if (buf_size < expected_fr_size) {
1110  av_log(avctx, AV_LOG_ERROR,
1111  "Frame too small (%d bytes). Truncated file?\n", buf_size);
1112  *got_frame_ptr = 0;
1113  return AVERROR_INVALIDDATA;
1114  }
1115 
1116  if (!ctx->fr_quality || ctx->fr_cur_mode > MODE_SID)
1117  av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n");
1118 
1119  if (ctx->fr_cur_mode == MODE_SID) { /* Comfort noise frame */
1120  avpriv_request_sample(avctx, "SID mode");
1121  return AVERROR_PATCHWELCOME;
1122  }
1123 
1124  ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame),
1125  buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]);
1126 
1127  /* Decode the quantized ISF vector */
1128  if (ctx->fr_cur_mode == MODE_6k60) {
1130  } else {
1132  }
1133 
1136 
1137  stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final);
1138 
1139  ctx->isf_cur[LP_ORDER - 1] *= 2.0;
1140  ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER);
1141 
1142  /* Generate a ISP vector for each subframe */
1143  if (ctx->first_frame) {
1144  ctx->first_frame = 0;
1145  memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double));
1146  }
1147  interpolate_isp(ctx->isp, ctx->isp_sub4_past);
1148 
1149  for (sub = 0; sub < 4; sub++)
1150  ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER);
1151 
1152  for (sub = 0; sub < 4; sub++) {
1153  const AMRWBSubFrame *cur_subframe = &cf->subframe[sub];
1154  float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k;
1155 
1156  /* Decode adaptive codebook (pitch vector) */
1157  decode_pitch_vector(ctx, cur_subframe, sub);
1158  /* Decode innovative codebook (fixed vector) */
1159  decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih,
1160  cur_subframe->pul_il, ctx->fr_cur_mode);
1161 
1162  pitch_sharpening(ctx, ctx->fixed_vector);
1163 
1164  decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode,
1165  &fixed_gain_factor, &ctx->pitch_gain[0]);
1166 
1167  ctx->fixed_gain[0] =
1168  ff_amr_set_fixed_gain(fixed_gain_factor,
1170  ctx->fixed_vector,
1171  AMRWB_SFR_SIZE) /
1173  ctx->prediction_error,
1175 
1176  /* Calculate voice factor and store tilt for next subframe */
1177  voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],
1178  ctx->fixed_vector, ctx->fixed_gain[0]);
1179  ctx->tilt_coef = voice_fac * 0.25 + 0.25;
1180 
1181  /* Construct current excitation */
1182  for (i = 0; i < AMRWB_SFR_SIZE; i++) {
1183  ctx->excitation[i] *= ctx->pitch_gain[0];
1184  ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i];
1185  ctx->excitation[i] = truncf(ctx->excitation[i]);
1186  }
1187 
1188  /* Post-processing of excitation elements */
1189  synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain,
1190  voice_fac, stab_fac);
1191 
1192  synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector,
1193  spare_vector);
1194 
1195  pitch_enhancer(synth_fixed_vector, voice_fac);
1196 
1197  synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain,
1198  synth_fixed_vector, &ctx->samples_az[LP_ORDER]);
1199 
1200  /* Synthesis speech post-processing */
1202  &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);
1203 
1206  hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);
1207 
1208  upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],
1209  AMRWB_SFR_SIZE_16k);
1210 
1211  /* High frequency band (6.4 - 7.0 kHz) generation part */
1214  hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);
1215 
1216  hb_gain = find_hb_gain(ctx, hb_samples,
1217  cur_subframe->hb_gain, cf->vad);
1218 
1219  scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain);
1220 
1221  hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k],
1222  hb_exc, ctx->isf_cur, ctx->isf_past_final);
1223 
1224  /* High-band post-processing filters */
1225  hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem,
1226  &ctx->samples_hb[LP_ORDER_16k]);
1227 
1228  if (ctx->fr_cur_mode == MODE_23k85)
1229  hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem,
1230  hb_samples);
1231 
1232  /* Add the low and high frequency bands */
1233  for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
1234  sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15));
1235 
1236  /* Update buffers and history */
1237  update_sub_state(ctx);
1238  }
1239 
1240  /* update state for next frame */
1241  memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0]));
1242  memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float));
1243 
1244  *got_frame_ptr = 1;
1245 
1246  return expected_fr_size;
1247 }
1248 
1250  .name = "amrwb",
1251  .long_name = NULL_IF_CONFIG_SMALL("AMR-WB (Adaptive Multi-Rate WideBand)"),
1252  .type = AVMEDIA_TYPE_AUDIO,
1253  .id = AV_CODEC_ID_AMR_WB,
1254  .priv_data_size = sizeof(AMRWBContext),
1257  .capabilities = CODEC_CAP_DR1,
1258  .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,
1260 };