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 "internal.h"
31 #include "get_bits.h"
32 #include "lsp.h"
33 #include "celp_math.h"
34 #include "celp_filters.h"
35 #include "acelp_filters.h"
36 #include "acelp_vectors.h"
37 #include "acelp_pitch_delay.h"
38 
39 #define AMR_USE_16BIT_TABLES
40 #include "amr.h"
41 
42 #include "amrwbdata.h"
43 
44 typedef struct {
48  uint8_t fr_quality;
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 
57  uint8_t base_pitch_lag;
58  uint8_t pitch_lag_int;
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 
86  uint8_t first_frame;
87 } AMRWBContext;
88 
90 {
91  AMRWBContext *ctx = avctx->priv_data;
92  int i;
93 
95 
96  av_lfg_init(&ctx->prng, 1);
97 
99  ctx->first_frame = 1;
100 
101  for (i = 0; i < LP_ORDER; i++)
102  ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15));
103 
104  for (i = 0; i < 4; i++)
105  ctx->prediction_error[i] = MIN_ENERGY;
106 
108  avctx->coded_frame = &ctx->avframe;
109 
110  return 0;
111 }
112 
122 static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf)
123 {
124  GetBitContext gb;
125  init_get_bits(&gb, buf, 8);
126 
127  /* Decode frame header (1st octet) */
128  skip_bits(&gb, 1); // padding bit
129  ctx->fr_cur_mode = get_bits(&gb, 4);
130  ctx->fr_quality = get_bits1(&gb);
131  skip_bits(&gb, 2); // padding bits
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) ff_dot_productf(p_vector, p_vector,
595  AMRWB_SFR_SIZE) * p_gain * p_gain;
596  double f_ener = (double) ff_dot_productf(f_vector, f_vector,
597  AMRWB_SFR_SIZE) * f_gain * f_gain;
598 
599  return (p_ener - f_ener) / (p_ener + f_ener);
600 }
601 
612 static float *anti_sparseness(AMRWBContext *ctx,
613  float *fixed_vector, float *buf)
614 {
615  int ir_filter_nr;
616 
617  if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes
618  return fixed_vector;
619 
620  if (ctx->pitch_gain[0] < 0.6) {
621  ir_filter_nr = 0; // strong filtering
622  } else if (ctx->pitch_gain[0] < 0.9) {
623  ir_filter_nr = 1; // medium filtering
624  } else
625  ir_filter_nr = 2; // no filtering
626 
627  /* detect 'onset' */
628  if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) {
629  if (ir_filter_nr < 2)
630  ir_filter_nr++;
631  } else {
632  int i, count = 0;
633 
634  for (i = 0; i < 6; i++)
635  if (ctx->pitch_gain[i] < 0.6)
636  count++;
637 
638  if (count > 2)
639  ir_filter_nr = 0;
640 
641  if (ir_filter_nr > ctx->prev_ir_filter_nr + 1)
642  ir_filter_nr--;
643  }
644 
645  /* update ir filter strength history */
646  ctx->prev_ir_filter_nr = ir_filter_nr;
647 
648  ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85);
649 
650  if (ir_filter_nr < 2) {
651  int i;
652  const float *coef = ir_filters_lookup[ir_filter_nr];
653 
654  /* Circular convolution code in the reference
655  * decoder was modified to avoid using one
656  * extra array. The filtered vector is given by:
657  *
658  * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) }
659  */
660 
661  memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE);
662  for (i = 0; i < AMRWB_SFR_SIZE; i++)
663  if (fixed_vector[i])
664  ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i],
665  AMRWB_SFR_SIZE);
666  fixed_vector = buf;
667  }
668 
669  return fixed_vector;
670 }
671 
676 static float stability_factor(const float *isf, const float *isf_past)
677 {
678  int i;
679  float acc = 0.0;
680 
681  for (i = 0; i < LP_ORDER - 1; i++)
682  acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]);
683 
684  // XXX: This part is not so clear from the reference code
685  // the result is more accurate changing the "/ 256" to "* 512"
686  return FFMAX(0.0, 1.25 - acc * 0.8 * 512);
687 }
688 
700 static float noise_enhancer(float fixed_gain, float *prev_tr_gain,
701  float voice_fac, float stab_fac)
702 {
703  float sm_fac = 0.5 * (1 - voice_fac) * stab_fac;
704  float g0;
705 
706  // XXX: the following fixed-point constants used to in(de)crement
707  // gain by 1.5dB were taken from the reference code, maybe it could
708  // be simpler
709  if (fixed_gain < *prev_tr_gain) {
710  g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain *
711  (6226 * (1.0f / (1 << 15)))); // +1.5 dB
712  } else
713  g0 = FFMAX(*prev_tr_gain, fixed_gain *
714  (27536 * (1.0f / (1 << 15)))); // -1.5 dB
715 
716  *prev_tr_gain = g0; // update next frame threshold
717 
718  return sm_fac * g0 + (1 - sm_fac) * fixed_gain;
719 }
720 
727 static void pitch_enhancer(float *fixed_vector, float voice_fac)
728 {
729  int i;
730  float cpe = 0.125 * (1 + voice_fac);
731  float last = fixed_vector[0]; // holds c(i - 1)
732 
733  fixed_vector[0] -= cpe * fixed_vector[1];
734 
735  for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) {
736  float cur = fixed_vector[i];
737 
738  fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]);
739  last = cur;
740  }
741 
742  fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last;
743 }
744 
755 static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
756  float fixed_gain, const float *fixed_vector,
757  float *samples)
758 {
759  ff_weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
760  ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);
761 
762  /* emphasize pitch vector contribution in low bitrate modes */
763  if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {
764  int i;
765  float energy = ff_dot_productf(excitation, excitation,
767 
768  // XXX: Weird part in both ref code and spec. A unknown parameter
769  // {beta} seems to be identical to the current pitch gain
770  float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0];
771 
772  for (i = 0; i < AMRWB_SFR_SIZE; i++)
773  excitation[i] += pitch_factor * ctx->pitch_vector[i];
774 
775  ff_scale_vector_to_given_sum_of_squares(excitation, excitation,
776  energy, AMRWB_SFR_SIZE);
777  }
778 
779  ff_celp_lp_synthesis_filterf(samples, lpc, excitation,
781 }
782 
792 static void de_emphasis(float *out, float *in, float m, float mem[1])
793 {
794  int i;
795 
796  out[0] = in[0] + m * mem[0];
797 
798  for (i = 1; i < AMRWB_SFR_SIZE; i++)
799  out[i] = in[i] + out[i - 1] * m;
800 
801  mem[0] = out[AMRWB_SFR_SIZE - 1];
802 }
803 
812 static void upsample_5_4(float *out, const float *in, int o_size)
813 {
814  const float *in0 = in - UPS_FIR_SIZE + 1;
815  int i, j, k;
816  int int_part = 0, frac_part;
817 
818  i = 0;
819  for (j = 0; j < o_size / 5; j++) {
820  out[i] = in[int_part];
821  frac_part = 4;
822  i++;
823 
824  for (k = 1; k < 5; k++) {
825  out[i] = ff_dot_productf(in0 + int_part, upsample_fir[4 - frac_part],
826  UPS_MEM_SIZE);
827  int_part++;
828  frac_part--;
829  i++;
830  }
831  }
832 }
833 
843 static float find_hb_gain(AMRWBContext *ctx, const float *synth,
844  uint16_t hb_idx, uint8_t vad)
845 {
846  int wsp = (vad > 0);
847  float tilt;
848 
849  if (ctx->fr_cur_mode == MODE_23k85)
850  return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));
851 
852  tilt = ff_dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1) /
853  ff_dot_productf(synth, synth, AMRWB_SFR_SIZE);
854 
855  /* return gain bounded by [0.1, 1.0] */
856  return av_clipf((1.0 - FFMAX(0.0, tilt)) * (1.25 - 0.25 * wsp), 0.1, 1.0);
857 }
858 
868 static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,
869  const float *synth_exc, float hb_gain)
870 {
871  int i;
872  float energy = ff_dot_productf(synth_exc, synth_exc, AMRWB_SFR_SIZE);
873 
874  /* Generate a white-noise excitation */
875  for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
876  hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng);
877 
879  energy * hb_gain * hb_gain,
880  AMRWB_SFR_SIZE_16k);
881 }
882 
886 static float auto_correlation(float *diff_isf, float mean, int lag)
887 {
888  int i;
889  float sum = 0.0;
890 
891  for (i = 7; i < LP_ORDER - 2; i++) {
892  float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean);
893  sum += prod * prod;
894  }
895  return sum;
896 }
897 
905 static void extrapolate_isf(float isf[LP_ORDER_16k])
906 {
907  float diff_isf[LP_ORDER - 2], diff_mean;
908  float *diff_hi = diff_isf - LP_ORDER + 1; // diff array for extrapolated indexes
909  float corr_lag[3];
910  float est, scale;
911  int i, i_max_corr;
912 
913  isf[LP_ORDER_16k - 1] = isf[LP_ORDER - 1];
914 
915  /* Calculate the difference vector */
916  for (i = 0; i < LP_ORDER - 2; i++)
917  diff_isf[i] = isf[i + 1] - isf[i];
918 
919  diff_mean = 0.0;
920  for (i = 2; i < LP_ORDER - 2; i++)
921  diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4));
922 
923  /* Find which is the maximum autocorrelation */
924  i_max_corr = 0;
925  for (i = 0; i < 3; i++) {
926  corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2);
927 
928  if (corr_lag[i] > corr_lag[i_max_corr])
929  i_max_corr = i;
930  }
931  i_max_corr++;
932 
933  for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
934  isf[i] = isf[i - 1] + isf[i - 1 - i_max_corr]
935  - isf[i - 2 - i_max_corr];
936 
937  /* Calculate an estimate for ISF(18) and scale ISF based on the error */
938  est = 7965 + (isf[2] - isf[3] - isf[4]) / 6.0;
939  scale = 0.5 * (FFMIN(est, 7600) - isf[LP_ORDER - 2]) /
940  (isf[LP_ORDER_16k - 2] - isf[LP_ORDER - 2]);
941 
942  for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
943  diff_hi[i] = scale * (isf[i] - isf[i - 1]);
944 
945  /* Stability insurance */
946  for (i = LP_ORDER; i < LP_ORDER_16k - 1; i++)
947  if (diff_hi[i] + diff_hi[i - 1] < 5.0) {
948  if (diff_hi[i] > diff_hi[i - 1]) {
949  diff_hi[i - 1] = 5.0 - diff_hi[i];
950  } else
951  diff_hi[i] = 5.0 - diff_hi[i - 1];
952  }
953 
954  for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
955  isf[i] = isf[i - 1] + diff_hi[i] * (1.0f / (1 << 15));
956 
957  /* Scale the ISF vector for 16000 Hz */
958  for (i = 0; i < LP_ORDER_16k - 1; i++)
959  isf[i] *= 0.8;
960 }
961 
971 static void lpc_weighting(float *out, const float *lpc, float gamma, int size)
972 {
973  int i;
974  float fac = gamma;
975 
976  for (i = 0; i < size; i++) {
977  out[i] = lpc[i] * fac;
978  fac *= gamma;
979  }
980 }
981 
993 static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
994  const float *exc, const float *isf, const float *isf_past)
995 {
996  float hb_lpc[LP_ORDER_16k];
997  enum Mode mode = ctx->fr_cur_mode;
998 
999  if (mode == MODE_6k60) {
1000  float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation
1001  double e_isp[LP_ORDER_16k];
1002 
1003  ff_weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
1004  1.0 - isfp_inter[subframe], LP_ORDER);
1005 
1006  extrapolate_isf(e_isf);
1007 
1008  e_isf[LP_ORDER_16k - 1] *= 2.0;
1009  ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k);
1010  ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k);
1011 
1012  lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k);
1013  } else {
1014  lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);
1015  }
1016 
1017  ff_celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
1018  (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);
1019 }
1020 
1032 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
1033  float mem[HB_FIR_SIZE], const float *in)
1034 {
1035  int i, j;
1036  float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples
1037 
1038  memcpy(data, mem, HB_FIR_SIZE * sizeof(float));
1039  memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float));
1040 
1041  for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) {
1042  out[i] = 0.0;
1043  for (j = 0; j <= HB_FIR_SIZE; j++)
1044  out[i] += data[i + j] * fir_coef[j];
1045  }
1046 
1047  memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));
1048 }
1049 
1054 {
1055  memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE],
1056  (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float));
1057 
1058  memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float));
1059  memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float));
1060 
1061  memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE],
1062  LP_ORDER * sizeof(float));
1063  memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE],
1064  UPS_MEM_SIZE * sizeof(float));
1065  memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k],
1066  LP_ORDER_16k * sizeof(float));
1067 }
1068 
1069 static int amrwb_decode_frame(AVCodecContext *avctx, void *data,
1070  int *got_frame_ptr, AVPacket *avpkt)
1071 {
1072  AMRWBContext *ctx = avctx->priv_data;
1073  AMRWBFrame *cf = &ctx->frame;
1074  const uint8_t *buf = avpkt->data;
1075  int buf_size = avpkt->size;
1076  int expected_fr_size, header_size;
1077  float *buf_out;
1078  float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing
1079  float fixed_gain_factor; // fixed gain correction factor (gamma)
1080  float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
1081  float synth_fixed_gain; // the fixed gain that synthesis should use
1082  float voice_fac, stab_fac; // parameters used for gain smoothing
1083  float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis
1084  float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band
1085  float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis
1086  float hb_gain;
1087  int sub, i, ret;
1088 
1089  /* get output buffer */
1091  if ((ret = ff_get_buffer(avctx, &ctx->avframe)) < 0) {
1092  av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1093  return ret;
1094  }
1095  buf_out = (float *)ctx->avframe.data[0];
1096 
1097  header_size = decode_mime_header(ctx, buf);
1098  if (ctx->fr_cur_mode > MODE_SID) {
1099  av_log(avctx, AV_LOG_ERROR,
1100  "Invalid mode %d\n", ctx->fr_cur_mode);
1101  return AVERROR_INVALIDDATA;
1102  }
1103  expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1;
1104 
1105  if (buf_size < expected_fr_size) {
1106  av_log(avctx, AV_LOG_ERROR,
1107  "Frame too small (%d bytes). Truncated file?\n", buf_size);
1108  *got_frame_ptr = 0;
1109  return AVERROR_INVALIDDATA;
1110  }
1111 
1112  if (!ctx->fr_quality || ctx->fr_cur_mode > MODE_SID)
1113  av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n");
1114 
1115  if (ctx->fr_cur_mode == MODE_SID) { /* Comfort noise frame */
1116  av_log_missing_feature(avctx, "SID mode", 1);
1117  return -1;
1118  }
1119 
1120  ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame),
1121  buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]);
1122 
1123  /* Decode the quantized ISF vector */
1124  if (ctx->fr_cur_mode == MODE_6k60) {
1126  } else {
1128  }
1129 
1132 
1133  stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final);
1134 
1135  ctx->isf_cur[LP_ORDER - 1] *= 2.0;
1136  ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER);
1137 
1138  /* Generate a ISP vector for each subframe */
1139  if (ctx->first_frame) {
1140  ctx->first_frame = 0;
1141  memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double));
1142  }
1143  interpolate_isp(ctx->isp, ctx->isp_sub4_past);
1144 
1145  for (sub = 0; sub < 4; sub++)
1146  ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER);
1147 
1148  for (sub = 0; sub < 4; sub++) {
1149  const AMRWBSubFrame *cur_subframe = &cf->subframe[sub];
1150  float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k;
1151 
1152  /* Decode adaptive codebook (pitch vector) */
1153  decode_pitch_vector(ctx, cur_subframe, sub);
1154  /* Decode innovative codebook (fixed vector) */
1155  decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih,
1156  cur_subframe->pul_il, ctx->fr_cur_mode);
1157 
1158  pitch_sharpening(ctx, ctx->fixed_vector);
1159 
1160  decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode,
1161  &fixed_gain_factor, &ctx->pitch_gain[0]);
1162 
1163  ctx->fixed_gain[0] =
1164  ff_amr_set_fixed_gain(fixed_gain_factor,
1167  ctx->prediction_error,
1169 
1170  /* Calculate voice factor and store tilt for next subframe */
1171  voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],
1172  ctx->fixed_vector, ctx->fixed_gain[0]);
1173  ctx->tilt_coef = voice_fac * 0.25 + 0.25;
1174 
1175  /* Construct current excitation */
1176  for (i = 0; i < AMRWB_SFR_SIZE; i++) {
1177  ctx->excitation[i] *= ctx->pitch_gain[0];
1178  ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i];
1179  ctx->excitation[i] = truncf(ctx->excitation[i]);
1180  }
1181 
1182  /* Post-processing of excitation elements */
1183  synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain,
1184  voice_fac, stab_fac);
1185 
1186  synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector,
1187  spare_vector);
1188 
1189  pitch_enhancer(synth_fixed_vector, voice_fac);
1190 
1191  synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain,
1192  synth_fixed_vector, &ctx->samples_az[LP_ORDER]);
1193 
1194  /* Synthesis speech post-processing */
1196  &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);
1197 
1200  hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);
1201 
1202  upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],
1203  AMRWB_SFR_SIZE_16k);
1204 
1205  /* High frequency band (6.4 - 7.0 kHz) generation part */
1208  hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);
1209 
1210  hb_gain = find_hb_gain(ctx, hb_samples,
1211  cur_subframe->hb_gain, cf->vad);
1212 
1213  scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain);
1214 
1215  hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k],
1216  hb_exc, ctx->isf_cur, ctx->isf_past_final);
1217 
1218  /* High-band post-processing filters */
1219  hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem,
1220  &ctx->samples_hb[LP_ORDER_16k]);
1221 
1222  if (ctx->fr_cur_mode == MODE_23k85)
1223  hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem,
1224  hb_samples);
1225 
1226  /* Add the low and high frequency bands */
1227  for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
1228  sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15));
1229 
1230  /* Update buffers and history */
1231  update_sub_state(ctx);
1232  }
1233 
1234  /* update state for next frame */
1235  memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0]));
1236  memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float));
1237 
1238  *got_frame_ptr = 1;
1239  *(AVFrame *)data = ctx->avframe;
1240 
1241  return expected_fr_size;
1242 }
1243 
1245  .name = "amrwb",
1246  .type = AVMEDIA_TYPE_AUDIO,
1247  .id = CODEC_ID_AMR_WB,
1248  .priv_data_size = sizeof(AMRWBContext),
1251  .capabilities = CODEC_CAP_DR1,
1252  .long_name = NULL_IF_CONFIG_SMALL("Adaptive Multi-Rate WideBand"),
1253  .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_FLT,AV_SAMPLE_FMT_NONE},
1254 };