FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
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 FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A particular PURPOSE. See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21 
22 /**
23  * @file
24  * AMR wideband decoder
25  */
26 
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 "celp_math.h"
36 #include "acelp_filters.h"
37 #include "acelp_vectors.h"
38 #include "acelp_pitch_delay.h"
39 #include "internal.h"
40 
41 #define AMR_USE_16BIT_TABLES
42 #include "amr.h"
43 
44 #include "amrwbdata.h"
45 #include "mips/amrwbdec_mips.h"
46 
47 typedef struct {
48  AMRWBFrame frame; ///< AMRWB parameters decoded from bitstream
49  enum Mode fr_cur_mode; ///< mode index of current frame
50  uint8_t fr_quality; ///< frame quality index (FQI)
51  float isf_cur[LP_ORDER]; ///< working ISF vector from current frame
52  float isf_q_past[LP_ORDER]; ///< quantized ISF vector of the previous frame
53  float isf_past_final[LP_ORDER]; ///< final processed ISF vector of the previous frame
54  double isp[4][LP_ORDER]; ///< ISP vectors from current frame
55  double isp_sub4_past[LP_ORDER]; ///< ISP vector for the 4th subframe of the previous frame
56 
57  float lp_coef[4][LP_ORDER]; ///< Linear Prediction Coefficients from ISP vector
58 
59  uint8_t base_pitch_lag; ///< integer part of pitch lag for the next relative subframe
60  uint8_t pitch_lag_int; ///< integer part of pitch lag of the previous subframe
61 
62  float excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 2 + AMRWB_SFR_SIZE]; ///< current excitation and all necessary excitation history
63  float *excitation; ///< points to current excitation in excitation_buf[]
64 
65  float pitch_vector[AMRWB_SFR_SIZE]; ///< adaptive codebook (pitch) vector for current subframe
66  float fixed_vector[AMRWB_SFR_SIZE]; ///< algebraic codebook (fixed) vector for current subframe
67 
68  float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
69  float pitch_gain[6]; ///< quantified pitch gains for the current and previous five subframes
70  float fixed_gain[2]; ///< quantified fixed gains for the current and previous subframes
71 
72  float tilt_coef; ///< {beta_1} related to the voicing of the previous subframe
73 
74  float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness to determine "onset"
75  uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
76  float prev_tr_gain; ///< previous initial gain used by noise enhancer for threshold
77 
78  float samples_az[LP_ORDER + AMRWB_SFR_SIZE]; ///< low-band samples and memory from synthesis at 12.8kHz
79  float samples_up[UPS_MEM_SIZE + AMRWB_SFR_SIZE]; ///< low-band samples and memory processed for upsampling
80  float samples_hb[LP_ORDER_16k + AMRWB_SFR_SIZE_16k]; ///< high-band samples and memory from synthesis at 16kHz
81 
82  float hpf_31_mem[2], hpf_400_mem[2]; ///< previous values in the high pass filters
83  float demph_mem[1]; ///< previous value in the de-emphasis filter
84  float bpf_6_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band band pass filter
85  float lpf_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band low pass filter
86 
87  AVLFG prng; ///< random number generator for white noise excitation
88  uint8_t first_frame; ///< flag active during decoding of the first frame
89  ACELPFContext acelpf_ctx; ///< context for filters for ACELP-based codecs
90  ACELPVContext acelpv_ctx; ///< context for vector operations for ACELP-based codecs
91  CELPFContext celpf_ctx; ///< context for filters for CELP-based codecs
92  CELPMContext celpm_ctx; ///< context for fixed point math operations
93 
94 } AMRWBContext;
95 
97 {
98  AMRWBContext *ctx = avctx->priv_data;
99  int i;
100 
101  if (avctx->channels > 1) {
102  avpriv_report_missing_feature(avctx, "multi-channel AMR");
103  return AVERROR_PATCHWELCOME;
104  }
105 
106  avctx->channels = 1;
108  if (!avctx->sample_rate)
109  avctx->sample_rate = 16000;
110  avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
111 
112  av_lfg_init(&ctx->prng, 1);
113 
115  ctx->first_frame = 1;
116 
117  for (i = 0; i < LP_ORDER; i++)
118  ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15));
119 
120  for (i = 0; i < 4; i++)
121  ctx->prediction_error[i] = MIN_ENERGY;
122 
127 
128  return 0;
129 }
130 
131 /**
132  * Decode the frame header in the "MIME/storage" format. This format
133  * is simpler and does not carry the auxiliary frame information.
134  *
135  * @param[in] ctx The Context
136  * @param[in] buf Pointer to the input buffer
137  *
138  * @return The decoded header length in bytes
139  */
140 static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf)
141 {
142  /* Decode frame header (1st octet) */
143  ctx->fr_cur_mode = buf[0] >> 3 & 0x0F;
144  ctx->fr_quality = (buf[0] & 0x4) == 0x4;
145 
146  return 1;
147 }
148 
149 /**
150  * Decode quantized ISF vectors using 36-bit indexes (6K60 mode only).
151  *
152  * @param[in] ind Array of 5 indexes
153  * @param[out] isf_q Buffer for isf_q[LP_ORDER]
154  *
155  */
156 static void decode_isf_indices_36b(uint16_t *ind, float *isf_q)
157 {
158  int i;
159 
160  for (i = 0; i < 9; i++)
161  isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
162 
163  for (i = 0; i < 7; i++)
164  isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
165 
166  for (i = 0; i < 5; i++)
167  isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15));
168 
169  for (i = 0; i < 4; i++)
170  isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15));
171 
172  for (i = 0; i < 7; i++)
173  isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15));
174 }
175 
176 /**
177  * Decode quantized ISF vectors using 46-bit indexes (except 6K60 mode).
178  *
179  * @param[in] ind Array of 7 indexes
180  * @param[out] isf_q Buffer for isf_q[LP_ORDER]
181  *
182  */
183 static void decode_isf_indices_46b(uint16_t *ind, float *isf_q)
184 {
185  int i;
186 
187  for (i = 0; i < 9; i++)
188  isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
189 
190  for (i = 0; i < 7; i++)
191  isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
192 
193  for (i = 0; i < 3; i++)
194  isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15));
195 
196  for (i = 0; i < 3; i++)
197  isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15));
198 
199  for (i = 0; i < 3; i++)
200  isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15));
201 
202  for (i = 0; i < 3; i++)
203  isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15));
204 
205  for (i = 0; i < 4; i++)
206  isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15));
207 }
208 
209 /**
210  * Apply mean and past ISF values using the prediction factor.
211  * Updates past ISF vector.
212  *
213  * @param[in,out] isf_q Current quantized ISF
214  * @param[in,out] isf_past Past quantized ISF
215  *
216  */
217 static void isf_add_mean_and_past(float *isf_q, float *isf_past)
218 {
219  int i;
220  float tmp;
221 
222  for (i = 0; i < LP_ORDER; i++) {
223  tmp = isf_q[i];
224  isf_q[i] += isf_mean[i] * (1.0f / (1 << 15));
225  isf_q[i] += PRED_FACTOR * isf_past[i];
226  isf_past[i] = tmp;
227  }
228 }
229 
230 /**
231  * Interpolate the fourth ISP vector from current and past frames
232  * to obtain an ISP vector for each subframe.
233  *
234  * @param[in,out] isp_q ISPs for each subframe
235  * @param[in] isp4_past Past ISP for subframe 4
236  */
237 static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past)
238 {
239  int i, k;
240 
241  for (k = 0; k < 3; k++) {
242  float c = isfp_inter[k];
243  for (i = 0; i < LP_ORDER; i++)
244  isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i];
245  }
246 }
247 
248 /**
249  * Decode an adaptive codebook index into pitch lag (except 6k60, 8k85 modes).
250  * Calculate integer lag and fractional lag always using 1/4 resolution.
251  * In 1st and 3rd subframes the index is relative to last subframe integer lag.
252  *
253  * @param[out] lag_int Decoded integer pitch lag
254  * @param[out] lag_frac Decoded fractional pitch lag
255  * @param[in] pitch_index Adaptive codebook pitch index
256  * @param[in,out] base_lag_int Base integer lag used in relative subframes
257  * @param[in] subframe Current subframe index (0 to 3)
258  */
259 static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index,
260  uint8_t *base_lag_int, int subframe)
261 {
262  if (subframe == 0 || subframe == 2) {
263  if (pitch_index < 376) {
264  *lag_int = (pitch_index + 137) >> 2;
265  *lag_frac = pitch_index - (*lag_int << 2) + 136;
266  } else if (pitch_index < 440) {
267  *lag_int = (pitch_index + 257 - 376) >> 1;
268  *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) << 1;
269  /* the actual resolution is 1/2 but expressed as 1/4 */
270  } else {
271  *lag_int = pitch_index - 280;
272  *lag_frac = 0;
273  }
274  /* minimum lag for next subframe */
275  *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
277  // XXX: the spec states clearly that *base_lag_int should be
278  // the nearest integer to *lag_int (minus 8), but the ref code
279  // actually always uses its floor, I'm following the latter
280  } else {
281  *lag_int = (pitch_index + 1) >> 2;
282  *lag_frac = pitch_index - (*lag_int << 2);
283  *lag_int += *base_lag_int;
284  }
285 }
286 
287 /**
288  * Decode an adaptive codebook index into pitch lag for 8k85 and 6k60 modes.
289  * The description is analogous to decode_pitch_lag_high, but in 6k60 the
290  * relative index is used for all subframes except the first.
291  */
292 static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index,
293  uint8_t *base_lag_int, int subframe, enum Mode mode)
294 {
295  if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) {
296  if (pitch_index < 116) {
297  *lag_int = (pitch_index + 69) >> 1;
298  *lag_frac = (pitch_index - (*lag_int << 1) + 68) << 1;
299  } else {
300  *lag_int = pitch_index - 24;
301  *lag_frac = 0;
302  }
303  // XXX: same problem as before
304  *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
306  } else {
307  *lag_int = (pitch_index + 1) >> 1;
308  *lag_frac = (pitch_index - (*lag_int << 1)) << 1;
309  *lag_int += *base_lag_int;
310  }
311 }
312 
313 /**
314  * Find the pitch vector by interpolating the past excitation at the
315  * pitch delay, which is obtained in this function.
316  *
317  * @param[in,out] ctx The context
318  * @param[in] amr_subframe Current subframe data
319  * @param[in] subframe Current subframe index (0 to 3)
320  */
322  const AMRWBSubFrame *amr_subframe,
323  const int subframe)
324 {
325  int pitch_lag_int, pitch_lag_frac;
326  int i;
327  float *exc = ctx->excitation;
328  enum Mode mode = ctx->fr_cur_mode;
329 
330  if (mode <= MODE_8k85) {
331  decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
332  &ctx->base_pitch_lag, subframe, mode);
333  } else
334  decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
335  &ctx->base_pitch_lag, subframe);
336 
337  ctx->pitch_lag_int = pitch_lag_int;
338  pitch_lag_int += pitch_lag_frac > 0;
339 
340  /* Calculate the pitch vector by interpolating the past excitation at the
341  pitch lag using a hamming windowed sinc function */
343  exc + 1 - pitch_lag_int,
344  ac_inter, 4,
345  pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4),
346  LP_ORDER, AMRWB_SFR_SIZE + 1);
347 
348  /* Check which pitch signal path should be used
349  * 6k60 and 8k85 modes have the ltp flag set to 0 */
350  if (amr_subframe->ltp) {
351  memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float));
352  } else {
353  for (i = 0; i < AMRWB_SFR_SIZE; i++)
354  ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] +
355  0.18 * exc[i + 1];
356  memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float));
357  }
358 }
359 
360 /** Get x bits in the index interval [lsb,lsb+len-1] inclusive */
361 #define BIT_STR(x,lsb,len) (((x) >> (lsb)) & ((1 << (len)) - 1))
362 
363 /** Get the bit at specified position */
364 #define BIT_POS(x, p) (((x) >> (p)) & 1)
365 
366 /**
367  * The next six functions decode_[i]p_track decode exactly i pulses
368  * positions and amplitudes (-1 or 1) in a subframe track using
369  * an encoded pulse indexing (TS 26.190 section 5.8.2).
370  *
371  * The results are given in out[], in which a negative number means
372  * amplitude -1 and vice versa (i.e., ampl(x) = x / abs(x) ).
373  *
374  * @param[out] out Output buffer (writes i elements)
375  * @param[in] code Pulse index (no. of bits varies, see below)
376  * @param[in] m (log2) Number of potential positions
377  * @param[in] off Offset for decoded positions
378  */
379 static inline void decode_1p_track(int *out, int code, int m, int off)
380 {
381  int pos = BIT_STR(code, 0, m) + off; ///code: m+1 bits
382 
383  out[0] = BIT_POS(code, m) ? -pos : pos;
384 }
385 
386 static inline void decode_2p_track(int *out, int code, int m, int off) ///code: 2m+1 bits
387 {
388  int pos0 = BIT_STR(code, m, m) + off;
389  int pos1 = BIT_STR(code, 0, m) + off;
390 
391  out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0;
392  out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1;
393  out[1] = pos0 > pos1 ? -out[1] : out[1];
394 }
395 
396 static void decode_3p_track(int *out, int code, int m, int off) ///code: 3m+1 bits
397 {
398  int half_2p = BIT_POS(code, 2*m - 1) << (m - 1);
399 
400  decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
401  m - 1, off + half_2p);
402  decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off);
403 }
404 
405 static void decode_4p_track(int *out, int code, int m, int off) ///code: 4m bits
406 {
407  int half_4p, subhalf_2p;
408  int b_offset = 1 << (m - 1);
409 
410  switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */
411  case 0: /* 0 pulses in A, 4 pulses in B or vice versa */
412  half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses
413  subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2);
414 
415  decode_2p_track(out, BIT_STR(code, 0, 2*m - 3),
416  m - 2, off + half_4p + subhalf_2p);
417  decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1),
418  m - 1, off + half_4p);
419  break;
420  case 1: /* 1 pulse in A, 3 pulses in B */
421  decode_1p_track(out, BIT_STR(code, 3*m - 2, m),
422  m - 1, off);
423  decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2),
424  m - 1, off + b_offset);
425  break;
426  case 2: /* 2 pulses in each half */
427  decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1),
428  m - 1, off);
429  decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1),
430  m - 1, off + b_offset);
431  break;
432  case 3: /* 3 pulses in A, 1 pulse in B */
433  decode_3p_track(out, BIT_STR(code, m, 3*m - 2),
434  m - 1, off);
435  decode_1p_track(out + 3, BIT_STR(code, 0, m),
436  m - 1, off + b_offset);
437  break;
438  }
439 }
440 
441 static void decode_5p_track(int *out, int code, int m, int off) ///code: 5m bits
442 {
443  int half_3p = BIT_POS(code, 5*m - 1) << (m - 1);
444 
445  decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2),
446  m - 1, off + half_3p);
447 
448  decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off);
449 }
450 
451 static void decode_6p_track(int *out, int code, int m, int off) ///code: 6m-2 bits
452 {
453  int b_offset = 1 << (m - 1);
454  /* which half has more pulses in cases 0 to 2 */
455  int half_more = BIT_POS(code, 6*m - 5) << (m - 1);
456  int half_other = b_offset - half_more;
457 
458  switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */
459  case 0: /* 0 pulses in A, 6 pulses in B or vice versa */
460  decode_1p_track(out, BIT_STR(code, 0, m),
461  m - 1, off + half_more);
462  decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
463  m - 1, off + half_more);
464  break;
465  case 1: /* 1 pulse in A, 5 pulses in B or vice versa */
466  decode_1p_track(out, BIT_STR(code, 0, m),
467  m - 1, off + half_other);
468  decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
469  m - 1, off + half_more);
470  break;
471  case 2: /* 2 pulses in A, 4 pulses in B or vice versa */
472  decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
473  m - 1, off + half_other);
474  decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4),
475  m - 1, off + half_more);
476  break;
477  case 3: /* 3 pulses in A, 3 pulses in B */
478  decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2),
479  m - 1, off);
480  decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2),
481  m - 1, off + b_offset);
482  break;
483  }
484 }
485 
486 /**
487  * Decode the algebraic codebook index to pulse positions and signs,
488  * then construct the algebraic codebook vector.
489  *
490  * @param[out] fixed_vector Buffer for the fixed codebook excitation
491  * @param[in] pulse_hi MSBs part of the pulse index array (higher modes only)
492  * @param[in] pulse_lo LSBs part of the pulse index array
493  * @param[in] mode Mode of the current frame
494  */
495 static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi,
496  const uint16_t *pulse_lo, const enum Mode mode)
497 {
498  /* sig_pos stores for each track the decoded pulse position indexes
499  * (1-based) multiplied by its corresponding amplitude (+1 or -1) */
500  int sig_pos[4][6];
501  int spacing = (mode == MODE_6k60) ? 2 : 4;
502  int i, j;
503 
504  switch (mode) {
505  case MODE_6k60:
506  for (i = 0; i < 2; i++)
507  decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1);
508  break;
509  case MODE_8k85:
510  for (i = 0; i < 4; i++)
511  decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1);
512  break;
513  case MODE_12k65:
514  for (i = 0; i < 4; i++)
515  decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
516  break;
517  case MODE_14k25:
518  for (i = 0; i < 2; i++)
519  decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
520  for (i = 2; i < 4; i++)
521  decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
522  break;
523  case MODE_15k85:
524  for (i = 0; i < 4; i++)
525  decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
526  break;
527  case MODE_18k25:
528  for (i = 0; i < 4; i++)
529  decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
530  ((int) pulse_hi[i] << 14), 4, 1);
531  break;
532  case MODE_19k85:
533  for (i = 0; i < 2; i++)
534  decode_5p_track(sig_pos[i], (int) pulse_lo[i] +
535  ((int) pulse_hi[i] << 10), 4, 1);
536  for (i = 2; i < 4; i++)
537  decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
538  ((int) pulse_hi[i] << 14), 4, 1);
539  break;
540  case MODE_23k05:
541  case MODE_23k85:
542  for (i = 0; i < 4; i++)
543  decode_6p_track(sig_pos[i], (int) pulse_lo[i] +
544  ((int) pulse_hi[i] << 11), 4, 1);
545  break;
546  }
547 
548  memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE);
549 
550  for (i = 0; i < 4; i++)
551  for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) {
552  int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i;
553 
554  fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0;
555  }
556 }
557 
558 /**
559  * Decode pitch gain and fixed gain correction factor.
560  *
561  * @param[in] vq_gain Vector-quantized index for gains
562  * @param[in] mode Mode of the current frame
563  * @param[out] fixed_gain_factor Decoded fixed gain correction factor
564  * @param[out] pitch_gain Decoded pitch gain
565  */
566 static void decode_gains(const uint8_t vq_gain, const enum Mode mode,
567  float *fixed_gain_factor, float *pitch_gain)
568 {
569  const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] :
570  qua_gain_7b[vq_gain]);
571 
572  *pitch_gain = gains[0] * (1.0f / (1 << 14));
573  *fixed_gain_factor = gains[1] * (1.0f / (1 << 11));
574 }
575 
576 /**
577  * Apply pitch sharpening filters to the fixed codebook vector.
578  *
579  * @param[in] ctx The context
580  * @param[in,out] fixed_vector Fixed codebook excitation
581  */
582 // XXX: Spec states this procedure should be applied when the pitch
583 // lag is less than 64, but this checking seems absent in reference and AMR-NB
584 static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector)
585 {
586  int i;
587 
588  /* Tilt part */
589  for (i = AMRWB_SFR_SIZE - 1; i != 0; i--)
590  fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef;
591 
592  /* Periodicity enhancement part */
593  for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++)
594  fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85;
595 }
596 
597 /**
598  * Calculate the voicing factor (-1.0 = unvoiced to 1.0 = voiced).
599  *
600  * @param[in] p_vector, f_vector Pitch and fixed excitation vectors
601  * @param[in] p_gain, f_gain Pitch and fixed gains
602  * @param[in] ctx The context
603  */
604 // XXX: There is something wrong with the precision here! The magnitudes
605 // of the energies are not correct. Please check the reference code carefully
606 static float voice_factor(float *p_vector, float p_gain,
607  float *f_vector, float f_gain,
608  CELPMContext *ctx)
609 {
610  double p_ener = (double) ctx->dot_productf(p_vector, p_vector,
611  AMRWB_SFR_SIZE) *
612  p_gain * p_gain;
613  double f_ener = (double) ctx->dot_productf(f_vector, f_vector,
614  AMRWB_SFR_SIZE) *
615  f_gain * f_gain;
616 
617  return (p_ener - f_ener) / (p_ener + f_ener);
618 }
619 
620 /**
621  * Reduce fixed vector sparseness by smoothing with one of three IR filters,
622  * also known as "adaptive phase dispersion".
623  *
624  * @param[in] ctx The context
625  * @param[in,out] fixed_vector Unfiltered fixed vector
626  * @param[out] buf Space for modified vector if necessary
627  *
628  * @return The potentially overwritten filtered fixed vector address
629  */
630 static float *anti_sparseness(AMRWBContext *ctx,
631  float *fixed_vector, float *buf)
632 {
633  int ir_filter_nr;
634 
635  if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes
636  return fixed_vector;
637 
638  if (ctx->pitch_gain[0] < 0.6) {
639  ir_filter_nr = 0; // strong filtering
640  } else if (ctx->pitch_gain[0] < 0.9) {
641  ir_filter_nr = 1; // medium filtering
642  } else
643  ir_filter_nr = 2; // no filtering
644 
645  /* detect 'onset' */
646  if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) {
647  if (ir_filter_nr < 2)
648  ir_filter_nr++;
649  } else {
650  int i, count = 0;
651 
652  for (i = 0; i < 6; i++)
653  if (ctx->pitch_gain[i] < 0.6)
654  count++;
655 
656  if (count > 2)
657  ir_filter_nr = 0;
658 
659  if (ir_filter_nr > ctx->prev_ir_filter_nr + 1)
660  ir_filter_nr--;
661  }
662 
663  /* update ir filter strength history */
664  ctx->prev_ir_filter_nr = ir_filter_nr;
665 
666  ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85);
667 
668  if (ir_filter_nr < 2) {
669  int i;
670  const float *coef = ir_filters_lookup[ir_filter_nr];
671 
672  /* Circular convolution code in the reference
673  * decoder was modified to avoid using one
674  * extra array. The filtered vector is given by:
675  *
676  * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) }
677  */
678 
679  memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE);
680  for (i = 0; i < AMRWB_SFR_SIZE; i++)
681  if (fixed_vector[i])
682  ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i],
683  AMRWB_SFR_SIZE);
684  fixed_vector = buf;
685  }
686 
687  return fixed_vector;
688 }
689 
690 /**
691  * Calculate a stability factor {teta} based on distance between
692  * current and past isf. A value of 1 shows maximum signal stability.
693  */
694 static float stability_factor(const float *isf, const float *isf_past)
695 {
696  int i;
697  float acc = 0.0;
698 
699  for (i = 0; i < LP_ORDER - 1; i++)
700  acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]);
701 
702  // XXX: This part is not so clear from the reference code
703  // the result is more accurate changing the "/ 256" to "* 512"
704  return FFMAX(0.0, 1.25 - acc * 0.8 * 512);
705 }
706 
707 /**
708  * Apply a non-linear fixed gain smoothing in order to reduce
709  * fluctuation in the energy of excitation.
710  *
711  * @param[in] fixed_gain Unsmoothed fixed gain
712  * @param[in,out] prev_tr_gain Previous threshold gain (updated)
713  * @param[in] voice_fac Frame voicing factor
714  * @param[in] stab_fac Frame stability factor
715  *
716  * @return The smoothed gain
717  */
718 static float noise_enhancer(float fixed_gain, float *prev_tr_gain,
719  float voice_fac, float stab_fac)
720 {
721  float sm_fac = 0.5 * (1 - voice_fac) * stab_fac;
722  float g0;
723 
724  // XXX: the following fixed-point constants used to in(de)crement
725  // gain by 1.5dB were taken from the reference code, maybe it could
726  // be simpler
727  if (fixed_gain < *prev_tr_gain) {
728  g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain *
729  (6226 * (1.0f / (1 << 15)))); // +1.5 dB
730  } else
731  g0 = FFMAX(*prev_tr_gain, fixed_gain *
732  (27536 * (1.0f / (1 << 15)))); // -1.5 dB
733 
734  *prev_tr_gain = g0; // update next frame threshold
735 
736  return sm_fac * g0 + (1 - sm_fac) * fixed_gain;
737 }
738 
739 /**
740  * Filter the fixed_vector to emphasize the higher frequencies.
741  *
742  * @param[in,out] fixed_vector Fixed codebook vector
743  * @param[in] voice_fac Frame voicing factor
744  */
745 static void pitch_enhancer(float *fixed_vector, float voice_fac)
746 {
747  int i;
748  float cpe = 0.125 * (1 + voice_fac);
749  float last = fixed_vector[0]; // holds c(i - 1)
750 
751  fixed_vector[0] -= cpe * fixed_vector[1];
752 
753  for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) {
754  float cur = fixed_vector[i];
755 
756  fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]);
757  last = cur;
758  }
759 
760  fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last;
761 }
762 
763 /**
764  * Conduct 16th order linear predictive coding synthesis from excitation.
765  *
766  * @param[in] ctx Pointer to the AMRWBContext
767  * @param[in] lpc Pointer to the LPC coefficients
768  * @param[out] excitation Buffer for synthesis final excitation
769  * @param[in] fixed_gain Fixed codebook gain for synthesis
770  * @param[in] fixed_vector Algebraic codebook vector
771  * @param[in,out] samples Pointer to the output samples and memory
772  */
773 static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
774  float fixed_gain, const float *fixed_vector,
775  float *samples)
776 {
777  ctx->acelpv_ctx.weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
778  ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);
779 
780  /* emphasize pitch vector contribution in low bitrate modes */
781  if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {
782  int i;
783  float energy = ctx->celpm_ctx.dot_productf(excitation, excitation,
785 
786  // XXX: Weird part in both ref code and spec. A unknown parameter
787  // {beta} seems to be identical to the current pitch gain
788  float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0];
789 
790  for (i = 0; i < AMRWB_SFR_SIZE; i++)
791  excitation[i] += pitch_factor * ctx->pitch_vector[i];
792 
793  ff_scale_vector_to_given_sum_of_squares(excitation, excitation,
794  energy, AMRWB_SFR_SIZE);
795  }
796 
797  ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation,
799 }
800 
801 /**
802  * Apply to synthesis a de-emphasis filter of the form:
803  * H(z) = 1 / (1 - m * z^-1)
804  *
805  * @param[out] out Output buffer
806  * @param[in] in Input samples array with in[-1]
807  * @param[in] m Filter coefficient
808  * @param[in,out] mem State from last filtering
809  */
810 static void de_emphasis(float *out, float *in, float m, float mem[1])
811 {
812  int i;
813 
814  out[0] = in[0] + m * mem[0];
815 
816  for (i = 1; i < AMRWB_SFR_SIZE; i++)
817  out[i] = in[i] + out[i - 1] * m;
818 
819  mem[0] = out[AMRWB_SFR_SIZE - 1];
820 }
821 
822 /**
823  * Upsample a signal by 5/4 ratio (from 12.8kHz to 16kHz) using
824  * a FIR interpolation filter. Uses past data from before *in address.
825  *
826  * @param[out] out Buffer for interpolated signal
827  * @param[in] in Current signal data (length 0.8*o_size)
828  * @param[in] o_size Output signal length
829  * @param[in] ctx The context
830  */
831 static void upsample_5_4(float *out, const float *in, int o_size, CELPMContext *ctx)
832 {
833  const float *in0 = in - UPS_FIR_SIZE + 1;
834  int i, j, k;
835  int int_part = 0, frac_part;
836 
837  i = 0;
838  for (j = 0; j < o_size / 5; j++) {
839  out[i] = in[int_part];
840  frac_part = 4;
841  i++;
842 
843  for (k = 1; k < 5; k++) {
844  out[i] = ctx->dot_productf(in0 + int_part,
845  upsample_fir[4 - frac_part],
846  UPS_MEM_SIZE);
847  int_part++;
848  frac_part--;
849  i++;
850  }
851  }
852 }
853 
854 /**
855  * Calculate the high-band gain based on encoded index (23k85 mode) or
856  * on the low-band speech signal and the Voice Activity Detection flag.
857  *
858  * @param[in] ctx The context
859  * @param[in] synth LB speech synthesis at 12.8k
860  * @param[in] hb_idx Gain index for mode 23k85 only
861  * @param[in] vad VAD flag for the frame
862  */
863 static float find_hb_gain(AMRWBContext *ctx, const float *synth,
864  uint16_t hb_idx, uint8_t vad)
865 {
866  int wsp = (vad > 0);
867  float tilt;
868 
869  if (ctx->fr_cur_mode == MODE_23k85)
870  return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));
871 
872  tilt = ctx->celpm_ctx.dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1) /
873  ctx->celpm_ctx.dot_productf(synth, synth, AMRWB_SFR_SIZE);
874 
875  /* return gain bounded by [0.1, 1.0] */
876  return av_clipf((1.0 - FFMAX(0.0, tilt)) * (1.25 - 0.25 * wsp), 0.1, 1.0);
877 }
878 
879 /**
880  * Generate the high-band excitation with the same energy from the lower
881  * one and scaled by the given gain.
882  *
883  * @param[in] ctx The context
884  * @param[out] hb_exc Buffer for the excitation
885  * @param[in] synth_exc Low-band excitation used for synthesis
886  * @param[in] hb_gain Wanted excitation gain
887  */
888 static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,
889  const float *synth_exc, float hb_gain)
890 {
891  int i;
892  float energy = ctx->celpm_ctx.dot_productf(synth_exc, synth_exc,
894 
895  /* Generate a white-noise excitation */
896  for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
897  hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng);
898 
900  energy * hb_gain * hb_gain,
901  AMRWB_SFR_SIZE_16k);
902 }
903 
904 /**
905  * Calculate the auto-correlation for the ISF difference vector.
906  */
907 static float auto_correlation(float *diff_isf, float mean, int lag)
908 {
909  int i;
910  float sum = 0.0;
911 
912  for (i = 7; i < LP_ORDER - 2; i++) {
913  float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean);
914  sum += prod * prod;
915  }
916  return sum;
917 }
918 
919 /**
920  * Extrapolate a ISF vector to the 16kHz range (20th order LP)
921  * used at mode 6k60 LP filter for the high frequency band.
922  *
923  * @param[out] isf Buffer for extrapolated isf; contains LP_ORDER
924  * values on input
925  */
926 static void extrapolate_isf(float isf[LP_ORDER_16k])
927 {
928  float diff_isf[LP_ORDER - 2], diff_mean;
929  float corr_lag[3];
930  float est, scale;
931  int i, j, i_max_corr;
932 
933  isf[LP_ORDER_16k - 1] = isf[LP_ORDER - 1];
934 
935  /* Calculate the difference vector */
936  for (i = 0; i < LP_ORDER - 2; i++)
937  diff_isf[i] = isf[i + 1] - isf[i];
938 
939  diff_mean = 0.0;
940  for (i = 2; i < LP_ORDER - 2; i++)
941  diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4));
942 
943  /* Find which is the maximum autocorrelation */
944  i_max_corr = 0;
945  for (i = 0; i < 3; i++) {
946  corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2);
947 
948  if (corr_lag[i] > corr_lag[i_max_corr])
949  i_max_corr = i;
950  }
951  i_max_corr++;
952 
953  for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
954  isf[i] = isf[i - 1] + isf[i - 1 - i_max_corr]
955  - isf[i - 2 - i_max_corr];
956 
957  /* Calculate an estimate for ISF(18) and scale ISF based on the error */
958  est = 7965 + (isf[2] - isf[3] - isf[4]) / 6.0;
959  scale = 0.5 * (FFMIN(est, 7600) - isf[LP_ORDER - 2]) /
960  (isf[LP_ORDER_16k - 2] - isf[LP_ORDER - 2]);
961 
962  for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
963  diff_isf[j] = scale * (isf[i] - isf[i - 1]);
964 
965  /* Stability insurance */
966  for (i = 1; i < LP_ORDER_16k - LP_ORDER; i++)
967  if (diff_isf[i] + diff_isf[i - 1] < 5.0) {
968  if (diff_isf[i] > diff_isf[i - 1]) {
969  diff_isf[i - 1] = 5.0 - diff_isf[i];
970  } else
971  diff_isf[i] = 5.0 - diff_isf[i - 1];
972  }
973 
974  for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
975  isf[i] = isf[i - 1] + diff_isf[j] * (1.0f / (1 << 15));
976 
977  /* Scale the ISF vector for 16000 Hz */
978  for (i = 0; i < LP_ORDER_16k - 1; i++)
979  isf[i] *= 0.8;
980 }
981 
982 /**
983  * Spectral expand the LP coefficients using the equation:
984  * y[i] = x[i] * (gamma ** i)
985  *
986  * @param[out] out Output buffer (may use input array)
987  * @param[in] lpc LP coefficients array
988  * @param[in] gamma Weighting factor
989  * @param[in] size LP array size
990  */
991 static void lpc_weighting(float *out, const float *lpc, float gamma, int size)
992 {
993  int i;
994  float fac = gamma;
995 
996  for (i = 0; i < size; i++) {
997  out[i] = lpc[i] * fac;
998  fac *= gamma;
999  }
1000 }
1001 
1002 /**
1003  * Conduct 20th order linear predictive coding synthesis for the high
1004  * frequency band excitation at 16kHz.
1005  *
1006  * @param[in] ctx The context
1007  * @param[in] subframe Current subframe index (0 to 3)
1008  * @param[in,out] samples Pointer to the output speech samples
1009  * @param[in] exc Generated white-noise scaled excitation
1010  * @param[in] isf Current frame isf vector
1011  * @param[in] isf_past Past frame final isf vector
1012  */
1013 static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
1014  const float *exc, const float *isf, const float *isf_past)
1015 {
1016  float hb_lpc[LP_ORDER_16k];
1017  enum Mode mode = ctx->fr_cur_mode;
1018 
1019  if (mode == MODE_6k60) {
1020  float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation
1021  double e_isp[LP_ORDER_16k];
1022 
1023  ctx->acelpv_ctx.weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
1024  1.0 - isfp_inter[subframe], LP_ORDER);
1025 
1026  extrapolate_isf(e_isf);
1027 
1028  e_isf[LP_ORDER_16k - 1] *= 2.0;
1029  ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k);
1030  ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k);
1031 
1032  lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k);
1033  } else {
1034  lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);
1035  }
1036 
1037  ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
1038  (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);
1039 }
1040 
1041 /**
1042  * Apply a 15th order filter to high-band samples.
1043  * The filter characteristic depends on the given coefficients.
1044  *
1045  * @param[out] out Buffer for filtered output
1046  * @param[in] fir_coef Filter coefficients
1047  * @param[in,out] mem State from last filtering (updated)
1048  * @param[in] in Input speech data (high-band)
1049  *
1050  * @remark It is safe to pass the same array in in and out parameters
1051  */
1052 
1053 #ifndef hb_fir_filter
1054 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
1055  float mem[HB_FIR_SIZE], const float *in)
1056 {
1057  int i, j;
1058  float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples
1059 
1060  memcpy(data, mem, HB_FIR_SIZE * sizeof(float));
1061  memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float));
1062 
1063  for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) {
1064  out[i] = 0.0;
1065  for (j = 0; j <= HB_FIR_SIZE; j++)
1066  out[i] += data[i + j] * fir_coef[j];
1067  }
1068 
1069  memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));
1070 }
1071 #endif /* hb_fir_filter */
1072 
1073 /**
1074  * Update context state before the next subframe.
1075  */
1077 {
1078  memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE],
1079  (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float));
1080 
1081  memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float));
1082  memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float));
1083 
1084  memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE],
1085  LP_ORDER * sizeof(float));
1086  memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE],
1087  UPS_MEM_SIZE * sizeof(float));
1088  memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k],
1089  LP_ORDER_16k * sizeof(float));
1090 }
1091 
1092 static int amrwb_decode_frame(AVCodecContext *avctx, void *data,
1093  int *got_frame_ptr, AVPacket *avpkt)
1094 {
1095  AMRWBContext *ctx = avctx->priv_data;
1096  AVFrame *frame = data;
1097  AMRWBFrame *cf = &ctx->frame;
1098  const uint8_t *buf = avpkt->data;
1099  int buf_size = avpkt->size;
1100  int expected_fr_size, header_size;
1101  float *buf_out;
1102  float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing
1103  float fixed_gain_factor; // fixed gain correction factor (gamma)
1104  float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
1105  float synth_fixed_gain; // the fixed gain that synthesis should use
1106  float voice_fac, stab_fac; // parameters used for gain smoothing
1107  float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis
1108  float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band
1109  float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis
1110  float hb_gain;
1111  int sub, i, ret;
1112 
1113  /* get output buffer */
1114  frame->nb_samples = 4 * AMRWB_SFR_SIZE_16k;
1115  if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
1116  return ret;
1117  buf_out = (float *)frame->data[0];
1118 
1119  header_size = decode_mime_header(ctx, buf);
1120  if (ctx->fr_cur_mode > MODE_SID) {
1121  av_log(avctx, AV_LOG_ERROR,
1122  "Invalid mode %d\n", ctx->fr_cur_mode);
1123  return AVERROR_INVALIDDATA;
1124  }
1125  expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1;
1126 
1127  if (buf_size < expected_fr_size) {
1128  av_log(avctx, AV_LOG_ERROR,
1129  "Frame too small (%d bytes). Truncated file?\n", buf_size);
1130  *got_frame_ptr = 0;
1131  return AVERROR_INVALIDDATA;
1132  }
1133 
1134  if (!ctx->fr_quality || ctx->fr_cur_mode > MODE_SID)
1135  av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n");
1136 
1137  if (ctx->fr_cur_mode == MODE_SID) { /* Comfort noise frame */
1138  avpriv_request_sample(avctx, "SID mode");
1139  return AVERROR_PATCHWELCOME;
1140  }
1141 
1142  ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame),
1143  buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]);
1144 
1145  /* Decode the quantized ISF vector */
1146  if (ctx->fr_cur_mode == MODE_6k60) {
1148  } else {
1150  }
1151 
1154 
1155  stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final);
1156 
1157  ctx->isf_cur[LP_ORDER - 1] *= 2.0;
1158  ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER);
1159 
1160  /* Generate a ISP vector for each subframe */
1161  if (ctx->first_frame) {
1162  ctx->first_frame = 0;
1163  memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double));
1164  }
1165  interpolate_isp(ctx->isp, ctx->isp_sub4_past);
1166 
1167  for (sub = 0; sub < 4; sub++)
1168  ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER);
1169 
1170  for (sub = 0; sub < 4; sub++) {
1171  const AMRWBSubFrame *cur_subframe = &cf->subframe[sub];
1172  float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k;
1173 
1174  /* Decode adaptive codebook (pitch vector) */
1175  decode_pitch_vector(ctx, cur_subframe, sub);
1176  /* Decode innovative codebook (fixed vector) */
1177  decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih,
1178  cur_subframe->pul_il, ctx->fr_cur_mode);
1179 
1180  pitch_sharpening(ctx, ctx->fixed_vector);
1181 
1182  decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode,
1183  &fixed_gain_factor, &ctx->pitch_gain[0]);
1184 
1185  ctx->fixed_gain[0] =
1186  ff_amr_set_fixed_gain(fixed_gain_factor,
1188  ctx->fixed_vector,
1189  AMRWB_SFR_SIZE) /
1191  ctx->prediction_error,
1193 
1194  /* Calculate voice factor and store tilt for next subframe */
1195  voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],
1196  ctx->fixed_vector, ctx->fixed_gain[0],
1197  &ctx->celpm_ctx);
1198  ctx->tilt_coef = voice_fac * 0.25 + 0.25;
1199 
1200  /* Construct current excitation */
1201  for (i = 0; i < AMRWB_SFR_SIZE; i++) {
1202  ctx->excitation[i] *= ctx->pitch_gain[0];
1203  ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i];
1204  ctx->excitation[i] = truncf(ctx->excitation[i]);
1205  }
1206 
1207  /* Post-processing of excitation elements */
1208  synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain,
1209  voice_fac, stab_fac);
1210 
1211  synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector,
1212  spare_vector);
1213 
1214  pitch_enhancer(synth_fixed_vector, voice_fac);
1215 
1216  synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain,
1217  synth_fixed_vector, &ctx->samples_az[LP_ORDER]);
1218 
1219  /* Synthesis speech post-processing */
1221  &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);
1222 
1225  hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);
1226 
1227  upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],
1228  AMRWB_SFR_SIZE_16k, &ctx->celpm_ctx);
1229 
1230  /* High frequency band (6.4 - 7.0 kHz) generation part */
1233  hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);
1234 
1235  hb_gain = find_hb_gain(ctx, hb_samples,
1236  cur_subframe->hb_gain, cf->vad);
1237 
1238  scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain);
1239 
1240  hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k],
1241  hb_exc, ctx->isf_cur, ctx->isf_past_final);
1242 
1243  /* High-band post-processing filters */
1244  hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem,
1245  &ctx->samples_hb[LP_ORDER_16k]);
1246 
1247  if (ctx->fr_cur_mode == MODE_23k85)
1248  hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem,
1249  hb_samples);
1250 
1251  /* Add the low and high frequency bands */
1252  for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
1253  sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15));
1254 
1255  /* Update buffers and history */
1256  update_sub_state(ctx);
1257  }
1258 
1259  /* update state for next frame */
1260  memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0]));
1261  memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float));
1262 
1263  *got_frame_ptr = 1;
1264 
1265  return expected_fr_size;
1266 }
1267 
1269  .name = "amrwb",
1270  .long_name = NULL_IF_CONFIG_SMALL("AMR-WB (Adaptive Multi-Rate WideBand)"),
1271  .type = AVMEDIA_TYPE_AUDIO,
1272  .id = AV_CODEC_ID_AMR_WB,
1273  .priv_data_size = sizeof(AMRWBContext),
1276  .capabilities = CODEC_CAP_DR1,
1277  .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,
1279 };