FFmpeg  4.4.6
opus_silk.c
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2012 Andrew D'Addesio
3  * Copyright (c) 2013-2014 Mozilla Corporation
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  * Opus SILK decoder
25  */
26 
27 #include <stdint.h>
28 
29 #include "opus.h"
30 #include "opustab.h"
31 
32 typedef struct SilkFrame {
33  int coded;
34  int log_gain;
35  int16_t nlsf[16];
36  float lpc[16];
37 
38  float output [2 * SILK_HISTORY];
41 
43 } SilkFrame;
44 
45 struct SilkContext {
48 
49  int midonly;
50  int subframes;
51  int sflength;
52  int flength;
54 
56  int wb;
57 
60  float stereo_weights[2];
61 
63 };
64 
65 static inline void silk_stabilize_lsf(int16_t nlsf[16], int order, const uint16_t min_delta[17])
66 {
67  int pass, i;
68  for (pass = 0; pass < 20; pass++) {
69  int k, min_diff = 0;
70  for (i = 0; i < order+1; i++) {
71  int low = i != 0 ? nlsf[i-1] : 0;
72  int high = i != order ? nlsf[i] : 32768;
73  int diff = (high - low) - (min_delta[i]);
74 
75  if (diff < min_diff) {
76  min_diff = diff;
77  k = i;
78 
79  if (pass == 20)
80  break;
81  }
82  }
83  if (min_diff == 0) /* no issues; stabilized */
84  return;
85 
86  /* wiggle one or two LSFs */
87  if (k == 0) {
88  /* repel away from lower bound */
89  nlsf[0] = min_delta[0];
90  } else if (k == order) {
91  /* repel away from higher bound */
92  nlsf[order-1] = 32768 - min_delta[order];
93  } else {
94  /* repel away from current position */
95  int min_center = 0, max_center = 32768, center_val;
96 
97  /* lower extent */
98  for (i = 0; i < k; i++)
99  min_center += min_delta[i];
100  min_center += min_delta[k] >> 1;
101 
102  /* upper extent */
103  for (i = order; i > k; i--)
104  max_center -= min_delta[i];
105  max_center -= min_delta[k] >> 1;
106 
107  /* move apart */
108  center_val = nlsf[k - 1] + nlsf[k];
109  center_val = (center_val >> 1) + (center_val & 1); // rounded divide by 2
110  center_val = FFMIN(max_center, FFMAX(min_center, center_val));
111 
112  nlsf[k - 1] = center_val - (min_delta[k] >> 1);
113  nlsf[k] = nlsf[k - 1] + min_delta[k];
114  }
115  }
116 
117  /* resort to the fall-back method, the standard method for LSF stabilization */
118 
119  /* sort; as the LSFs should be nearly sorted, use insertion sort */
120  for (i = 1; i < order; i++) {
121  int j, value = nlsf[i];
122  for (j = i - 1; j >= 0 && nlsf[j] > value; j--)
123  nlsf[j + 1] = nlsf[j];
124  nlsf[j + 1] = value;
125  }
126 
127  /* push forwards to increase distance */
128  if (nlsf[0] < min_delta[0])
129  nlsf[0] = min_delta[0];
130  for (i = 1; i < order; i++)
131  nlsf[i] = FFMAX(nlsf[i], FFMIN(nlsf[i - 1] + min_delta[i], 32767));
132 
133  /* push backwards to increase distance */
134  if (nlsf[order-1] > 32768 - min_delta[order])
135  nlsf[order-1] = 32768 - min_delta[order];
136  for (i = order-2; i >= 0; i--)
137  if (nlsf[i] > nlsf[i + 1] - min_delta[i+1])
138  nlsf[i] = nlsf[i + 1] - min_delta[i+1];
139 
140  return;
141 }
142 
143 static inline int silk_is_lpc_stable(const int16_t lpc[16], int order)
144 {
145  int k, j, DC_resp = 0;
146  int32_t lpc32[2][16]; // Q24
147  int totalinvgain = 1 << 30; // 1.0 in Q30
148  int32_t *row = lpc32[0], *prevrow;
149 
150  /* initialize the first row for the Levinson recursion */
151  for (k = 0; k < order; k++) {
152  DC_resp += lpc[k];
153  row[k] = lpc[k] * 4096;
154  }
155 
156  if (DC_resp >= 4096)
157  return 0;
158 
159  /* check if prediction gain pushes any coefficients too far */
160  for (k = order - 1; 1; k--) {
161  int rc; // Q31; reflection coefficient
162  int gaindiv; // Q30; inverse of the gain (the divisor)
163  int gain; // gain for this reflection coefficient
164  int fbits; // fractional bits used for the gain
165  int error; // Q29; estimate of the error of our partial estimate of 1/gaindiv
166 
167  if (FFABS(row[k]) > 16773022)
168  return 0;
169 
170  rc = -(row[k] * 128);
171  gaindiv = (1 << 30) - MULH(rc, rc);
172 
173  totalinvgain = MULH(totalinvgain, gaindiv) << 2;
174  if (k == 0)
175  return (totalinvgain >= 107374);
176 
177  /* approximate 1.0/gaindiv */
178  fbits = opus_ilog(gaindiv);
179  gain = ((1 << 29) - 1) / (gaindiv >> (fbits + 1 - 16)); // Q<fbits-16>
180  error = (1 << 29) - MULL(gaindiv << (15 + 16 - fbits), gain, 16);
181  gain = ((gain << 16) + (error * gain >> 13));
182 
183  /* switch to the next row of the LPC coefficients */
184  prevrow = row;
185  row = lpc32[k & 1];
186 
187  for (j = 0; j < k; j++) {
188  int x = av_sat_sub32(prevrow[j], ROUND_MULL(prevrow[k - j - 1], rc, 31));
189  int64_t tmp = ROUND_MULL(x, gain, fbits);
190 
191  /* per RFC 8251 section 6, if this calculation overflows, the filter
192  is considered unstable. */
193  if (tmp < INT32_MIN || tmp > INT32_MAX)
194  return 0;
195 
196  row[j] = (int32_t)tmp;
197  }
198  }
199 }
200 
201 static void silk_lsp2poly(const int32_t lsp[/* 2 * half_order - 1 */],
202  int32_t pol[/* half_order + 1 */], int half_order)
203 {
204  int i, j;
205 
206  pol[0] = 65536; // 1.0 in Q16
207  pol[1] = -lsp[0];
208 
209  for (i = 1; i < half_order; i++) {
210  pol[i + 1] = pol[i - 1] * 2 - ROUND_MULL(lsp[2 * i], pol[i], 16);
211  for (j = i; j > 1; j--)
212  pol[j] += pol[j - 2] - ROUND_MULL(lsp[2 * i], pol[j - 1], 16);
213 
214  pol[1] -= lsp[2 * i];
215  }
216 }
217 
218 static void silk_lsf2lpc(const int16_t nlsf[16], float lpcf[16], int order)
219 {
220  int i, k;
221  int32_t lsp[16]; // Q17; 2*cos(LSF)
222  int32_t p[9], q[9]; // Q16
223  int32_t lpc32[16]; // Q17
224  int16_t lpc[16]; // Q12
225 
226  /* convert the LSFs to LSPs, i.e. 2*cos(LSF) */
227  for (k = 0; k < order; k++) {
228  int index = nlsf[k] >> 8;
229  int offset = nlsf[k] & 255;
230  int k2 = (order == 10) ? ff_silk_lsf_ordering_nbmb[k] : ff_silk_lsf_ordering_wb[k];
231 
232  /* interpolate and round */
233  lsp[k2] = ff_silk_cosine[index] * 256;
234  lsp[k2] += (ff_silk_cosine[index + 1] - ff_silk_cosine[index]) * offset;
235  lsp[k2] = (lsp[k2] + 4) >> 3;
236  }
237 
238  silk_lsp2poly(lsp , p, order >> 1);
239  silk_lsp2poly(lsp + 1, q, order >> 1);
240 
241  /* reconstruct A(z) */
242  for (k = 0; k < order>>1; k++) {
243  int32_t p_tmp = p[k + 1] + p[k];
244  int32_t q_tmp = q[k + 1] - q[k];
245  lpc32[k] = -q_tmp - p_tmp;
246  lpc32[order-k-1] = q_tmp - p_tmp;
247  }
248 
249  /* limit the range of the LPC coefficients to each fit within an int16_t */
250  for (i = 0; i < 10; i++) {
251  int j;
252  unsigned int maxabs = 0;
253  for (j = 0, k = 0; j < order; j++) {
254  unsigned int x = FFABS(lpc32[k]);
255  if (x > maxabs) {
256  maxabs = x; // Q17
257  k = j;
258  }
259  }
260 
261  maxabs = (maxabs + 16) >> 5; // convert to Q12
262 
263  if (maxabs > 32767) {
264  /* perform bandwidth expansion */
265  unsigned int chirp, chirp_base; // Q16
266  maxabs = FFMIN(maxabs, 163838); // anything above this overflows chirp's numerator
267  chirp_base = chirp = 65470 - ((maxabs - 32767) << 14) / ((maxabs * (k+1)) >> 2);
268 
269  for (k = 0; k < order; k++) {
270  lpc32[k] = ROUND_MULL(lpc32[k], chirp, 16);
271  chirp = (chirp_base * chirp + 32768) >> 16;
272  }
273  } else break;
274  }
275 
276  if (i == 10) {
277  /* time's up: just clamp */
278  for (k = 0; k < order; k++) {
279  int x = (lpc32[k] + 16) >> 5;
280  lpc[k] = av_clip_int16(x);
281  lpc32[k] = lpc[k] << 5; // shortcut mandated by the spec; drops lower 5 bits
282  }
283  } else {
284  for (k = 0; k < order; k++)
285  lpc[k] = (lpc32[k] + 16) >> 5;
286  }
287 
288  /* if the prediction gain causes the LPC filter to become unstable,
289  apply further bandwidth expansion on the Q17 coefficients */
290  for (i = 1; i <= 16 && !silk_is_lpc_stable(lpc, order); i++) {
291  unsigned int chirp, chirp_base;
292  chirp_base = chirp = 65536 - (1 << i);
293 
294  for (k = 0; k < order; k++) {
295  lpc32[k] = ROUND_MULL(lpc32[k], chirp, 16);
296  lpc[k] = (lpc32[k] + 16) >> 5;
297  chirp = (chirp_base * chirp + 32768) >> 16;
298  }
299  }
300 
301  for (i = 0; i < order; i++)
302  lpcf[i] = lpc[i] / 4096.0f;
303 }
304 
306  OpusRangeCoder *rc,
307  float lpc_leadin[16], float lpc[16],
308  int *lpc_order, int *has_lpc_leadin, int voiced)
309 {
310  int i;
311  int order; // order of the LP polynomial; 10 for NB/MB and 16 for WB
312  int8_t lsf_i1, lsf_i2[16]; // stage-1 and stage-2 codebook indices
313  int16_t lsf_res[16]; // residual as a Q10 value
314  int16_t nlsf[16]; // Q15
315 
316  *lpc_order = order = s->wb ? 16 : 10;
317 
318  /* obtain LSF stage-1 and stage-2 indices */
319  lsf_i1 = ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_s1[s->wb][voiced]);
320  for (i = 0; i < order; i++) {
321  int index = s->wb ? ff_silk_lsf_s2_model_sel_wb [lsf_i1][i] :
323  lsf_i2[i] = ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_s2[index]) - 4;
324  if (lsf_i2[i] == -4)
326  else if (lsf_i2[i] == 4)
328  }
329 
330  /* reverse the backwards-prediction step */
331  for (i = order - 1; i >= 0; i--) {
332  int qstep = s->wb ? 9830 : 11796;
333 
334  lsf_res[i] = lsf_i2[i] * 1024;
335  if (lsf_i2[i] < 0) lsf_res[i] += 102;
336  else if (lsf_i2[i] > 0) lsf_res[i] -= 102;
337  lsf_res[i] = (lsf_res[i] * qstep) >> 16;
338 
339  if (i + 1 < order) {
342  lsf_res[i] += (lsf_res[i+1] * weight) >> 8;
343  }
344  }
345 
346  /* reconstruct the NLSF coefficients from the supplied indices */
347  for (i = 0; i < order; i++) {
348  const uint8_t * codebook = s->wb ? ff_silk_lsf_codebook_wb [lsf_i1] :
350  int cur, prev, next, weight_sq, weight, ipart, fpart, y, value;
351 
352  /* find the weight of the residual */
353  /* TODO: precompute */
354  cur = codebook[i];
355  prev = i ? codebook[i - 1] : 0;
356  next = i + 1 < order ? codebook[i + 1] : 256;
357  weight_sq = (1024 / (cur - prev) + 1024 / (next - cur)) << 16;
358 
359  /* approximate square-root with mandated fixed-point arithmetic */
360  ipart = opus_ilog(weight_sq);
361  fpart = (weight_sq >> (ipart-8)) & 127;
362  y = ((ipart & 1) ? 32768 : 46214) >> ((32 - ipart)>>1);
363  weight = y + ((213 * fpart * y) >> 16);
364 
365  value = cur * 128 + (lsf_res[i] * 16384) / weight;
366  nlsf[i] = av_clip_uintp2(value, 15);
367  }
368 
369  /* stabilize the NLSF coefficients */
370  silk_stabilize_lsf(nlsf, order, s->wb ? ff_silk_lsf_min_spacing_wb :
372 
373  /* produce an interpolation for the first 2 subframes, */
374  /* and then convert both sets of NLSFs to LPC coefficients */
375  *has_lpc_leadin = 0;
376  if (s->subframes == 4) {
378  if (offset != 4 && frame->coded) {
379  *has_lpc_leadin = 1;
380  if (offset != 0) {
381  int16_t nlsf_leadin[16];
382  for (i = 0; i < order; i++)
383  nlsf_leadin[i] = frame->nlsf[i] +
384  ((nlsf[i] - frame->nlsf[i]) * offset >> 2);
385  silk_lsf2lpc(nlsf_leadin, lpc_leadin, order);
386  } else /* avoid re-computation for a (roughly) 1-in-4 occurrence */
387  memcpy(lpc_leadin, frame->lpc, 16 * sizeof(float));
388  } else
389  offset = 4;
390  s->nlsf_interp_factor = offset;
391 
392  silk_lsf2lpc(nlsf, lpc, order);
393  } else {
394  s->nlsf_interp_factor = 4;
395  silk_lsf2lpc(nlsf, lpc, order);
396  }
397 
398  memcpy(frame->nlsf, nlsf, order * sizeof(nlsf[0]));
399  memcpy(frame->lpc, lpc, order * sizeof(lpc[0]));
400 }
401 
402 static inline void silk_count_children(OpusRangeCoder *rc, int model, int32_t total,
403  int32_t child[2])
404 {
405  if (total != 0) {
406  child[0] = ff_opus_rc_dec_cdf(rc,
407  ff_silk_model_pulse_location[model] + (((total - 1 + 5) * (total - 1)) >> 1));
408  child[1] = total - child[0];
409  } else {
410  child[0] = 0;
411  child[1] = 0;
412  }
413 }
414 
416  float* excitationf,
417  int qoffset_high, int active, int voiced)
418 {
419  int i;
420  uint32_t seed;
421  int shellblocks;
422  int ratelevel;
423  uint8_t pulsecount[20]; // total pulses in each shell block
424  uint8_t lsbcount[20] = {0}; // raw lsbits defined for each pulse in each shell block
425  int32_t excitation[320]; // Q23
426 
427  /* excitation parameters */
429  shellblocks = ff_silk_shell_blocks[s->bandwidth][s->subframes >> 2];
430  ratelevel = ff_opus_rc_dec_cdf(rc, ff_silk_model_exc_rate[voiced]);
431 
432  for (i = 0; i < shellblocks; i++) {
433  pulsecount[i] = ff_opus_rc_dec_cdf(rc, ff_silk_model_pulse_count[ratelevel]);
434  if (pulsecount[i] == 17) {
435  while (pulsecount[i] == 17 && ++lsbcount[i] != 10)
436  pulsecount[i] = ff_opus_rc_dec_cdf(rc, ff_silk_model_pulse_count[9]);
437  if (lsbcount[i] == 10)
438  pulsecount[i] = ff_opus_rc_dec_cdf(rc, ff_silk_model_pulse_count[10]);
439  }
440  }
441 
442  /* decode pulse locations using PVQ */
443  for (i = 0; i < shellblocks; i++) {
444  if (pulsecount[i] != 0) {
445  int a, b, c, d;
446  int32_t * location = excitation + 16*i;
447  int32_t branch[4][2];
448  branch[0][0] = pulsecount[i];
449 
450  /* unrolled tail recursion */
451  for (a = 0; a < 1; a++) {
452  silk_count_children(rc, 0, branch[0][a], branch[1]);
453  for (b = 0; b < 2; b++) {
454  silk_count_children(rc, 1, branch[1][b], branch[2]);
455  for (c = 0; c < 2; c++) {
456  silk_count_children(rc, 2, branch[2][c], branch[3]);
457  for (d = 0; d < 2; d++) {
458  silk_count_children(rc, 3, branch[3][d], location);
459  location += 2;
460  }
461  }
462  }
463  }
464  } else
465  memset(excitation + 16*i, 0, 16*sizeof(int32_t));
466  }
467 
468  /* decode least significant bits */
469  for (i = 0; i < shellblocks << 4; i++) {
470  int bit;
471  for (bit = 0; bit < lsbcount[i >> 4]; bit++)
472  excitation[i] = (excitation[i] << 1) |
474  }
475 
476  /* decode signs */
477  for (i = 0; i < shellblocks << 4; i++) {
478  if (excitation[i] != 0) {
479  int sign = ff_opus_rc_dec_cdf(rc, ff_silk_model_excitation_sign[active +
480  voiced][qoffset_high][FFMIN(pulsecount[i >> 4], 6)]);
481  if (sign == 0)
482  excitation[i] *= -1;
483  }
484  }
485 
486  /* assemble the excitation */
487  for (i = 0; i < shellblocks << 4; i++) {
488  int value = excitation[i];
489  excitation[i] = value * 256 | ff_silk_quant_offset[voiced][qoffset_high];
490  if (value < 0) excitation[i] += 20;
491  else if (value > 0) excitation[i] -= 20;
492 
493  /* invert samples pseudorandomly */
494  seed = 196314165 * seed + 907633515;
495  if (seed & 0x80000000)
496  excitation[i] *= -1;
497  seed += value;
498 
499  excitationf[i] = excitation[i] / 8388608.0f;
500  }
501 }
502 
503 /** Maximum residual history according to 4.2.7.6.1 */
504 #define SILK_MAX_LAG (288 + LTP_ORDER / 2)
505 
506 /** Order of the LTP filter */
507 #define LTP_ORDER 5
508 
510  int frame_num, int channel, int coded_channels,
511  int active, int active1, int redundant)
512 {
513  /* per frame */
514  int voiced; // combines with active to indicate inactive, active, or active+voiced
515  int qoffset_high;
516  int order; // order of the LPC coefficients
517  float lpc_leadin[16], lpc_body[16], residual[SILK_MAX_LAG + SILK_HISTORY];
518  int has_lpc_leadin;
519  float ltpscale;
520 
521  /* per subframe */
522  struct {
523  float gain;
524  int pitchlag;
525  float ltptaps[5];
526  } sf[4];
527 
528  SilkFrame * const frame = s->frame + channel;
529 
530  int i;
531 
532  /* obtain stereo weights */
533  if (coded_channels == 2 && channel == 0) {
534  int n, wi[2], ws[2], w[2];
536  wi[0] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s2) + 3 * (n / 5);
538  wi[1] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s2) + 3 * (n % 5);
540 
541  for (i = 0; i < 2; i++)
542  w[i] = ff_silk_stereo_weights[wi[i]] +
543  (((ff_silk_stereo_weights[wi[i] + 1] - ff_silk_stereo_weights[wi[i]]) * 6554) >> 16)
544  * (ws[i]*2 + 1);
545 
546  s->stereo_weights[0] = (w[0] - w[1]) / 8192.0;
547  s->stereo_weights[1] = w[1] / 8192.0;
548 
549  /* and read the mid-only flag */
550  s->midonly = active1 ? 0 : ff_opus_rc_dec_cdf(rc, ff_silk_model_mid_only);
551  }
552 
553  /* obtain frame type */
554  if (!active) {
556  voiced = 0;
557  } else {
559  qoffset_high = type & 1;
560  voiced = type >> 1;
561  }
562 
563  /* obtain subframe quantization gains */
564  for (i = 0; i < s->subframes; i++) {
565  int log_gain; //Q7
566  int ipart, fpart, lingain;
567 
568  if (i == 0 && (frame_num == 0 || !frame->coded)) {
569  /* gain is coded absolute */
570  int x = ff_opus_rc_dec_cdf(rc, ff_silk_model_gain_highbits[active + voiced]);
571  log_gain = (x<<3) | ff_opus_rc_dec_cdf(rc, ff_silk_model_gain_lowbits);
572 
573  if (frame->coded)
574  log_gain = FFMAX(log_gain, frame->log_gain - 16);
575  } else {
576  /* gain is coded relative */
577  int delta_gain = ff_opus_rc_dec_cdf(rc, ff_silk_model_gain_delta);
578  log_gain = av_clip_uintp2(FFMAX((delta_gain<<1) - 16,
579  frame->log_gain + delta_gain - 4), 6);
580  }
581 
582  frame->log_gain = log_gain;
583 
584  /* approximate 2**(x/128) with a Q7 (i.e. non-integer) input */
585  log_gain = (log_gain * 0x1D1C71 >> 16) + 2090;
586  ipart = log_gain >> 7;
587  fpart = log_gain & 127;
588  lingain = (1 << ipart) + ((-174 * fpart * (128-fpart) >>16) + fpart) * ((1<<ipart) >> 7);
589  sf[i].gain = lingain / 65536.0f;
590  }
591 
592  /* obtain LPC filter coefficients */
593  silk_decode_lpc(s, frame, rc, lpc_leadin, lpc_body, &order, &has_lpc_leadin, voiced);
594 
595  /* obtain pitch lags, if this is a voiced frame */
596  if (voiced) {
597  int lag_absolute = (!frame_num || !frame->prev_voiced);
598  int primarylag; // primary pitch lag for the entire SILK frame
599  int ltpfilter;
600  const int8_t * offsets;
601 
602  if (!lag_absolute) {
604  if (delta)
605  primarylag = frame->primarylag + delta - 9;
606  else
607  lag_absolute = 1;
608  }
609 
610  if (lag_absolute) {
611  /* primary lag is coded absolute */
612  int highbits, lowbits;
613  static const uint16_t * const model[] = {
616  };
618  lowbits = ff_opus_rc_dec_cdf(rc, model[s->bandwidth]);
619 
620  primarylag = ff_silk_pitch_min_lag[s->bandwidth] +
621  highbits*ff_silk_pitch_scale[s->bandwidth] + lowbits;
622  }
623  frame->primarylag = primarylag;
624 
625  if (s->subframes == 2)
626  offsets = (s->bandwidth == OPUS_BANDWIDTH_NARROWBAND)
631  else
632  offsets = (s->bandwidth == OPUS_BANDWIDTH_NARROWBAND)
637 
638  for (i = 0; i < s->subframes; i++)
639  sf[i].pitchlag = av_clip(primarylag + offsets[i],
640  ff_silk_pitch_min_lag[s->bandwidth],
641  ff_silk_pitch_max_lag[s->bandwidth]);
642 
643  /* obtain LTP filter coefficients */
645  for (i = 0; i < s->subframes; i++) {
646  int index, j;
647  static const uint16_t * const filter_sel[] = {
650  };
651  static const int8_t (* const filter_taps[])[5] = {
653  };
654  index = ff_opus_rc_dec_cdf(rc, filter_sel[ltpfilter]);
655  for (j = 0; j < 5; j++)
656  sf[i].ltptaps[j] = filter_taps[ltpfilter][index][j] / 128.0f;
657  }
658  }
659 
660  /* obtain LTP scale factor */
661  if (voiced && frame_num == 0)
663  ff_silk_model_ltp_scale_index)] / 16384.0f;
664  else ltpscale = 15565.0f/16384.0f;
665 
666  /* generate the excitation signal for the entire frame */
667  silk_decode_excitation(s, rc, residual + SILK_MAX_LAG, qoffset_high,
668  active, voiced);
669 
670  /* skip synthesising the output if we do not need it */
671  // TODO: implement error recovery
672  if (s->output_channels == channel || redundant)
673  return;
674 
675  /* generate the output signal */
676  for (i = 0; i < s->subframes; i++) {
677  const float * lpc_coeff = (i < 2 && has_lpc_leadin) ? lpc_leadin : lpc_body;
678  float *dst = frame->output + SILK_HISTORY + i * s->sflength;
679  float *resptr = residual + SILK_MAX_LAG + i * s->sflength;
680  float *lpc = frame->lpc_history + SILK_HISTORY + i * s->sflength;
681  float sum;
682  int j, k;
683 
684  if (voiced) {
685  int out_end;
686  float scale;
687 
688  if (i < 2 || s->nlsf_interp_factor == 4) {
689  out_end = -i * s->sflength;
690  scale = ltpscale;
691  } else {
692  out_end = -(i - 2) * s->sflength;
693  scale = 1.0f;
694  }
695 
696  /* when the LPC coefficients change, a re-whitening filter is used */
697  /* to produce a residual that accounts for the change */
698  for (j = - sf[i].pitchlag - LTP_ORDER/2; j < out_end; j++) {
699  sum = dst[j];
700  for (k = 0; k < order; k++)
701  sum -= lpc_coeff[k] * dst[j - k - 1];
702  resptr[j] = av_clipf(sum, -1.0f, 1.0f) * scale / sf[i].gain;
703  }
704 
705  if (out_end) {
706  float rescale = sf[i-1].gain / sf[i].gain;
707  for (j = out_end; j < 0; j++)
708  resptr[j] *= rescale;
709  }
710 
711  /* LTP synthesis */
712  for (j = 0; j < s->sflength; j++) {
713  sum = resptr[j];
714  for (k = 0; k < LTP_ORDER; k++)
715  sum += sf[i].ltptaps[k] * resptr[j - sf[i].pitchlag + LTP_ORDER/2 - k];
716  resptr[j] = sum;
717  }
718  }
719 
720  /* LPC synthesis */
721  for (j = 0; j < s->sflength; j++) {
722  sum = resptr[j] * sf[i].gain;
723  for (k = 1; k <= order; k++)
724  sum += lpc_coeff[k - 1] * lpc[j - k];
725 
726  lpc[j] = sum;
727  dst[j] = av_clipf(sum, -1.0f, 1.0f);
728  }
729  }
730 
731  frame->prev_voiced = voiced;
732  memmove(frame->lpc_history, frame->lpc_history + s->flength, SILK_HISTORY * sizeof(float));
733  memmove(frame->output, frame->output + s->flength, SILK_HISTORY * sizeof(float));
734 
735  frame->coded = 1;
736 }
737 
738 static void silk_unmix_ms(SilkContext *s, float *l, float *r)
739 {
740  float *mid = s->frame[0].output + SILK_HISTORY - s->flength;
741  float *side = s->frame[1].output + SILK_HISTORY - s->flength;
742  float w0_prev = s->prev_stereo_weights[0];
743  float w1_prev = s->prev_stereo_weights[1];
744  float w0 = s->stereo_weights[0];
745  float w1 = s->stereo_weights[1];
746  int n1 = ff_silk_stereo_interp_len[s->bandwidth];
747  int i;
748 
749  for (i = 0; i < n1; i++) {
750  float interp0 = w0_prev + i * (w0 - w0_prev) / n1;
751  float interp1 = w1_prev + i * (w1 - w1_prev) / n1;
752  float p0 = 0.25 * (mid[i - 2] + 2 * mid[i - 1] + mid[i]);
753 
754  l[i] = av_clipf((1 + interp1) * mid[i - 1] + side[i - 1] + interp0 * p0, -1.0, 1.0);
755  r[i] = av_clipf((1 - interp1) * mid[i - 1] - side[i - 1] - interp0 * p0, -1.0, 1.0);
756  }
757 
758  for (; i < s->flength; i++) {
759  float p0 = 0.25 * (mid[i - 2] + 2 * mid[i - 1] + mid[i]);
760 
761  l[i] = av_clipf((1 + w1) * mid[i - 1] + side[i - 1] + w0 * p0, -1.0, 1.0);
762  r[i] = av_clipf((1 - w1) * mid[i - 1] - side[i - 1] - w0 * p0, -1.0, 1.0);
763  }
764 
765  memcpy(s->prev_stereo_weights, s->stereo_weights, sizeof(s->stereo_weights));
766 }
767 
769 {
770  if (!frame->coded)
771  return;
772 
773  memset(frame->output, 0, sizeof(frame->output));
774  memset(frame->lpc_history, 0, sizeof(frame->lpc_history));
775 
776  memset(frame->lpc, 0, sizeof(frame->lpc));
777  memset(frame->nlsf, 0, sizeof(frame->nlsf));
778 
779  frame->log_gain = 0;
780 
781  frame->primarylag = 0;
782  frame->prev_voiced = 0;
783  frame->coded = 0;
784 }
785 
787  float *output[2],
788  enum OpusBandwidth bandwidth,
789  int coded_channels,
790  int duration_ms)
791 {
792  int active[2][6], redundancy[2];
793  int nb_frames, i, j;
794 
795  if (bandwidth > OPUS_BANDWIDTH_WIDEBAND ||
796  coded_channels > 2 || duration_ms > 60) {
797  av_log(s->avctx, AV_LOG_ERROR, "Invalid parameters passed "
798  "to the SILK decoder.\n");
799  return AVERROR(EINVAL);
800  }
801 
802  nb_frames = 1 + (duration_ms > 20) + (duration_ms > 40);
803  s->subframes = duration_ms / nb_frames / 5; // 5ms subframes
804  s->sflength = 20 * (bandwidth + 2);
805  s->flength = s->sflength * s->subframes;
806  s->bandwidth = bandwidth;
807  s->wb = bandwidth == OPUS_BANDWIDTH_WIDEBAND;
808 
809  /* make sure to flush the side channel when switching from mono to stereo */
810  if (coded_channels > s->prev_coded_channels)
811  silk_flush_frame(&s->frame[1]);
812  s->prev_coded_channels = coded_channels;
813 
814  /* read the LP-layer header bits */
815  for (i = 0; i < coded_channels; i++) {
816  for (j = 0; j < nb_frames; j++)
817  active[i][j] = ff_opus_rc_dec_log(rc, 1);
818 
819  redundancy[i] = ff_opus_rc_dec_log(rc, 1);
820  }
821 
822  /* read the per-frame LBRR flags */
823  for (i = 0; i < coded_channels; i++)
824  if (redundancy[i] && duration_ms > 20) {
825  redundancy[i] = ff_opus_rc_dec_cdf(rc, duration_ms == 40 ?
827  }
828 
829  /* decode the LBRR frames */
830  for (i = 0; i < nb_frames; i++) {
831  for (j = 0; j < coded_channels; j++)
832  if (redundancy[j] & (1 << i)) {
833  int active1 = (j == 0 && !(redundancy[1] & (1 << i))) ? 0 : 1;
834  silk_decode_frame(s, rc, i, j, coded_channels, 1, active1, 1);
835  }
836  }
837 
838  for (i = 0; i < nb_frames; i++) {
839  for (j = 0; j < coded_channels && !s->midonly; j++)
840  silk_decode_frame(s, rc, i, j, coded_channels, active[j][i], active[1][i], 0);
841 
842  /* reset the side channel if it is not coded */
843  if (s->midonly && s->frame[1].coded)
844  silk_flush_frame(&s->frame[1]);
845 
846  if (coded_channels == 1 || s->output_channels == 1) {
847  for (j = 0; j < s->output_channels; j++) {
848  memcpy(output[j] + i * s->flength,
849  s->frame[0].output + SILK_HISTORY - s->flength - 2,
850  s->flength * sizeof(float));
851  }
852  } else {
853  silk_unmix_ms(s, output[0] + i * s->flength, output[1] + i * s->flength);
854  }
855 
856  s->midonly = 0;
857  }
858 
859  return nb_frames * s->flength;
860 }
861 
863 {
864  av_freep(ps);
865 }
866 
868 {
869  silk_flush_frame(&s->frame[0]);
870  silk_flush_frame(&s->frame[1]);
871 
872  memset(s->prev_stereo_weights, 0, sizeof(s->prev_stereo_weights));
873 }
874 
875 int ff_silk_init(AVCodecContext *avctx, SilkContext **ps, int output_channels)
876 {
877  SilkContext *s;
878 
879  if (output_channels != 1 && output_channels != 2) {
880  av_log(avctx, AV_LOG_ERROR, "Invalid number of output channels: %d\n",
881  output_channels);
882  return AVERROR(EINVAL);
883  }
884 
885  s = av_mallocz(sizeof(*s));
886  if (!s)
887  return AVERROR(ENOMEM);
888 
889  s->avctx = avctx;
890  s->output_channels = output_channels;
891 
892  ff_silk_flush(s);
893 
894  *ps = s;
895 
896  return 0;
897 }
uint8_t
int32_t
#define bit(string, value)
Definition: cbs_mpeg2.c:58
#define s(width, name)
Definition: cbs_vp9.c:257
static const unsigned codebook[256][2]
Definition: cfhdenc.c:42
#define FFMIN(a, b)
Definition: common.h:105
#define av_clip
Definition: common.h:122
#define av_clip_int16
Definition: common.h:137
#define FFMAX(a, b)
Definition: common.h:103
#define av_clip_uintp2
Definition: common.h:146
#define av_clipf
Definition: common.h:170
#define FFABS(a)
Absolute value, Note, INT_MIN / INT64_MIN result in undefined behavior as they are not representable ...
Definition: common.h:72
#define av_sat_sub32
Definition: common.h:158
long long int64_t
Definition: coverity.c:34
static AVFrame * frame
channel
Use these values when setting the channel map with ebur128_set_channel().
Definition: ebur128.h:39
double value
Definition: eval.c:98
#define AVERROR(e)
Definition: error.h:43
#define AV_LOG_ERROR
Something went wrong and cannot losslessly be recovered.
Definition: log.h:194
void * av_mallocz(size_t size)
Allocate a memory block with alignment suitable for all memory accesses (including vectors if availab...
Definition: mem.c:237
int index
Definition: gxfenc.c:89
static const int offsets[]
Definition: hevc_pel.c:34
cl_device_type type
int i
Definition: input.c:407
static int weight(int i, int blen, int offset)
Definition: diracdec.c:1561
static void rescale(GDVContext *gdv, uint8_t *dst, int w, int h, int scale_v, int scale_h)
Definition: gdv.c:130
uint8_t w
Definition: llviddspenc.c:39
#define MULL(a, b, s)
Definition: mathops.h:58
#define ROUND_MULL(a, b, s)
Definition: opus.h:51
OpusBandwidth
Definition: opus.h:71
@ OPUS_BANDWIDTH_NARROWBAND
Definition: opus.h:72
@ OPUS_BANDWIDTH_WIDEBAND
Definition: opus.h:74
#define SILK_HISTORY
Definition: opus.h:48
uint32_t ff_opus_rc_dec_log(OpusRangeCoder *rc, uint32_t bits)
Definition: opus_rc.c:114
uint32_t ff_opus_rc_dec_cdf(OpusRangeCoder *rc, const uint16_t *cdf)
Definition: opus_rc.c:90
#define opus_ilog(i)
Definition: opus_rc.h:31
static void silk_decode_lpc(SilkContext *s, SilkFrame *frame, OpusRangeCoder *rc, float lpc_leadin[16], float lpc[16], int *lpc_order, int *has_lpc_leadin, int voiced)
Definition: opus_silk.c:305
#define LTP_ORDER
Order of the LTP filter.
Definition: opus_silk.c:507
int ff_silk_decode_superframe(SilkContext *s, OpusRangeCoder *rc, float *output[2], enum OpusBandwidth bandwidth, int coded_channels, int duration_ms)
Decode the LP layer of one Opus frame (which may correspond to several SILK frames).
Definition: opus_silk.c:786
static void silk_flush_frame(SilkFrame *frame)
Definition: opus_silk.c:768
#define SILK_MAX_LAG
Maximum residual history according to 4.2.7.6.1.
Definition: opus_silk.c:504
static void silk_unmix_ms(SilkContext *s, float *l, float *r)
Definition: opus_silk.c:738
static void silk_stabilize_lsf(int16_t nlsf[16], int order, const uint16_t min_delta[17])
Definition: opus_silk.c:65
void ff_silk_flush(SilkContext *s)
Definition: opus_silk.c:867
static void silk_decode_excitation(SilkContext *s, OpusRangeCoder *rc, float *excitationf, int qoffset_high, int active, int voiced)
Definition: opus_silk.c:415
void ff_silk_free(SilkContext **ps)
Definition: opus_silk.c:862
static void silk_lsp2poly(const int32_t lsp[], int32_t pol[], int half_order)
Definition: opus_silk.c:201
static void silk_count_children(OpusRangeCoder *rc, int model, int32_t total, int32_t child[2])
Definition: opus_silk.c:402
static int silk_is_lpc_stable(const int16_t lpc[16], int order)
Definition: opus_silk.c:143
int ff_silk_init(AVCodecContext *avctx, SilkContext **ps, int output_channels)
Definition: opus_silk.c:875
static void silk_decode_frame(SilkContext *s, OpusRangeCoder *rc, int frame_num, int channel, int coded_channels, int active, int active1, int redundant)
Definition: opus_silk.c:509
static void silk_lsf2lpc(const int16_t nlsf[16], float lpcf[16], int order)
Definition: opus_silk.c:218
const uint8_t ff_silk_lsf_weight_sel_nbmb[32][9]
Definition: opustab.c:406
const uint16_t ff_silk_model_stereo_s3[]
Definition: opustab.c:41
const uint16_t ff_silk_model_excitation_sign[3][2][7][3]
Definition: opustab.c:263
const uint8_t ff_silk_lsf_s2_model_sel_wb[32][16]
Definition: opustab.c:361
const uint16_t ff_silk_model_ltp_filter[]
Definition: opustab.c:140
const uint16_t ff_silk_lsf_min_spacing_wb[]
Definition: opustab.c:550
const uint16_t ff_silk_pitch_scale[]
Definition: opustab.c:598
const uint16_t ff_silk_model_pitch_contour_nb10ms[]
Definition: opustab.c:124
const uint16_t ff_silk_ltp_scale_factor[]
Definition: opustab.c:741
const uint16_t ff_silk_model_stereo_s1[]
Definition: opustab.c:34
const int16_t ff_silk_cosine[]
Definition: opustab.c:562
const uint8_t ff_silk_lsf_codebook_nbmb[32][10]
Definition: opustab.c:476
const uint16_t ff_silk_model_lsf_interpolation_offset[]
Definition: opustab.c:106
const uint16_t ff_silk_model_pitch_lowbits_mb[]
Definition: opustab.c:115
const uint16_t ff_silk_model_lbrr_flags_60[]
Definition: opustab.c:32
const uint8_t ff_silk_lsf_pred_weights_wb[2][15]
Definition: opustab.c:401
const uint8_t ff_silk_lsf_codebook_wb[32][16]
Definition: opustab.c:511
const uint16_t ff_silk_model_mid_only[]
Definition: opustab.c:43
const uint16_t ff_silk_model_ltp_filter2_sel[]
Definition: opustab.c:150
const uint16_t ff_silk_pitch_max_lag[]
Definition: opustab.c:602
const uint8_t ff_silk_lsf_weight_sel_wb[32][15]
Definition: opustab.c:441
const uint16_t ff_silk_model_stereo_s2[]
Definition: opustab.c:39
const uint16_t ff_silk_model_pitch_lowbits_wb[]
Definition: opustab.c:117
const uint8_t ff_silk_lsf_ordering_nbmb[]
Definition: opustab.c:554
const int8_t ff_silk_pitch_offset_mbwb10ms[12][2]
Definition: opustab.c:624
const uint16_t ff_silk_lsf_min_spacing_nbmb[]
Definition: opustab.c:546
const uint16_t ff_silk_model_pulse_location[4][168]
Definition: opustab.c:189
const uint8_t ff_silk_lsf_pred_weights_nbmb[2][9]
Definition: opustab.c:396
const uint8_t ff_silk_lsf_ordering_wb[]
Definition: opustab.c:558
const uint16_t ff_silk_model_lbrr_flags_40[]
Definition: opustab.c:31
const uint16_t ff_silk_model_gain_delta[]
Definition: opustab.c:57
const uint16_t ff_silk_model_gain_lowbits[]
Definition: opustab.c:55
const uint16_t ff_silk_model_ltp_filter1_sel[]
Definition: opustab.c:146
const uint16_t ff_silk_model_pitch_contour_mbwb10ms[]
Definition: opustab.c:130
const uint16_t ff_silk_model_pitch_contour_mbwb20ms[]
Definition: opustab.c:134
const int8_t ff_silk_pitch_offset_nb20ms[11][4]
Definition: opustab.c:610
const uint16_t ff_silk_model_frame_type_inactive[]
Definition: opustab.c:45
const int ff_silk_stereo_interp_len[3]
Definition: opustab.c:754
const uint16_t ff_silk_model_excitation_lsb[]
Definition: opustab.c:261
const int8_t ff_silk_ltp_filter1_taps[16][5]
Definition: opustab.c:687
const int8_t ff_silk_ltp_filter2_taps[32][5]
Definition: opustab.c:706
const uint16_t ff_silk_model_frame_type_active[]
Definition: opustab.c:47
const uint8_t ff_silk_shell_blocks[3][2]
Definition: opustab.c:743
const uint8_t ff_silk_lsf_s2_model_sel_nbmb[32][10]
Definition: opustab.c:326
const uint8_t ff_silk_quant_offset[2][2]
Definition: opustab.c:749
const uint16_t ff_silk_model_pulse_count[11][19]
Definition: opustab.c:164
const uint16_t ff_silk_pitch_min_lag[]
Definition: opustab.c:600
const uint16_t ff_silk_model_gain_highbits[3][9]
Definition: opustab.c:49
const uint16_t ff_silk_model_pitch_delta[]
Definition: opustab.c:119
const uint16_t ff_silk_model_lsf_s2[32][10]
Definition: opustab.c:82
const uint16_t ff_silk_model_ltp_filter0_sel[]
Definition: opustab.c:142
const uint16_t ff_silk_model_pitch_lowbits_nb[]
Definition: opustab.c:113
const int8_t ff_silk_pitch_offset_nb10ms[3][2]
Definition: opustab.c:604
const uint16_t ff_silk_model_lsf_s2_ext[]
Definition: opustab.c:104
const uint16_t ff_silk_model_lcg_seed[]
Definition: opustab.c:157
const int16_t ff_silk_stereo_weights[]
Definition: opustab.c:321
const int8_t ff_silk_pitch_offset_mbwb20ms[34][4]
Definition: opustab.c:639
const uint16_t ff_silk_model_pitch_highbits[]
Definition: opustab.c:108
const uint16_t ff_silk_model_exc_rate[2][10]
Definition: opustab.c:159
const uint16_t ff_silk_model_pitch_contour_nb20ms[]
Definition: opustab.c:126
const uint16_t ff_silk_model_lsf_s1[2][2][33]
Definition: opustab.c:62
const int8_t ff_silk_ltp_filter0_taps[8][5]
Definition: opustab.c:676
const uint16_t ff_silk_model_ltp_scale_index[]
Definition: opustab.c:155
#define MULH
Definition: mathops.h:42
main external API structure.
Definition: avcodec.h:536
int output_channels
Definition: opus_silk.c:47
int midonly
Definition: opus_silk.c:49
float prev_stereo_weights[2]
Definition: opus_silk.c:59
SilkFrame frame[2]
Definition: opus_silk.c:58
int subframes
Definition: opus_silk.c:50
int nlsf_interp_factor
Definition: opus_silk.c:53
AVCodecContext * avctx
Definition: opus_silk.c:46
enum OpusBandwidth bandwidth
Definition: opus_silk.c:55
int flength
Definition: opus_silk.c:52
int sflength
Definition: opus_silk.c:51
float stereo_weights[2]
Definition: opus_silk.c:60
int prev_coded_channels
Definition: opus_silk.c:62
int16_t nlsf[16]
Definition: opus_silk.c:35
int primarylag
Definition: opus_silk.c:40
int coded
Definition: opus_silk.c:33
float lpc_history[2 *SILK_HISTORY]
Definition: opus_silk.c:39
float lpc[16]
Definition: opus_silk.c:36
int prev_voiced
Definition: opus_silk.c:42
int log_gain
Definition: opus_silk.c:34
float output[2 *SILK_HISTORY]
Definition: opus_silk.c:38
#define av_freep(p)
#define av_log(a,...)
static void error(const char *err)
static uint8_t tmp[11]
Definition: aes_ctr.c:27
#define pass
Definition: tx_template.c:347
const char * b
Definition: vf_curves.c:118
const char * r
Definition: vf_curves.c:116
static av_always_inline int diff(const uint32_t a, const uint32_t b)
static const uint8_t offset[127][2]
Definition: vf_spp.c:107
static unsigned int seed
Definition: videogen.c:78
float delta
static double c[64]