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];
39 float lpc_history[2 * SILK_HISTORY];
40 int primarylag;
41
42 int prev_voiced;
43 } SilkFrame;
44
45 struct SilkContext {
46 AVCodecContext *avctx;
47 int output_channels;
48
49 int midonly;
50 int subframes;
51 int sflength;
52 int flength;
53 int nlsf_interp_factor;
54
55 enum OpusBandwidth bandwidth;
56 int wb;
57
58 SilkFrame frame[2];
59 float prev_stereo_weights[2];
60 float stereo_weights[2];
61
62 int prev_coded_channels;
63 };
64
silk_stabilize_lsf(int16_t nlsf[16],int order,const uint16_t min_delta[17])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
silk_is_lpc_stable(const int16_t lpc[16],int order)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
silk_lsp2poly(const int32_t lsp[16],int32_t pol[16],int half_order)201 static void silk_lsp2poly(const int32_t lsp[16], int32_t pol[16], int half_order)
202 {
203 int i, j;
204
205 pol[0] = 65536; // 1.0 in Q16
206 pol[1] = -lsp[0];
207
208 for (i = 1; i < half_order; i++) {
209 pol[i + 1] = pol[i - 1] * 2 - ROUND_MULL(lsp[2 * i], pol[i], 16);
210 for (j = i; j > 1; j--)
211 pol[j] += pol[j - 2] - ROUND_MULL(lsp[2 * i], pol[j - 1], 16);
212
213 pol[1] -= lsp[2 * i];
214 }
215 }
216
silk_lsf2lpc(const int16_t nlsf[16],float lpcf[16],int order)217 static void silk_lsf2lpc(const int16_t nlsf[16], float lpcf[16], int order)
218 {
219 int i, k;
220 int32_t lsp[16]; // Q17; 2*cos(LSF)
221 int32_t p[9], q[9]; // Q16
222 int32_t lpc32[16]; // Q17
223 int16_t lpc[16]; // Q12
224
225 /* convert the LSFs to LSPs, i.e. 2*cos(LSF) */
226 for (k = 0; k < order; k++) {
227 int index = nlsf[k] >> 8;
228 int offset = nlsf[k] & 255;
229 int k2 = (order == 10) ? ff_silk_lsf_ordering_nbmb[k] : ff_silk_lsf_ordering_wb[k];
230
231 /* interpolate and round */
232 lsp[k2] = ff_silk_cosine[index] * 256;
233 lsp[k2] += (ff_silk_cosine[index + 1] - ff_silk_cosine[index]) * offset;
234 lsp[k2] = (lsp[k2] + 4) >> 3;
235 }
236
237 silk_lsp2poly(lsp , p, order >> 1);
238 silk_lsp2poly(lsp + 1, q, order >> 1);
239
240 /* reconstruct A(z) */
241 for (k = 0; k < order>>1; k++) {
242 int32_t p_tmp = p[k + 1] + p[k];
243 int32_t q_tmp = q[k + 1] - q[k];
244 lpc32[k] = -q_tmp - p_tmp;
245 lpc32[order-k-1] = q_tmp - p_tmp;
246 }
247
248 /* limit the range of the LPC coefficients to each fit within an int16_t */
249 for (i = 0; i < 10; i++) {
250 int j;
251 unsigned int maxabs = 0;
252 for (j = 0, k = 0; j < order; j++) {
253 unsigned int x = FFABS(lpc32[k]);
254 if (x > maxabs) {
255 maxabs = x; // Q17
256 k = j;
257 }
258 }
259
260 maxabs = (maxabs + 16) >> 5; // convert to Q12
261
262 if (maxabs > 32767) {
263 /* perform bandwidth expansion */
264 unsigned int chirp, chirp_base; // Q16
265 maxabs = FFMIN(maxabs, 163838); // anything above this overflows chirp's numerator
266 chirp_base = chirp = 65470 - ((maxabs - 32767) << 14) / ((maxabs * (k+1)) >> 2);
267
268 for (k = 0; k < order; k++) {
269 lpc32[k] = ROUND_MULL(lpc32[k], chirp, 16);
270 chirp = (chirp_base * chirp + 32768) >> 16;
271 }
272 } else break;
273 }
274
275 if (i == 10) {
276 /* time's up: just clamp */
277 for (k = 0; k < order; k++) {
278 int x = (lpc32[k] + 16) >> 5;
279 lpc[k] = av_clip_int16(x);
280 lpc32[k] = lpc[k] << 5; // shortcut mandated by the spec; drops lower 5 bits
281 }
282 } else {
283 for (k = 0; k < order; k++)
284 lpc[k] = (lpc32[k] + 16) >> 5;
285 }
286
287 /* if the prediction gain causes the LPC filter to become unstable,
288 apply further bandwidth expansion on the Q17 coefficients */
289 for (i = 1; i <= 16 && !silk_is_lpc_stable(lpc, order); i++) {
290 unsigned int chirp, chirp_base;
291 chirp_base = chirp = 65536 - (1 << i);
292
293 for (k = 0; k < order; k++) {
294 lpc32[k] = ROUND_MULL(lpc32[k], chirp, 16);
295 lpc[k] = (lpc32[k] + 16) >> 5;
296 chirp = (chirp_base * chirp + 32768) >> 16;
297 }
298 }
299
300 for (i = 0; i < order; i++)
301 lpcf[i] = lpc[i] / 4096.0f;
302 }
303
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)304 static inline void silk_decode_lpc(SilkContext *s, SilkFrame *frame,
305 OpusRangeCoder *rc,
306 float lpc_leadin[16], float lpc[16],
307 int *lpc_order, int *has_lpc_leadin, int voiced)
308 {
309 int i;
310 int order; // order of the LP polynomial; 10 for NB/MB and 16 for WB
311 int8_t lsf_i1, lsf_i2[16]; // stage-1 and stage-2 codebook indices
312 int16_t lsf_res[16]; // residual as a Q10 value
313 int16_t nlsf[16]; // Q15
314
315 *lpc_order = order = s->wb ? 16 : 10;
316
317 /* obtain LSF stage-1 and stage-2 indices */
318 lsf_i1 = ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_s1[s->wb][voiced]);
319 for (i = 0; i < order; i++) {
320 int index = s->wb ? ff_silk_lsf_s2_model_sel_wb [lsf_i1][i] :
321 ff_silk_lsf_s2_model_sel_nbmb[lsf_i1][i];
322 lsf_i2[i] = ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_s2[index]) - 4;
323 if (lsf_i2[i] == -4)
324 lsf_i2[i] -= ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_s2_ext);
325 else if (lsf_i2[i] == 4)
326 lsf_i2[i] += ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_s2_ext);
327 }
328
329 /* reverse the backwards-prediction step */
330 for (i = order - 1; i >= 0; i--) {
331 int qstep = s->wb ? 9830 : 11796;
332
333 lsf_res[i] = lsf_i2[i] * 1024;
334 if (lsf_i2[i] < 0) lsf_res[i] += 102;
335 else if (lsf_i2[i] > 0) lsf_res[i] -= 102;
336 lsf_res[i] = (lsf_res[i] * qstep) >> 16;
337
338 if (i + 1 < order) {
339 int weight = s->wb ? ff_silk_lsf_pred_weights_wb [ff_silk_lsf_weight_sel_wb [lsf_i1][i]][i] :
340 ff_silk_lsf_pred_weights_nbmb[ff_silk_lsf_weight_sel_nbmb[lsf_i1][i]][i];
341 lsf_res[i] += (lsf_res[i+1] * weight) >> 8;
342 }
343 }
344
345 /* reconstruct the NLSF coefficients from the supplied indices */
346 for (i = 0; i < order; i++) {
347 const uint8_t * codebook = s->wb ? ff_silk_lsf_codebook_wb [lsf_i1] :
348 ff_silk_lsf_codebook_nbmb[lsf_i1];
349 int cur, prev, next, weight_sq, weight, ipart, fpart, y, value;
350
351 /* find the weight of the residual */
352 /* TODO: precompute */
353 cur = codebook[i];
354 prev = i ? codebook[i - 1] : 0;
355 next = i + 1 < order ? codebook[i + 1] : 256;
356 weight_sq = (1024 / (cur - prev) + 1024 / (next - cur)) << 16;
357
358 /* approximate square-root with mandated fixed-point arithmetic */
359 ipart = opus_ilog(weight_sq);
360 fpart = (weight_sq >> (ipart-8)) & 127;
361 y = ((ipart & 1) ? 32768 : 46214) >> ((32 - ipart)>>1);
362 weight = y + ((213 * fpart * y) >> 16);
363
364 value = cur * 128 + (lsf_res[i] * 16384) / weight;
365 nlsf[i] = av_clip_uintp2(value, 15);
366 }
367
368 /* stabilize the NLSF coefficients */
369 silk_stabilize_lsf(nlsf, order, s->wb ? ff_silk_lsf_min_spacing_wb :
370 ff_silk_lsf_min_spacing_nbmb);
371
372 /* produce an interpolation for the first 2 subframes, */
373 /* and then convert both sets of NLSFs to LPC coefficients */
374 *has_lpc_leadin = 0;
375 if (s->subframes == 4) {
376 int offset = ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_interpolation_offset);
377 if (offset != 4 && frame->coded) {
378 *has_lpc_leadin = 1;
379 if (offset != 0) {
380 int16_t nlsf_leadin[16];
381 for (i = 0; i < order; i++)
382 nlsf_leadin[i] = frame->nlsf[i] +
383 ((nlsf[i] - frame->nlsf[i]) * offset >> 2);
384 silk_lsf2lpc(nlsf_leadin, lpc_leadin, order);
385 } else /* avoid re-computation for a (roughly) 1-in-4 occurrence */
386 memcpy(lpc_leadin, frame->lpc, 16 * sizeof(float));
387 } else
388 offset = 4;
389 s->nlsf_interp_factor = offset;
390
391 silk_lsf2lpc(nlsf, lpc, order);
392 } else {
393 s->nlsf_interp_factor = 4;
394 silk_lsf2lpc(nlsf, lpc, order);
395 }
396
397 memcpy(frame->nlsf, nlsf, order * sizeof(nlsf[0]));
398 memcpy(frame->lpc, lpc, order * sizeof(lpc[0]));
399 }
400
silk_count_children(OpusRangeCoder * rc,int model,int32_t total,int32_t child[2])401 static inline void silk_count_children(OpusRangeCoder *rc, int model, int32_t total,
402 int32_t child[2])
403 {
404 if (total != 0) {
405 child[0] = ff_opus_rc_dec_cdf(rc,
406 ff_silk_model_pulse_location[model] + (((total - 1 + 5) * (total - 1)) >> 1));
407 child[1] = total - child[0];
408 } else {
409 child[0] = 0;
410 child[1] = 0;
411 }
412 }
413
silk_decode_excitation(SilkContext * s,OpusRangeCoder * rc,float * excitationf,int qoffset_high,int active,int voiced)414 static inline void silk_decode_excitation(SilkContext *s, OpusRangeCoder *rc,
415 float* excitationf,
416 int qoffset_high, int active, int voiced)
417 {
418 int i;
419 uint32_t seed;
420 int shellblocks;
421 int ratelevel;
422 uint8_t pulsecount[20]; // total pulses in each shell block
423 uint8_t lsbcount[20] = {0}; // raw lsbits defined for each pulse in each shell block
424 int32_t excitation[320]; // Q23
425
426 /* excitation parameters */
427 seed = ff_opus_rc_dec_cdf(rc, ff_silk_model_lcg_seed);
428 shellblocks = ff_silk_shell_blocks[s->bandwidth][s->subframes >> 2];
429 ratelevel = ff_opus_rc_dec_cdf(rc, ff_silk_model_exc_rate[voiced]);
430
431 for (i = 0; i < shellblocks; i++) {
432 pulsecount[i] = ff_opus_rc_dec_cdf(rc, ff_silk_model_pulse_count[ratelevel]);
433 if (pulsecount[i] == 17) {
434 while (pulsecount[i] == 17 && ++lsbcount[i] != 10)
435 pulsecount[i] = ff_opus_rc_dec_cdf(rc, ff_silk_model_pulse_count[9]);
436 if (lsbcount[i] == 10)
437 pulsecount[i] = ff_opus_rc_dec_cdf(rc, ff_silk_model_pulse_count[10]);
438 }
439 }
440
441 /* decode pulse locations using PVQ */
442 for (i = 0; i < shellblocks; i++) {
443 if (pulsecount[i] != 0) {
444 int a, b, c, d;
445 int32_t * location = excitation + 16*i;
446 int32_t branch[4][2];
447 branch[0][0] = pulsecount[i];
448
449 /* unrolled tail recursion */
450 for (a = 0; a < 1; a++) {
451 silk_count_children(rc, 0, branch[0][a], branch[1]);
452 for (b = 0; b < 2; b++) {
453 silk_count_children(rc, 1, branch[1][b], branch[2]);
454 for (c = 0; c < 2; c++) {
455 silk_count_children(rc, 2, branch[2][c], branch[3]);
456 for (d = 0; d < 2; d++) {
457 silk_count_children(rc, 3, branch[3][d], location);
458 location += 2;
459 }
460 }
461 }
462 }
463 } else
464 memset(excitation + 16*i, 0, 16*sizeof(int32_t));
465 }
466
467 /* decode least significant bits */
468 for (i = 0; i < shellblocks << 4; i++) {
469 int bit;
470 for (bit = 0; bit < lsbcount[i >> 4]; bit++)
471 excitation[i] = (excitation[i] << 1) |
472 ff_opus_rc_dec_cdf(rc, ff_silk_model_excitation_lsb);
473 }
474
475 /* decode signs */
476 for (i = 0; i < shellblocks << 4; i++) {
477 if (excitation[i] != 0) {
478 int sign = ff_opus_rc_dec_cdf(rc, ff_silk_model_excitation_sign[active +
479 voiced][qoffset_high][FFMIN(pulsecount[i >> 4], 6)]);
480 if (sign == 0)
481 excitation[i] *= -1;
482 }
483 }
484
485 /* assemble the excitation */
486 for (i = 0; i < shellblocks << 4; i++) {
487 int value = excitation[i];
488 excitation[i] = value * 256 | ff_silk_quant_offset[voiced][qoffset_high];
489 if (value < 0) excitation[i] += 20;
490 else if (value > 0) excitation[i] -= 20;
491
492 /* invert samples pseudorandomly */
493 seed = 196314165 * seed + 907633515;
494 if (seed & 0x80000000)
495 excitation[i] *= -1;
496 seed += value;
497
498 excitationf[i] = excitation[i] / 8388608.0f;
499 }
500 }
501
502 /** Maximum residual history according to 4.2.7.6.1 */
503 #define SILK_MAX_LAG (288 + LTP_ORDER / 2)
504
505 /** Order of the LTP filter */
506 #define LTP_ORDER 5
507
silk_decode_frame(SilkContext * s,OpusRangeCoder * rc,int frame_num,int channel,int coded_channels,int active,int active1,int redundant)508 static void silk_decode_frame(SilkContext *s, OpusRangeCoder *rc,
509 int frame_num, int channel, int coded_channels,
510 int active, int active1, int redundant)
511 {
512 /* per frame */
513 int voiced; // combines with active to indicate inactive, active, or active+voiced
514 int qoffset_high;
515 int order; // order of the LPC coefficients
516 float lpc_leadin[16], lpc_body[16], residual[SILK_MAX_LAG + SILK_HISTORY];
517 int has_lpc_leadin;
518 float ltpscale;
519
520 /* per subframe */
521 struct {
522 float gain;
523 int pitchlag;
524 float ltptaps[5];
525 } sf[4];
526
527 SilkFrame * const frame = s->frame + channel;
528
529 int i;
530
531 /* obtain stereo weights */
532 if (coded_channels == 2 && channel == 0) {
533 int n, wi[2], ws[2], w[2];
534 n = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s1);
535 wi[0] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s2) + 3 * (n / 5);
536 ws[0] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s3);
537 wi[1] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s2) + 3 * (n % 5);
538 ws[1] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s3);
539
540 for (i = 0; i < 2; i++)
541 w[i] = ff_silk_stereo_weights[wi[i]] +
542 (((ff_silk_stereo_weights[wi[i] + 1] - ff_silk_stereo_weights[wi[i]]) * 6554) >> 16)
543 * (ws[i]*2 + 1);
544
545 s->stereo_weights[0] = (w[0] - w[1]) / 8192.0;
546 s->stereo_weights[1] = w[1] / 8192.0;
547
548 /* and read the mid-only flag */
549 s->midonly = active1 ? 0 : ff_opus_rc_dec_cdf(rc, ff_silk_model_mid_only);
550 }
551
552 /* obtain frame type */
553 if (!active) {
554 qoffset_high = ff_opus_rc_dec_cdf(rc, ff_silk_model_frame_type_inactive);
555 voiced = 0;
556 } else {
557 int type = ff_opus_rc_dec_cdf(rc, ff_silk_model_frame_type_active);
558 qoffset_high = type & 1;
559 voiced = type >> 1;
560 }
561
562 /* obtain subframe quantization gains */
563 for (i = 0; i < s->subframes; i++) {
564 int log_gain; //Q7
565 int ipart, fpart, lingain;
566
567 if (i == 0 && (frame_num == 0 || !frame->coded)) {
568 /* gain is coded absolute */
569 int x = ff_opus_rc_dec_cdf(rc, ff_silk_model_gain_highbits[active + voiced]);
570 log_gain = (x<<3) | ff_opus_rc_dec_cdf(rc, ff_silk_model_gain_lowbits);
571
572 if (frame->coded)
573 log_gain = FFMAX(log_gain, frame->log_gain - 16);
574 } else {
575 /* gain is coded relative */
576 int delta_gain = ff_opus_rc_dec_cdf(rc, ff_silk_model_gain_delta);
577 log_gain = av_clip_uintp2(FFMAX((delta_gain<<1) - 16,
578 frame->log_gain + delta_gain - 4), 6);
579 }
580
581 frame->log_gain = log_gain;
582
583 /* approximate 2**(x/128) with a Q7 (i.e. non-integer) input */
584 log_gain = (log_gain * 0x1D1C71 >> 16) + 2090;
585 ipart = log_gain >> 7;
586 fpart = log_gain & 127;
587 lingain = (1 << ipart) + ((-174 * fpart * (128-fpart) >>16) + fpart) * ((1<<ipart) >> 7);
588 sf[i].gain = lingain / 65536.0f;
589 }
590
591 /* obtain LPC filter coefficients */
592 silk_decode_lpc(s, frame, rc, lpc_leadin, lpc_body, &order, &has_lpc_leadin, voiced);
593
594 /* obtain pitch lags, if this is a voiced frame */
595 if (voiced) {
596 int lag_absolute = (!frame_num || !frame->prev_voiced);
597 int primarylag; // primary pitch lag for the entire SILK frame
598 int ltpfilter;
599 const int8_t * offsets;
600
601 if (!lag_absolute) {
602 int delta = ff_opus_rc_dec_cdf(rc, ff_silk_model_pitch_delta);
603 if (delta)
604 primarylag = frame->primarylag + delta - 9;
605 else
606 lag_absolute = 1;
607 }
608
609 if (lag_absolute) {
610 /* primary lag is coded absolute */
611 int highbits, lowbits;
612 static const uint16_t * const model[] = {
613 ff_silk_model_pitch_lowbits_nb, ff_silk_model_pitch_lowbits_mb,
614 ff_silk_model_pitch_lowbits_wb
615 };
616 highbits = ff_opus_rc_dec_cdf(rc, ff_silk_model_pitch_highbits);
617 lowbits = ff_opus_rc_dec_cdf(rc, model[s->bandwidth]);
618
619 primarylag = ff_silk_pitch_min_lag[s->bandwidth] +
620 highbits*ff_silk_pitch_scale[s->bandwidth] + lowbits;
621 }
622 frame->primarylag = primarylag;
623
624 if (s->subframes == 2)
625 offsets = (s->bandwidth == OPUS_BANDWIDTH_NARROWBAND)
626 ? ff_silk_pitch_offset_nb10ms[ff_opus_rc_dec_cdf(rc,
627 ff_silk_model_pitch_contour_nb10ms)]
628 : ff_silk_pitch_offset_mbwb10ms[ff_opus_rc_dec_cdf(rc,
629 ff_silk_model_pitch_contour_mbwb10ms)];
630 else
631 offsets = (s->bandwidth == OPUS_BANDWIDTH_NARROWBAND)
632 ? ff_silk_pitch_offset_nb20ms[ff_opus_rc_dec_cdf(rc,
633 ff_silk_model_pitch_contour_nb20ms)]
634 : ff_silk_pitch_offset_mbwb20ms[ff_opus_rc_dec_cdf(rc,
635 ff_silk_model_pitch_contour_mbwb20ms)];
636
637 for (i = 0; i < s->subframes; i++)
638 sf[i].pitchlag = av_clip(primarylag + offsets[i],
639 ff_silk_pitch_min_lag[s->bandwidth],
640 ff_silk_pitch_max_lag[s->bandwidth]);
641
642 /* obtain LTP filter coefficients */
643 ltpfilter = ff_opus_rc_dec_cdf(rc, ff_silk_model_ltp_filter);
644 for (i = 0; i < s->subframes; i++) {
645 int index, j;
646 static const uint16_t * const filter_sel[] = {
647 ff_silk_model_ltp_filter0_sel, ff_silk_model_ltp_filter1_sel,
648 ff_silk_model_ltp_filter2_sel
649 };
650 static const int8_t (* const filter_taps[])[5] = {
651 ff_silk_ltp_filter0_taps, ff_silk_ltp_filter1_taps, ff_silk_ltp_filter2_taps
652 };
653 index = ff_opus_rc_dec_cdf(rc, filter_sel[ltpfilter]);
654 for (j = 0; j < 5; j++)
655 sf[i].ltptaps[j] = filter_taps[ltpfilter][index][j] / 128.0f;
656 }
657 }
658
659 /* obtain LTP scale factor */
660 if (voiced && frame_num == 0)
661 ltpscale = ff_silk_ltp_scale_factor[ff_opus_rc_dec_cdf(rc,
662 ff_silk_model_ltp_scale_index)] / 16384.0f;
663 else ltpscale = 15565.0f/16384.0f;
664
665 /* generate the excitation signal for the entire frame */
666 silk_decode_excitation(s, rc, residual + SILK_MAX_LAG, qoffset_high,
667 active, voiced);
668
669 /* skip synthesising the output if we do not need it */
670 // TODO: implement error recovery
671 if (s->output_channels == channel || redundant)
672 return;
673
674 /* generate the output signal */
675 for (i = 0; i < s->subframes; i++) {
676 const float * lpc_coeff = (i < 2 && has_lpc_leadin) ? lpc_leadin : lpc_body;
677 float *dst = frame->output + SILK_HISTORY + i * s->sflength;
678 float *resptr = residual + SILK_MAX_LAG + i * s->sflength;
679 float *lpc = frame->lpc_history + SILK_HISTORY + i * s->sflength;
680 float sum;
681 int j, k;
682
683 if (voiced) {
684 int out_end;
685 float scale;
686
687 if (i < 2 || s->nlsf_interp_factor == 4) {
688 out_end = -i * s->sflength;
689 scale = ltpscale;
690 } else {
691 out_end = -(i - 2) * s->sflength;
692 scale = 1.0f;
693 }
694
695 /* when the LPC coefficients change, a re-whitening filter is used */
696 /* to produce a residual that accounts for the change */
697 for (j = - sf[i].pitchlag - LTP_ORDER/2; j < out_end; j++) {
698 sum = dst[j];
699 for (k = 0; k < order; k++)
700 sum -= lpc_coeff[k] * dst[j - k - 1];
701 resptr[j] = av_clipf(sum, -1.0f, 1.0f) * scale / sf[i].gain;
702 }
703
704 if (out_end) {
705 float rescale = sf[i-1].gain / sf[i].gain;
706 for (j = out_end; j < 0; j++)
707 resptr[j] *= rescale;
708 }
709
710 /* LTP synthesis */
711 for (j = 0; j < s->sflength; j++) {
712 sum = resptr[j];
713 for (k = 0; k < LTP_ORDER; k++)
714 sum += sf[i].ltptaps[k] * resptr[j - sf[i].pitchlag + LTP_ORDER/2 - k];
715 resptr[j] = sum;
716 }
717 }
718
719 /* LPC synthesis */
720 for (j = 0; j < s->sflength; j++) {
721 sum = resptr[j] * sf[i].gain;
722 for (k = 1; k <= order; k++)
723 sum += lpc_coeff[k - 1] * lpc[j - k];
724
725 lpc[j] = sum;
726 dst[j] = av_clipf(sum, -1.0f, 1.0f);
727 }
728 }
729
730 frame->prev_voiced = voiced;
731 memmove(frame->lpc_history, frame->lpc_history + s->flength, SILK_HISTORY * sizeof(float));
732 memmove(frame->output, frame->output + s->flength, SILK_HISTORY * sizeof(float));
733
734 frame->coded = 1;
735 }
736
silk_unmix_ms(SilkContext * s,float * l,float * r)737 static void silk_unmix_ms(SilkContext *s, float *l, float *r)
738 {
739 float *mid = s->frame[0].output + SILK_HISTORY - s->flength;
740 float *side = s->frame[1].output + SILK_HISTORY - s->flength;
741 float w0_prev = s->prev_stereo_weights[0];
742 float w1_prev = s->prev_stereo_weights[1];
743 float w0 = s->stereo_weights[0];
744 float w1 = s->stereo_weights[1];
745 int n1 = ff_silk_stereo_interp_len[s->bandwidth];
746 int i;
747
748 for (i = 0; i < n1; i++) {
749 float interp0 = w0_prev + i * (w0 - w0_prev) / n1;
750 float interp1 = w1_prev + i * (w1 - w1_prev) / n1;
751 float p0 = 0.25 * (mid[i - 2] + 2 * mid[i - 1] + mid[i]);
752
753 l[i] = av_clipf((1 + interp1) * mid[i - 1] + side[i - 1] + interp0 * p0, -1.0, 1.0);
754 r[i] = av_clipf((1 - interp1) * mid[i - 1] - side[i - 1] - interp0 * p0, -1.0, 1.0);
755 }
756
757 for (; i < s->flength; i++) {
758 float p0 = 0.25 * (mid[i - 2] + 2 * mid[i - 1] + mid[i]);
759
760 l[i] = av_clipf((1 + w1) * mid[i - 1] + side[i - 1] + w0 * p0, -1.0, 1.0);
761 r[i] = av_clipf((1 - w1) * mid[i - 1] - side[i - 1] - w0 * p0, -1.0, 1.0);
762 }
763
764 memcpy(s->prev_stereo_weights, s->stereo_weights, sizeof(s->stereo_weights));
765 }
766
silk_flush_frame(SilkFrame * frame)767 static void silk_flush_frame(SilkFrame *frame)
768 {
769 if (!frame->coded)
770 return;
771
772 memset(frame->output, 0, sizeof(frame->output));
773 memset(frame->lpc_history, 0, sizeof(frame->lpc_history));
774
775 memset(frame->lpc, 0, sizeof(frame->lpc));
776 memset(frame->nlsf, 0, sizeof(frame->nlsf));
777
778 frame->log_gain = 0;
779
780 frame->primarylag = 0;
781 frame->prev_voiced = 0;
782 frame->coded = 0;
783 }
784
ff_silk_decode_superframe(SilkContext * s,OpusRangeCoder * rc,float * output[2],enum OpusBandwidth bandwidth,int coded_channels,int duration_ms)785 int ff_silk_decode_superframe(SilkContext *s, OpusRangeCoder *rc,
786 float *output[2],
787 enum OpusBandwidth bandwidth,
788 int coded_channels,
789 int duration_ms)
790 {
791 int active[2][6], redundancy[2];
792 int nb_frames, i, j;
793
794 if (bandwidth > OPUS_BANDWIDTH_WIDEBAND ||
795 coded_channels > 2 || duration_ms > 60) {
796 av_log(s->avctx, AV_LOG_ERROR, "Invalid parameters passed "
797 "to the SILK decoder.\n");
798 return AVERROR(EINVAL);
799 }
800
801 nb_frames = 1 + (duration_ms > 20) + (duration_ms > 40);
802 s->subframes = duration_ms / nb_frames / 5; // 5ms subframes
803 s->sflength = 20 * (bandwidth + 2);
804 s->flength = s->sflength * s->subframes;
805 s->bandwidth = bandwidth;
806 s->wb = bandwidth == OPUS_BANDWIDTH_WIDEBAND;
807
808 /* make sure to flush the side channel when switching from mono to stereo */
809 if (coded_channels > s->prev_coded_channels)
810 silk_flush_frame(&s->frame[1]);
811 s->prev_coded_channels = coded_channels;
812
813 /* read the LP-layer header bits */
814 for (i = 0; i < coded_channels; i++) {
815 for (j = 0; j < nb_frames; j++)
816 active[i][j] = ff_opus_rc_dec_log(rc, 1);
817
818 redundancy[i] = ff_opus_rc_dec_log(rc, 1);
819 }
820
821 /* read the per-frame LBRR flags */
822 for (i = 0; i < coded_channels; i++)
823 if (redundancy[i] && duration_ms > 20) {
824 redundancy[i] = ff_opus_rc_dec_cdf(rc, duration_ms == 40 ?
825 ff_silk_model_lbrr_flags_40 : ff_silk_model_lbrr_flags_60);
826 }
827
828 /* decode the LBRR frames */
829 for (i = 0; i < nb_frames; i++) {
830 for (j = 0; j < coded_channels; j++)
831 if (redundancy[j] & (1 << i)) {
832 int active1 = (j == 0 && !(redundancy[1] & (1 << i))) ? 0 : 1;
833 silk_decode_frame(s, rc, i, j, coded_channels, 1, active1, 1);
834 }
835 }
836
837 for (i = 0; i < nb_frames; i++) {
838 for (j = 0; j < coded_channels && !s->midonly; j++)
839 silk_decode_frame(s, rc, i, j, coded_channels, active[j][i], active[1][i], 0);
840
841 /* reset the side channel if it is not coded */
842 if (s->midonly && s->frame[1].coded)
843 silk_flush_frame(&s->frame[1]);
844
845 if (coded_channels == 1 || s->output_channels == 1) {
846 for (j = 0; j < s->output_channels; j++) {
847 memcpy(output[j] + i * s->flength,
848 s->frame[0].output + SILK_HISTORY - s->flength - 2,
849 s->flength * sizeof(float));
850 }
851 } else {
852 silk_unmix_ms(s, output[0] + i * s->flength, output[1] + i * s->flength);
853 }
854
855 s->midonly = 0;
856 }
857
858 return nb_frames * s->flength;
859 }
860
ff_silk_free(SilkContext ** ps)861 void ff_silk_free(SilkContext **ps)
862 {
863 av_freep(ps);
864 }
865
ff_silk_flush(SilkContext * s)866 void ff_silk_flush(SilkContext *s)
867 {
868 silk_flush_frame(&s->frame[0]);
869 silk_flush_frame(&s->frame[1]);
870
871 memset(s->prev_stereo_weights, 0, sizeof(s->prev_stereo_weights));
872 }
873
ff_silk_init(AVCodecContext * avctx,SilkContext ** ps,int output_channels)874 int ff_silk_init(AVCodecContext *avctx, SilkContext **ps, int output_channels)
875 {
876 SilkContext *s;
877
878 if (output_channels != 1 && output_channels != 2) {
879 av_log(avctx, AV_LOG_ERROR, "Invalid number of output channels: %d\n",
880 output_channels);
881 return AVERROR(EINVAL);
882 }
883
884 s = av_mallocz(sizeof(*s));
885 if (!s)
886 return AVERROR(ENOMEM);
887
888 s->avctx = avctx;
889 s->output_channels = output_channels;
890
891 ff_silk_flush(s);
892
893 *ps = s;
894
895 return 0;
896 }
897