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[],int32_t pol[],int half_order)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
silk_lsf2lpc(const int16_t nlsf[16],float lpcf[16],int order)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
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)305 static inline void silk_decode_lpc(SilkContext *s, SilkFrame *frame,
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] :
322 ff_silk_lsf_s2_model_sel_nbmb[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)
325 lsf_i2[i] -= ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_s2_ext);
326 else if (lsf_i2[i] == 4)
327 lsf_i2[i] += ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_s2_ext);
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) {
340 int weight = s->wb ? ff_silk_lsf_pred_weights_wb [ff_silk_lsf_weight_sel_wb [lsf_i1][i]][i] :
341 ff_silk_lsf_pred_weights_nbmb[ff_silk_lsf_weight_sel_nbmb[lsf_i1][i]][i];
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] :
349 ff_silk_lsf_codebook_nbmb[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 :
371 ff_silk_lsf_min_spacing_nbmb);
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) {
377 int offset = ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_interpolation_offset);
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
silk_count_children(OpusRangeCoder * rc,int model,int32_t total,int32_t child[2])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
silk_decode_excitation(SilkContext * s,OpusRangeCoder * rc,float * excitationf,int qoffset_high,int active,int voiced)415 static inline void silk_decode_excitation(SilkContext *s, OpusRangeCoder *rc,
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 */
428 seed = ff_opus_rc_dec_cdf(rc, ff_silk_model_lcg_seed);
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) |
473 ff_opus_rc_dec_cdf(rc, ff_silk_model_excitation_lsb);
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
silk_decode_frame(SilkContext * s,OpusRangeCoder * rc,int frame_num,int channel,int coded_channels,int active,int active1,int redundant)509 static void silk_decode_frame(SilkContext *s, OpusRangeCoder *rc,
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];
535 n = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s1);
536 wi[0] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s2) + 3 * (n / 5);
537 ws[0] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s3);
538 wi[1] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s2) + 3 * (n % 5);
539 ws[1] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s3);
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) {
555 qoffset_high = ff_opus_rc_dec_cdf(rc, ff_silk_model_frame_type_inactive);
556 voiced = 0;
557 } else {
558 int type = ff_opus_rc_dec_cdf(rc, ff_silk_model_frame_type_active);
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) {
603 int delta = ff_opus_rc_dec_cdf(rc, ff_silk_model_pitch_delta);
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[] = {
614 ff_silk_model_pitch_lowbits_nb, ff_silk_model_pitch_lowbits_mb,
615 ff_silk_model_pitch_lowbits_wb
616 };
617 highbits = ff_opus_rc_dec_cdf(rc, ff_silk_model_pitch_highbits);
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)
627 ? ff_silk_pitch_offset_nb10ms[ff_opus_rc_dec_cdf(rc,
628 ff_silk_model_pitch_contour_nb10ms)]
629 : ff_silk_pitch_offset_mbwb10ms[ff_opus_rc_dec_cdf(rc,
630 ff_silk_model_pitch_contour_mbwb10ms)];
631 else
632 offsets = (s->bandwidth == OPUS_BANDWIDTH_NARROWBAND)
633 ? ff_silk_pitch_offset_nb20ms[ff_opus_rc_dec_cdf(rc,
634 ff_silk_model_pitch_contour_nb20ms)]
635 : ff_silk_pitch_offset_mbwb20ms[ff_opus_rc_dec_cdf(rc,
636 ff_silk_model_pitch_contour_mbwb20ms)];
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 */
644 ltpfilter = ff_opus_rc_dec_cdf(rc, ff_silk_model_ltp_filter);
645 for (i = 0; i < s->subframes; i++) {
646 int index, j;
647 static const uint16_t * const filter_sel[] = {
648 ff_silk_model_ltp_filter0_sel, ff_silk_model_ltp_filter1_sel,
649 ff_silk_model_ltp_filter2_sel
650 };
651 static const int8_t (* const filter_taps[])[5] = {
652 ff_silk_ltp_filter0_taps, ff_silk_ltp_filter1_taps, ff_silk_ltp_filter2_taps
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)
662 ltpscale = ff_silk_ltp_scale_factor[ff_opus_rc_dec_cdf(rc,
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
silk_unmix_ms(SilkContext * s,float * l,float * r)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
silk_flush_frame(SilkFrame * frame)768 static void silk_flush_frame(SilkFrame *frame)
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
ff_silk_decode_superframe(SilkContext * s,OpusRangeCoder * rc,float * output[2],enum OpusBandwidth bandwidth,int coded_channels,int duration_ms)786 int ff_silk_decode_superframe(SilkContext *s, OpusRangeCoder *rc,
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 ?
826 ff_silk_model_lbrr_flags_40 : ff_silk_model_lbrr_flags_60);
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
ff_silk_free(SilkContext ** ps)862 void ff_silk_free(SilkContext **ps)
863 {
864 av_freep(ps);
865 }
866
ff_silk_flush(SilkContext * s)867 void ff_silk_flush(SilkContext *s)
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
ff_silk_init(AVCodecContext * avctx,SilkContext ** ps,int output_channels)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 }
898