2 * Copyright (c) 2012 Andrew D'Addesio
3 * Copyright (c) 2013-2014 Mozilla Corporation
5 * This file is part of FFmpeg.
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.
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.
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
32 typedef struct SilkFrame {
38 float output [2 * SILK_HISTORY];
39 float lpc_history[2 * SILK_HISTORY];
46 AVCodecContext *avctx;
53 int nlsf_interp_factor;
55 enum OpusBandwidth bandwidth;
59 float prev_stereo_weights[2];
60 float stereo_weights[2];
62 int prev_coded_channels;
65 static inline void silk_stabilize_lsf(int16_t nlsf[16], int order, const uint16_t min_delta[17])
68 for (pass = 0; pass < 20; pass++) {
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]);
75 if (diff < min_diff) {
83 if (min_diff == 0) /* no issues; stabilized */
86 /* wiggle one or two LSFs */
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];
94 /* repel away from current position */
95 int min_center = 0, max_center = 32768, center_val;
98 for (i = 0; i < k; i++)
99 min_center += min_delta[i];
100 min_center += min_delta[k] >> 1;
103 for (i = order; i > k; i--)
104 max_center -= min_delta[i];
105 max_center -= min_delta[k] >> 1;
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));
112 nlsf[k - 1] = center_val - (min_delta[k] >> 1);
113 nlsf[k] = nlsf[k - 1] + min_delta[k];
117 /* resort to the fall-back method, the standard method for LSF stabilization */
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];
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));
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];
143 static inline int silk_is_lpc_stable(const int16_t lpc[16], int order)
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;
150 /* initialize the first row for the Levinson recursion */
151 for (k = 0; k < order; k++) {
153 row[k] = lpc[k] * 4096;
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
167 if (FFABS(row[k]) > 16773022)
170 rc = -(row[k] * 128);
171 gaindiv = (1 << 30) - MULH(rc, rc);
173 totalinvgain = MULH(totalinvgain, gaindiv) << 2;
175 return (totalinvgain >= 107374);
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));
183 /* switch to the next row of the LPC coefficients */
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);
191 /* per RFC 8251 section 6, if this calculation overflows, the filter
192 is considered unstable. */
193 if (tmp < INT32_MIN || tmp > INT32_MAX)
196 row[j] = (int32_t)tmp;
201 static void silk_lsp2poly(const int32_t lsp[16], int32_t pol[16], int half_order)
205 pol[0] = 65536; // 1.0 in Q16
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);
213 pol[1] -= lsp[2 * i];
217 static void silk_lsf2lpc(const int16_t nlsf[16], float lpcf[16], int order)
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
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];
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;
237 silk_lsp2poly(lsp , p, order >> 1);
238 silk_lsp2poly(lsp + 1, q, order >> 1);
240 /* reconstruct A(z) */
241 for (k = 0; k < order>>1; k++) {
242 lpc32[k] = -p[k + 1] - p[k] - q[k + 1] + q[k];
243 lpc32[order-k-1] = -p[k + 1] - p[k] + q[k + 1] - q[k];
246 /* limit the range of the LPC coefficients to each fit within an int16_t */
247 for (i = 0; i < 10; i++) {
249 unsigned int maxabs = 0;
250 for (j = 0, k = 0; j < order; j++) {
251 unsigned int x = FFABS(lpc32[k]);
258 maxabs = (maxabs + 16) >> 5; // convert to Q12
260 if (maxabs > 32767) {
261 /* perform bandwidth expansion */
262 unsigned int chirp, chirp_base; // Q16
263 maxabs = FFMIN(maxabs, 163838); // anything above this overflows chirp's numerator
264 chirp_base = chirp = 65470 - ((maxabs - 32767) << 14) / ((maxabs * (k+1)) >> 2);
266 for (k = 0; k < order; k++) {
267 lpc32[k] = ROUND_MULL(lpc32[k], chirp, 16);
268 chirp = (chirp_base * chirp + 32768) >> 16;
274 /* time's up: just clamp */
275 for (k = 0; k < order; k++) {
276 int x = (lpc32[k] + 16) >> 5;
277 lpc[k] = av_clip_int16(x);
278 lpc32[k] = lpc[k] << 5; // shortcut mandated by the spec; drops lower 5 bits
281 for (k = 0; k < order; k++)
282 lpc[k] = (lpc32[k] + 16) >> 5;
285 /* if the prediction gain causes the LPC filter to become unstable,
286 apply further bandwidth expansion on the Q17 coefficients */
287 for (i = 1; i <= 16 && !silk_is_lpc_stable(lpc, order); i++) {
288 unsigned int chirp, chirp_base;
289 chirp_base = chirp = 65536 - (1 << i);
291 for (k = 0; k < order; k++) {
292 lpc32[k] = ROUND_MULL(lpc32[k], chirp, 16);
293 lpc[k] = (lpc32[k] + 16) >> 5;
294 chirp = (chirp_base * chirp + 32768) >> 16;
298 for (i = 0; i < order; i++)
299 lpcf[i] = lpc[i] / 4096.0f;
302 static inline void silk_decode_lpc(SilkContext *s, SilkFrame *frame,
304 float lpc_leadin[16], float lpc[16],
305 int *lpc_order, int *has_lpc_leadin, int voiced)
308 int order; // order of the LP polynomial; 10 for NB/MB and 16 for WB
309 int8_t lsf_i1, lsf_i2[16]; // stage-1 and stage-2 codebook indices
310 int16_t lsf_res[16]; // residual as a Q10 value
311 int16_t nlsf[16]; // Q15
313 *lpc_order = order = s->wb ? 16 : 10;
315 /* obtain LSF stage-1 and stage-2 indices */
316 lsf_i1 = ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_s1[s->wb][voiced]);
317 for (i = 0; i < order; i++) {
318 int index = s->wb ? ff_silk_lsf_s2_model_sel_wb [lsf_i1][i] :
319 ff_silk_lsf_s2_model_sel_nbmb[lsf_i1][i];
320 lsf_i2[i] = ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_s2[index]) - 4;
322 lsf_i2[i] -= ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_s2_ext);
323 else if (lsf_i2[i] == 4)
324 lsf_i2[i] += ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_s2_ext);
327 /* reverse the backwards-prediction step */
328 for (i = order - 1; i >= 0; i--) {
329 int qstep = s->wb ? 9830 : 11796;
331 lsf_res[i] = lsf_i2[i] * 1024;
332 if (lsf_i2[i] < 0) lsf_res[i] += 102;
333 else if (lsf_i2[i] > 0) lsf_res[i] -= 102;
334 lsf_res[i] = (lsf_res[i] * qstep) >> 16;
337 int weight = s->wb ? ff_silk_lsf_pred_weights_wb [ff_silk_lsf_weight_sel_wb [lsf_i1][i]][i] :
338 ff_silk_lsf_pred_weights_nbmb[ff_silk_lsf_weight_sel_nbmb[lsf_i1][i]][i];
339 lsf_res[i] += (lsf_res[i+1] * weight) >> 8;
343 /* reconstruct the NLSF coefficients from the supplied indices */
344 for (i = 0; i < order; i++) {
345 const uint8_t * codebook = s->wb ? ff_silk_lsf_codebook_wb [lsf_i1] :
346 ff_silk_lsf_codebook_nbmb[lsf_i1];
347 int cur, prev, next, weight_sq, weight, ipart, fpart, y, value;
349 /* find the weight of the residual */
350 /* TODO: precompute */
352 prev = i ? codebook[i - 1] : 0;
353 next = i + 1 < order ? codebook[i + 1] : 256;
354 weight_sq = (1024 / (cur - prev) + 1024 / (next - cur)) << 16;
356 /* approximate square-root with mandated fixed-point arithmetic */
357 ipart = opus_ilog(weight_sq);
358 fpart = (weight_sq >> (ipart-8)) & 127;
359 y = ((ipart & 1) ? 32768 : 46214) >> ((32 - ipart)>>1);
360 weight = y + ((213 * fpart * y) >> 16);
362 value = cur * 128 + (lsf_res[i] * 16384) / weight;
363 nlsf[i] = av_clip_uintp2(value, 15);
366 /* stabilize the NLSF coefficients */
367 silk_stabilize_lsf(nlsf, order, s->wb ? ff_silk_lsf_min_spacing_wb :
368 ff_silk_lsf_min_spacing_nbmb);
370 /* produce an interpolation for the first 2 subframes, */
371 /* and then convert both sets of NLSFs to LPC coefficients */
373 if (s->subframes == 4) {
374 int offset = ff_opus_rc_dec_cdf(rc, ff_silk_model_lsf_interpolation_offset);
375 if (offset != 4 && frame->coded) {
378 int16_t nlsf_leadin[16];
379 for (i = 0; i < order; i++)
380 nlsf_leadin[i] = frame->nlsf[i] +
381 ((nlsf[i] - frame->nlsf[i]) * offset >> 2);
382 silk_lsf2lpc(nlsf_leadin, lpc_leadin, order);
383 } else /* avoid re-computation for a (roughly) 1-in-4 occurrence */
384 memcpy(lpc_leadin, frame->lpc, 16 * sizeof(float));
387 s->nlsf_interp_factor = offset;
389 silk_lsf2lpc(nlsf, lpc, order);
391 s->nlsf_interp_factor = 4;
392 silk_lsf2lpc(nlsf, lpc, order);
395 memcpy(frame->nlsf, nlsf, order * sizeof(nlsf[0]));
396 memcpy(frame->lpc, lpc, order * sizeof(lpc[0]));
399 static inline void silk_count_children(OpusRangeCoder *rc, int model, int32_t total,
403 child[0] = ff_opus_rc_dec_cdf(rc,
404 ff_silk_model_pulse_location[model] + (((total - 1 + 5) * (total - 1)) >> 1));
405 child[1] = total - child[0];
412 static inline void silk_decode_excitation(SilkContext *s, OpusRangeCoder *rc,
414 int qoffset_high, int active, int voiced)
420 uint8_t pulsecount[20]; // total pulses in each shell block
421 uint8_t lsbcount[20] = {0}; // raw lsbits defined for each pulse in each shell block
422 int32_t excitation[320]; // Q23
424 /* excitation parameters */
425 seed = ff_opus_rc_dec_cdf(rc, ff_silk_model_lcg_seed);
426 shellblocks = ff_silk_shell_blocks[s->bandwidth][s->subframes >> 2];
427 ratelevel = ff_opus_rc_dec_cdf(rc, ff_silk_model_exc_rate[voiced]);
429 for (i = 0; i < shellblocks; i++) {
430 pulsecount[i] = ff_opus_rc_dec_cdf(rc, ff_silk_model_pulse_count[ratelevel]);
431 if (pulsecount[i] == 17) {
432 while (pulsecount[i] == 17 && ++lsbcount[i] != 10)
433 pulsecount[i] = ff_opus_rc_dec_cdf(rc, ff_silk_model_pulse_count[9]);
434 if (lsbcount[i] == 10)
435 pulsecount[i] = ff_opus_rc_dec_cdf(rc, ff_silk_model_pulse_count[10]);
439 /* decode pulse locations using PVQ */
440 for (i = 0; i < shellblocks; i++) {
441 if (pulsecount[i] != 0) {
443 int32_t * location = excitation + 16*i;
444 int32_t branch[4][2];
445 branch[0][0] = pulsecount[i];
447 /* unrolled tail recursion */
448 for (a = 0; a < 1; a++) {
449 silk_count_children(rc, 0, branch[0][a], branch[1]);
450 for (b = 0; b < 2; b++) {
451 silk_count_children(rc, 1, branch[1][b], branch[2]);
452 for (c = 0; c < 2; c++) {
453 silk_count_children(rc, 2, branch[2][c], branch[3]);
454 for (d = 0; d < 2; d++) {
455 silk_count_children(rc, 3, branch[3][d], location);
462 memset(excitation + 16*i, 0, 16*sizeof(int32_t));
465 /* decode least significant bits */
466 for (i = 0; i < shellblocks << 4; i++) {
468 for (bit = 0; bit < lsbcount[i >> 4]; bit++)
469 excitation[i] = (excitation[i] << 1) |
470 ff_opus_rc_dec_cdf(rc, ff_silk_model_excitation_lsb);
474 for (i = 0; i < shellblocks << 4; i++) {
475 if (excitation[i] != 0) {
476 int sign = ff_opus_rc_dec_cdf(rc, ff_silk_model_excitation_sign[active +
477 voiced][qoffset_high][FFMIN(pulsecount[i >> 4], 6)]);
483 /* assemble the excitation */
484 for (i = 0; i < shellblocks << 4; i++) {
485 int value = excitation[i];
486 excitation[i] = value * 256 | ff_silk_quant_offset[voiced][qoffset_high];
487 if (value < 0) excitation[i] += 20;
488 else if (value > 0) excitation[i] -= 20;
490 /* invert samples pseudorandomly */
491 seed = 196314165 * seed + 907633515;
492 if (seed & 0x80000000)
496 excitationf[i] = excitation[i] / 8388608.0f;
500 /** Maximum residual history according to 4.2.7.6.1 */
501 #define SILK_MAX_LAG (288 + LTP_ORDER / 2)
503 /** Order of the LTP filter */
506 static void silk_decode_frame(SilkContext *s, OpusRangeCoder *rc,
507 int frame_num, int channel, int coded_channels, int active, int active1)
510 int voiced; // combines with active to indicate inactive, active, or active+voiced
512 int order; // order of the LPC coefficients
513 float lpc_leadin[16], lpc_body[16], residual[SILK_MAX_LAG + SILK_HISTORY];
524 SilkFrame * const frame = s->frame + channel;
528 /* obtain stereo weights */
529 if (coded_channels == 2 && channel == 0) {
530 int n, wi[2], ws[2], w[2];
531 n = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s1);
532 wi[0] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s2) + 3 * (n / 5);
533 ws[0] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s3);
534 wi[1] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s2) + 3 * (n % 5);
535 ws[1] = ff_opus_rc_dec_cdf(rc, ff_silk_model_stereo_s3);
537 for (i = 0; i < 2; i++)
538 w[i] = ff_silk_stereo_weights[wi[i]] +
539 (((ff_silk_stereo_weights[wi[i] + 1] - ff_silk_stereo_weights[wi[i]]) * 6554) >> 16)
542 s->stereo_weights[0] = (w[0] - w[1]) / 8192.0;
543 s->stereo_weights[1] = w[1] / 8192.0;
545 /* and read the mid-only flag */
546 s->midonly = active1 ? 0 : ff_opus_rc_dec_cdf(rc, ff_silk_model_mid_only);
549 /* obtain frame type */
551 qoffset_high = ff_opus_rc_dec_cdf(rc, ff_silk_model_frame_type_inactive);
554 int type = ff_opus_rc_dec_cdf(rc, ff_silk_model_frame_type_active);
555 qoffset_high = type & 1;
559 /* obtain subframe quantization gains */
560 for (i = 0; i < s->subframes; i++) {
562 int ipart, fpart, lingain;
564 if (i == 0 && (frame_num == 0 || !frame->coded)) {
565 /* gain is coded absolute */
566 int x = ff_opus_rc_dec_cdf(rc, ff_silk_model_gain_highbits[active + voiced]);
567 log_gain = (x<<3) | ff_opus_rc_dec_cdf(rc, ff_silk_model_gain_lowbits);
570 log_gain = FFMAX(log_gain, frame->log_gain - 16);
572 /* gain is coded relative */
573 int delta_gain = ff_opus_rc_dec_cdf(rc, ff_silk_model_gain_delta);
574 log_gain = av_clip_uintp2(FFMAX((delta_gain<<1) - 16,
575 frame->log_gain + delta_gain - 4), 6);
578 frame->log_gain = log_gain;
580 /* approximate 2**(x/128) with a Q7 (i.e. non-integer) input */
581 log_gain = (log_gain * 0x1D1C71 >> 16) + 2090;
582 ipart = log_gain >> 7;
583 fpart = log_gain & 127;
584 lingain = (1 << ipart) + ((-174 * fpart * (128-fpart) >>16) + fpart) * ((1<<ipart) >> 7);
585 sf[i].gain = lingain / 65536.0f;
588 /* obtain LPC filter coefficients */
589 silk_decode_lpc(s, frame, rc, lpc_leadin, lpc_body, &order, &has_lpc_leadin, voiced);
591 /* obtain pitch lags, if this is a voiced frame */
593 int lag_absolute = (!frame_num || !frame->prev_voiced);
594 int primarylag; // primary pitch lag for the entire SILK frame
596 const int8_t * offsets;
599 int delta = ff_opus_rc_dec_cdf(rc, ff_silk_model_pitch_delta);
601 primarylag = frame->primarylag + delta - 9;
607 /* primary lag is coded absolute */
608 int highbits, lowbits;
609 static const uint16_t * const model[] = {
610 ff_silk_model_pitch_lowbits_nb, ff_silk_model_pitch_lowbits_mb,
611 ff_silk_model_pitch_lowbits_wb
613 highbits = ff_opus_rc_dec_cdf(rc, ff_silk_model_pitch_highbits);
614 lowbits = ff_opus_rc_dec_cdf(rc, model[s->bandwidth]);
616 primarylag = ff_silk_pitch_min_lag[s->bandwidth] +
617 highbits*ff_silk_pitch_scale[s->bandwidth] + lowbits;
619 frame->primarylag = primarylag;
621 if (s->subframes == 2)
622 offsets = (s->bandwidth == OPUS_BANDWIDTH_NARROWBAND)
623 ? ff_silk_pitch_offset_nb10ms[ff_opus_rc_dec_cdf(rc,
624 ff_silk_model_pitch_contour_nb10ms)]
625 : ff_silk_pitch_offset_mbwb10ms[ff_opus_rc_dec_cdf(rc,
626 ff_silk_model_pitch_contour_mbwb10ms)];
628 offsets = (s->bandwidth == OPUS_BANDWIDTH_NARROWBAND)
629 ? ff_silk_pitch_offset_nb20ms[ff_opus_rc_dec_cdf(rc,
630 ff_silk_model_pitch_contour_nb20ms)]
631 : ff_silk_pitch_offset_mbwb20ms[ff_opus_rc_dec_cdf(rc,
632 ff_silk_model_pitch_contour_mbwb20ms)];
634 for (i = 0; i < s->subframes; i++)
635 sf[i].pitchlag = av_clip(primarylag + offsets[i],
636 ff_silk_pitch_min_lag[s->bandwidth],
637 ff_silk_pitch_max_lag[s->bandwidth]);
639 /* obtain LTP filter coefficients */
640 ltpfilter = ff_opus_rc_dec_cdf(rc, ff_silk_model_ltp_filter);
641 for (i = 0; i < s->subframes; i++) {
643 static const uint16_t * const filter_sel[] = {
644 ff_silk_model_ltp_filter0_sel, ff_silk_model_ltp_filter1_sel,
645 ff_silk_model_ltp_filter2_sel
647 static const int8_t (* const filter_taps[])[5] = {
648 ff_silk_ltp_filter0_taps, ff_silk_ltp_filter1_taps, ff_silk_ltp_filter2_taps
650 index = ff_opus_rc_dec_cdf(rc, filter_sel[ltpfilter]);
651 for (j = 0; j < 5; j++)
652 sf[i].ltptaps[j] = filter_taps[ltpfilter][index][j] / 128.0f;
656 /* obtain LTP scale factor */
657 if (voiced && frame_num == 0)
658 ltpscale = ff_silk_ltp_scale_factor[ff_opus_rc_dec_cdf(rc,
659 ff_silk_model_ltp_scale_index)] / 16384.0f;
660 else ltpscale = 15565.0f/16384.0f;
662 /* generate the excitation signal for the entire frame */
663 silk_decode_excitation(s, rc, residual + SILK_MAX_LAG, qoffset_high,
666 /* skip synthesising the side channel if we want mono-only */
667 if (s->output_channels == channel)
670 /* generate the output signal */
671 for (i = 0; i < s->subframes; i++) {
672 const float * lpc_coeff = (i < 2 && has_lpc_leadin) ? lpc_leadin : lpc_body;
673 float *dst = frame->output + SILK_HISTORY + i * s->sflength;
674 float *resptr = residual + SILK_MAX_LAG + i * s->sflength;
675 float *lpc = frame->lpc_history + SILK_HISTORY + i * s->sflength;
683 if (i < 2 || s->nlsf_interp_factor == 4) {
684 out_end = -i * s->sflength;
687 out_end = -(i - 2) * s->sflength;
691 /* when the LPC coefficients change, a re-whitening filter is used */
692 /* to produce a residual that accounts for the change */
693 for (j = - sf[i].pitchlag - LTP_ORDER/2; j < out_end; j++) {
695 for (k = 0; k < order; k++)
696 sum -= lpc_coeff[k] * dst[j - k - 1];
697 resptr[j] = av_clipf(sum, -1.0f, 1.0f) * scale / sf[i].gain;
701 float rescale = sf[i-1].gain / sf[i].gain;
702 for (j = out_end; j < 0; j++)
703 resptr[j] *= rescale;
707 for (j = 0; j < s->sflength; j++) {
709 for (k = 0; k < LTP_ORDER; k++)
710 sum += sf[i].ltptaps[k] * resptr[j - sf[i].pitchlag + LTP_ORDER/2 - k];
716 for (j = 0; j < s->sflength; j++) {
717 sum = resptr[j] * sf[i].gain;
718 for (k = 1; k <= order; k++)
719 sum += lpc_coeff[k - 1] * lpc[j - k];
722 dst[j] = av_clipf(sum, -1.0f, 1.0f);
726 frame->prev_voiced = voiced;
727 memmove(frame->lpc_history, frame->lpc_history + s->flength, SILK_HISTORY * sizeof(float));
728 memmove(frame->output, frame->output + s->flength, SILK_HISTORY * sizeof(float));
733 static void silk_unmix_ms(SilkContext *s, float *l, float *r)
735 float *mid = s->frame[0].output + SILK_HISTORY - s->flength;
736 float *side = s->frame[1].output + SILK_HISTORY - s->flength;
737 float w0_prev = s->prev_stereo_weights[0];
738 float w1_prev = s->prev_stereo_weights[1];
739 float w0 = s->stereo_weights[0];
740 float w1 = s->stereo_weights[1];
741 int n1 = ff_silk_stereo_interp_len[s->bandwidth];
744 for (i = 0; i < n1; i++) {
745 float interp0 = w0_prev + i * (w0 - w0_prev) / n1;
746 float interp1 = w1_prev + i * (w1 - w1_prev) / n1;
747 float p0 = 0.25 * (mid[i - 2] + 2 * mid[i - 1] + mid[i]);
749 l[i] = av_clipf((1 + interp1) * mid[i - 1] + side[i - 1] + interp0 * p0, -1.0, 1.0);
750 r[i] = av_clipf((1 - interp1) * mid[i - 1] - side[i - 1] - interp0 * p0, -1.0, 1.0);
753 for (; i < s->flength; i++) {
754 float p0 = 0.25 * (mid[i - 2] + 2 * mid[i - 1] + mid[i]);
756 l[i] = av_clipf((1 + w1) * mid[i - 1] + side[i - 1] + w0 * p0, -1.0, 1.0);
757 r[i] = av_clipf((1 - w1) * mid[i - 1] - side[i - 1] - w0 * p0, -1.0, 1.0);
760 memcpy(s->prev_stereo_weights, s->stereo_weights, sizeof(s->stereo_weights));
763 static void silk_flush_frame(SilkFrame *frame)
768 memset(frame->output, 0, sizeof(frame->output));
769 memset(frame->lpc_history, 0, sizeof(frame->lpc_history));
771 memset(frame->lpc, 0, sizeof(frame->lpc));
772 memset(frame->nlsf, 0, sizeof(frame->nlsf));
776 frame->primarylag = 0;
777 frame->prev_voiced = 0;
781 int ff_silk_decode_superframe(SilkContext *s, OpusRangeCoder *rc,
783 enum OpusBandwidth bandwidth,
787 int active[2][6], redundancy[2];
790 if (bandwidth > OPUS_BANDWIDTH_WIDEBAND ||
791 coded_channels > 2 || duration_ms > 60) {
792 av_log(s->avctx, AV_LOG_ERROR, "Invalid parameters passed "
793 "to the SILK decoder.\n");
794 return AVERROR(EINVAL);
797 nb_frames = 1 + (duration_ms > 20) + (duration_ms > 40);
798 s->subframes = duration_ms / nb_frames / 5; // 5ms subframes
799 s->sflength = 20 * (bandwidth + 2);
800 s->flength = s->sflength * s->subframes;
801 s->bandwidth = bandwidth;
802 s->wb = bandwidth == OPUS_BANDWIDTH_WIDEBAND;
804 /* make sure to flush the side channel when switching from mono to stereo */
805 if (coded_channels > s->prev_coded_channels)
806 silk_flush_frame(&s->frame[1]);
807 s->prev_coded_channels = coded_channels;
809 /* read the LP-layer header bits */
810 for (i = 0; i < coded_channels; i++) {
811 for (j = 0; j < nb_frames; j++)
812 active[i][j] = ff_opus_rc_dec_log(rc, 1);
814 redundancy[i] = ff_opus_rc_dec_log(rc, 1);
816 avpriv_report_missing_feature(s->avctx, "LBRR frames");
817 return AVERROR_PATCHWELCOME;
821 for (i = 0; i < nb_frames; i++) {
822 for (j = 0; j < coded_channels && !s->midonly; j++)
823 silk_decode_frame(s, rc, i, j, coded_channels, active[j][i], active[1][i]);
825 /* reset the side channel if it is not coded */
826 if (s->midonly && s->frame[1].coded)
827 silk_flush_frame(&s->frame[1]);
829 if (coded_channels == 1 || s->output_channels == 1) {
830 for (j = 0; j < s->output_channels; j++) {
831 memcpy(output[j] + i * s->flength,
832 s->frame[0].output + SILK_HISTORY - s->flength - 2,
833 s->flength * sizeof(float));
836 silk_unmix_ms(s, output[0] + i * s->flength, output[1] + i * s->flength);
842 return nb_frames * s->flength;
845 void ff_silk_free(SilkContext **ps)
850 void ff_silk_flush(SilkContext *s)
852 silk_flush_frame(&s->frame[0]);
853 silk_flush_frame(&s->frame[1]);
855 memset(s->prev_stereo_weights, 0, sizeof(s->prev_stereo_weights));
858 int ff_silk_init(AVCodecContext *avctx, SilkContext **ps, int output_channels)
862 if (output_channels != 1 && output_channels != 2) {
863 av_log(avctx, AV_LOG_ERROR, "Invalid number of output channels: %d\n",
865 return AVERROR(EINVAL);
868 s = av_mallocz(sizeof(*s));
870 return AVERROR(ENOMEM);
873 s->output_channels = output_channels;