1 /*
2 * Copyright (c) 2012 Andrew D'Addesio
3 * Copyright (c) 2013-2014 Mozilla Corporation
4 * Copyright (c) 2016 Rostislav Pehlivanov <atomnuker@gmail.com>
5 *
6 * This file is part of FFmpeg.
7 *
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22
23 /**
24 * @file
25 * Opus CELT decoder
26 */
27
28 #include "opus_celt.h"
29 #include "opustab.h"
30 #include "opus_pvq.h"
31
32 /* Use the 2D z-transform to apply prediction in both the time domain (alpha)
33 * and the frequency domain (beta) */
celt_decode_coarse_energy(CeltFrame * f,OpusRangeCoder * rc)34 static void celt_decode_coarse_energy(CeltFrame *f, OpusRangeCoder *rc)
35 {
36 int i, j;
37 float prev[2] = { 0 };
38 float alpha = ff_celt_alpha_coef[f->size];
39 float beta = ff_celt_beta_coef[f->size];
40 const uint8_t *model = ff_celt_coarse_energy_dist[f->size][0];
41
42 /* intra frame */
43 if (opus_rc_tell(rc) + 3 <= f->framebits && ff_opus_rc_dec_log(rc, 3)) {
44 alpha = 0.0f;
45 beta = 1.0f - (4915.0f/32768.0f);
46 model = ff_celt_coarse_energy_dist[f->size][1];
47 }
48
49 for (i = 0; i < CELT_MAX_BANDS; i++) {
50 for (j = 0; j < f->channels; j++) {
51 CeltBlock *block = &f->block[j];
52 float value;
53 int available;
54
55 if (i < f->start_band || i >= f->end_band) {
56 block->energy[i] = 0.0;
57 continue;
58 }
59
60 available = f->framebits - opus_rc_tell(rc);
61 if (available >= 15) {
62 /* decode using a Laplace distribution */
63 int k = FFMIN(i, 20) << 1;
64 value = ff_opus_rc_dec_laplace(rc, model[k] << 7, model[k+1] << 6);
65 } else if (available >= 2) {
66 int x = ff_opus_rc_dec_cdf(rc, ff_celt_model_energy_small);
67 value = (x>>1) ^ -(x&1);
68 } else if (available >= 1) {
69 value = -(float)ff_opus_rc_dec_log(rc, 1);
70 } else value = -1;
71
72 block->energy[i] = FFMAX(-9.0f, block->energy[i]) * alpha + prev[j] + value;
73 prev[j] += beta * value;
74 }
75 }
76 }
77
celt_decode_fine_energy(CeltFrame * f,OpusRangeCoder * rc)78 static void celt_decode_fine_energy(CeltFrame *f, OpusRangeCoder *rc)
79 {
80 int i;
81 for (i = f->start_band; i < f->end_band; i++) {
82 int j;
83 if (!f->fine_bits[i])
84 continue;
85
86 for (j = 0; j < f->channels; j++) {
87 CeltBlock *block = &f->block[j];
88 int q2;
89 float offset;
90 q2 = ff_opus_rc_get_raw(rc, f->fine_bits[i]);
91 offset = (q2 + 0.5f) * (1 << (14 - f->fine_bits[i])) / 16384.0f - 0.5f;
92 block->energy[i] += offset;
93 }
94 }
95 }
96
celt_decode_final_energy(CeltFrame * f,OpusRangeCoder * rc)97 static void celt_decode_final_energy(CeltFrame *f, OpusRangeCoder *rc)
98 {
99 int priority, i, j;
100 int bits_left = f->framebits - opus_rc_tell(rc);
101
102 for (priority = 0; priority < 2; priority++) {
103 for (i = f->start_band; i < f->end_band && bits_left >= f->channels; i++) {
104 if (f->fine_priority[i] != priority || f->fine_bits[i] >= CELT_MAX_FINE_BITS)
105 continue;
106
107 for (j = 0; j < f->channels; j++) {
108 int q2;
109 float offset;
110 q2 = ff_opus_rc_get_raw(rc, 1);
111 offset = (q2 - 0.5f) * (1 << (14 - f->fine_bits[i] - 1)) / 16384.0f;
112 f->block[j].energy[i] += offset;
113 bits_left--;
114 }
115 }
116 }
117 }
118
celt_decode_tf_changes(CeltFrame * f,OpusRangeCoder * rc)119 static void celt_decode_tf_changes(CeltFrame *f, OpusRangeCoder *rc)
120 {
121 int i, diff = 0, tf_select = 0, tf_changed = 0, tf_select_bit;
122 int consumed, bits = f->transient ? 2 : 4;
123
124 consumed = opus_rc_tell(rc);
125 tf_select_bit = (f->size != 0 && consumed+bits+1 <= f->framebits);
126
127 for (i = f->start_band; i < f->end_band; i++) {
128 if (consumed+bits+tf_select_bit <= f->framebits) {
129 diff ^= ff_opus_rc_dec_log(rc, bits);
130 consumed = opus_rc_tell(rc);
131 tf_changed |= diff;
132 }
133 f->tf_change[i] = diff;
134 bits = f->transient ? 4 : 5;
135 }
136
137 if (tf_select_bit && ff_celt_tf_select[f->size][f->transient][0][tf_changed] !=
138 ff_celt_tf_select[f->size][f->transient][1][tf_changed])
139 tf_select = ff_opus_rc_dec_log(rc, 1);
140
141 for (i = f->start_band; i < f->end_band; i++) {
142 f->tf_change[i] = ff_celt_tf_select[f->size][f->transient][tf_select][f->tf_change[i]];
143 }
144 }
145
celt_denormalize(CeltFrame * f,CeltBlock * block,float * data)146 static void celt_denormalize(CeltFrame *f, CeltBlock *block, float *data)
147 {
148 int i, j;
149
150 for (i = f->start_band; i < f->end_band; i++) {
151 float *dst = data + (ff_celt_freq_bands[i] << f->size);
152 float log_norm = block->energy[i] + ff_celt_mean_energy[i];
153 float norm = exp2f(FFMIN(log_norm, 32.0f));
154
155 for (j = 0; j < ff_celt_freq_range[i] << f->size; j++)
156 dst[j] *= norm;
157 }
158 }
159
celt_postfilter_apply_transition(CeltBlock * block,float * data)160 static void celt_postfilter_apply_transition(CeltBlock *block, float *data)
161 {
162 const int T0 = block->pf_period_old;
163 const int T1 = block->pf_period;
164
165 float g00, g01, g02;
166 float g10, g11, g12;
167
168 float x0, x1, x2, x3, x4;
169
170 int i;
171
172 if (block->pf_gains[0] == 0.0 &&
173 block->pf_gains_old[0] == 0.0)
174 return;
175
176 g00 = block->pf_gains_old[0];
177 g01 = block->pf_gains_old[1];
178 g02 = block->pf_gains_old[2];
179 g10 = block->pf_gains[0];
180 g11 = block->pf_gains[1];
181 g12 = block->pf_gains[2];
182
183 x1 = data[-T1 + 1];
184 x2 = data[-T1];
185 x3 = data[-T1 - 1];
186 x4 = data[-T1 - 2];
187
188 for (i = 0; i < CELT_OVERLAP; i++) {
189 float w = ff_celt_window2[i];
190 x0 = data[i - T1 + 2];
191
192 data[i] += (1.0 - w) * g00 * data[i - T0] +
193 (1.0 - w) * g01 * (data[i - T0 - 1] + data[i - T0 + 1]) +
194 (1.0 - w) * g02 * (data[i - T0 - 2] + data[i - T0 + 2]) +
195 w * g10 * x2 +
196 w * g11 * (x1 + x3) +
197 w * g12 * (x0 + x4);
198 x4 = x3;
199 x3 = x2;
200 x2 = x1;
201 x1 = x0;
202 }
203 }
204
celt_postfilter(CeltFrame * f,CeltBlock * block)205 static void celt_postfilter(CeltFrame *f, CeltBlock *block)
206 {
207 int len = f->blocksize * f->blocks;
208 const int filter_len = len - 2 * CELT_OVERLAP;
209
210 celt_postfilter_apply_transition(block, block->buf + 1024);
211
212 block->pf_period_old = block->pf_period;
213 memcpy(block->pf_gains_old, block->pf_gains, sizeof(block->pf_gains));
214
215 block->pf_period = block->pf_period_new;
216 memcpy(block->pf_gains, block->pf_gains_new, sizeof(block->pf_gains));
217
218 if (len > CELT_OVERLAP) {
219 celt_postfilter_apply_transition(block, block->buf + 1024 + CELT_OVERLAP);
220
221 if (block->pf_gains[0] > FLT_EPSILON && filter_len > 0)
222 f->opusdsp.postfilter(block->buf + 1024 + 2 * CELT_OVERLAP,
223 block->pf_period, block->pf_gains,
224 filter_len);
225
226 block->pf_period_old = block->pf_period;
227 memcpy(block->pf_gains_old, block->pf_gains, sizeof(block->pf_gains));
228 }
229
230 memmove(block->buf, block->buf + len, (1024 + CELT_OVERLAP / 2) * sizeof(float));
231 }
232
parse_postfilter(CeltFrame * f,OpusRangeCoder * rc,int consumed)233 static int parse_postfilter(CeltFrame *f, OpusRangeCoder *rc, int consumed)
234 {
235 int i;
236
237 memset(f->block[0].pf_gains_new, 0, sizeof(f->block[0].pf_gains_new));
238 memset(f->block[1].pf_gains_new, 0, sizeof(f->block[1].pf_gains_new));
239
240 if (f->start_band == 0 && consumed + 16 <= f->framebits) {
241 int has_postfilter = ff_opus_rc_dec_log(rc, 1);
242 if (has_postfilter) {
243 float gain;
244 int tapset, octave, period;
245
246 octave = ff_opus_rc_dec_uint(rc, 6);
247 period = (16 << octave) + ff_opus_rc_get_raw(rc, 4 + octave) - 1;
248 gain = 0.09375f * (ff_opus_rc_get_raw(rc, 3) + 1);
249 tapset = (opus_rc_tell(rc) + 2 <= f->framebits) ?
250 ff_opus_rc_dec_cdf(rc, ff_celt_model_tapset) : 0;
251
252 for (i = 0; i < 2; i++) {
253 CeltBlock *block = &f->block[i];
254
255 block->pf_period_new = FFMAX(period, CELT_POSTFILTER_MINPERIOD);
256 block->pf_gains_new[0] = gain * ff_celt_postfilter_taps[tapset][0];
257 block->pf_gains_new[1] = gain * ff_celt_postfilter_taps[tapset][1];
258 block->pf_gains_new[2] = gain * ff_celt_postfilter_taps[tapset][2];
259 }
260 }
261
262 consumed = opus_rc_tell(rc);
263 }
264
265 return consumed;
266 }
267
process_anticollapse(CeltFrame * f,CeltBlock * block,float * X)268 static void process_anticollapse(CeltFrame *f, CeltBlock *block, float *X)
269 {
270 int i, j, k;
271
272 for (i = f->start_band; i < f->end_band; i++) {
273 int renormalize = 0;
274 float *xptr;
275 float prev[2];
276 float Ediff, r;
277 float thresh, sqrt_1;
278 int depth;
279
280 /* depth in 1/8 bits */
281 depth = (1 + f->pulses[i]) / (ff_celt_freq_range[i] << f->size);
282 thresh = exp2f(-1.0 - 0.125f * depth);
283 sqrt_1 = 1.0f / sqrtf(ff_celt_freq_range[i] << f->size);
284
285 xptr = X + (ff_celt_freq_bands[i] << f->size);
286
287 prev[0] = block->prev_energy[0][i];
288 prev[1] = block->prev_energy[1][i];
289 if (f->channels == 1) {
290 CeltBlock *block1 = &f->block[1];
291
292 prev[0] = FFMAX(prev[0], block1->prev_energy[0][i]);
293 prev[1] = FFMAX(prev[1], block1->prev_energy[1][i]);
294 }
295 Ediff = block->energy[i] - FFMIN(prev[0], prev[1]);
296 Ediff = FFMAX(0, Ediff);
297
298 /* r needs to be multiplied by 2 or 2*sqrt(2) depending on LM because
299 short blocks don't have the same energy as long */
300 r = exp2f(1 - Ediff);
301 if (f->size == 3)
302 r *= M_SQRT2;
303 r = FFMIN(thresh, r) * sqrt_1;
304 for (k = 0; k < 1 << f->size; k++) {
305 /* Detect collapse */
306 if (!(block->collapse_masks[i] & 1 << k)) {
307 /* Fill with noise */
308 for (j = 0; j < ff_celt_freq_range[i]; j++)
309 xptr[(j << f->size) + k] = (celt_rng(f) & 0x8000) ? r : -r;
310 renormalize = 1;
311 }
312 }
313
314 /* We just added some energy, so we need to renormalize */
315 if (renormalize)
316 celt_renormalize_vector(xptr, ff_celt_freq_range[i] << f->size, 1.0f);
317 }
318 }
319
ff_celt_decode_frame(CeltFrame * f,OpusRangeCoder * rc,float ** output,int channels,int frame_size,int start_band,int end_band)320 int ff_celt_decode_frame(CeltFrame *f, OpusRangeCoder *rc,
321 float **output, int channels, int frame_size,
322 int start_band, int end_band)
323 {
324 int i, j, downmix = 0;
325 int consumed; // bits of entropy consumed thus far for this frame
326 MDCT15Context *imdct;
327
328 if (channels != 1 && channels != 2) {
329 av_log(f->avctx, AV_LOG_ERROR, "Invalid number of coded channels: %d\n",
330 channels);
331 return AVERROR_INVALIDDATA;
332 }
333 if (start_band < 0 || start_band > end_band || end_band > CELT_MAX_BANDS) {
334 av_log(f->avctx, AV_LOG_ERROR, "Invalid start/end band: %d %d\n",
335 start_band, end_band);
336 return AVERROR_INVALIDDATA;
337 }
338
339 f->silence = 0;
340 f->transient = 0;
341 f->anticollapse = 0;
342 f->flushed = 0;
343 f->channels = channels;
344 f->start_band = start_band;
345 f->end_band = end_band;
346 f->framebits = rc->rb.bytes * 8;
347
348 f->size = av_log2(frame_size / CELT_SHORT_BLOCKSIZE);
349 if (f->size > CELT_MAX_LOG_BLOCKS ||
350 frame_size != CELT_SHORT_BLOCKSIZE * (1 << f->size)) {
351 av_log(f->avctx, AV_LOG_ERROR, "Invalid CELT frame size: %d\n",
352 frame_size);
353 return AVERROR_INVALIDDATA;
354 }
355
356 if (!f->output_channels)
357 f->output_channels = channels;
358
359 for (i = 0; i < f->channels; i++) {
360 memset(f->block[i].coeffs, 0, sizeof(f->block[i].coeffs));
361 memset(f->block[i].collapse_masks, 0, sizeof(f->block[i].collapse_masks));
362 }
363
364 consumed = opus_rc_tell(rc);
365
366 /* obtain silence flag */
367 if (consumed >= f->framebits)
368 f->silence = 1;
369 else if (consumed == 1)
370 f->silence = ff_opus_rc_dec_log(rc, 15);
371
372
373 if (f->silence) {
374 consumed = f->framebits;
375 rc->total_bits += f->framebits - opus_rc_tell(rc);
376 }
377
378 /* obtain post-filter options */
379 consumed = parse_postfilter(f, rc, consumed);
380
381 /* obtain transient flag */
382 if (f->size != 0 && consumed+3 <= f->framebits)
383 f->transient = ff_opus_rc_dec_log(rc, 3);
384
385 f->blocks = f->transient ? 1 << f->size : 1;
386 f->blocksize = frame_size / f->blocks;
387
388 imdct = f->imdct[f->transient ? 0 : f->size];
389
390 if (channels == 1) {
391 for (i = 0; i < CELT_MAX_BANDS; i++)
392 f->block[0].energy[i] = FFMAX(f->block[0].energy[i], f->block[1].energy[i]);
393 }
394
395 celt_decode_coarse_energy(f, rc);
396 celt_decode_tf_changes (f, rc);
397 ff_celt_bitalloc (f, rc, 0);
398 celt_decode_fine_energy (f, rc);
399 ff_celt_quant_bands (f, rc);
400
401 if (f->anticollapse_needed)
402 f->anticollapse = ff_opus_rc_get_raw(rc, 1);
403
404 celt_decode_final_energy(f, rc);
405
406 /* apply anti-collapse processing and denormalization to
407 * each coded channel */
408 for (i = 0; i < f->channels; i++) {
409 CeltBlock *block = &f->block[i];
410
411 if (f->anticollapse)
412 process_anticollapse(f, block, f->block[i].coeffs);
413
414 celt_denormalize(f, block, f->block[i].coeffs);
415 }
416
417 /* stereo -> mono downmix */
418 if (f->output_channels < f->channels) {
419 f->dsp->vector_fmac_scalar(f->block[0].coeffs, f->block[1].coeffs, 1.0, FFALIGN(frame_size, 16));
420 downmix = 1;
421 } else if (f->output_channels > f->channels)
422 memcpy(f->block[1].coeffs, f->block[0].coeffs, frame_size * sizeof(float));
423
424 if (f->silence) {
425 for (i = 0; i < 2; i++) {
426 CeltBlock *block = &f->block[i];
427
428 for (j = 0; j < FF_ARRAY_ELEMS(block->energy); j++)
429 block->energy[j] = CELT_ENERGY_SILENCE;
430 }
431 memset(f->block[0].coeffs, 0, sizeof(f->block[0].coeffs));
432 memset(f->block[1].coeffs, 0, sizeof(f->block[1].coeffs));
433 }
434
435 /* transform and output for each output channel */
436 for (i = 0; i < f->output_channels; i++) {
437 CeltBlock *block = &f->block[i];
438
439 /* iMDCT and overlap-add */
440 for (j = 0; j < f->blocks; j++) {
441 float *dst = block->buf + 1024 + j * f->blocksize;
442
443 imdct->imdct_half(imdct, dst + CELT_OVERLAP / 2, f->block[i].coeffs + j,
444 f->blocks);
445 f->dsp->vector_fmul_window(dst, dst, dst + CELT_OVERLAP / 2,
446 ff_celt_window, CELT_OVERLAP / 2);
447 }
448
449 if (downmix)
450 f->dsp->vector_fmul_scalar(&block->buf[1024], &block->buf[1024], 0.5f, frame_size);
451
452 /* postfilter */
453 celt_postfilter(f, block);
454
455 /* deemphasis */
456 block->emph_coeff = f->opusdsp.deemphasis(output[i],
457 &block->buf[1024 - frame_size],
458 block->emph_coeff, frame_size);
459 }
460
461 if (channels == 1)
462 memcpy(f->block[1].energy, f->block[0].energy, sizeof(f->block[0].energy));
463
464 for (i = 0; i < 2; i++ ) {
465 CeltBlock *block = &f->block[i];
466
467 if (!f->transient) {
468 memcpy(block->prev_energy[1], block->prev_energy[0], sizeof(block->prev_energy[0]));
469 memcpy(block->prev_energy[0], block->energy, sizeof(block->prev_energy[0]));
470 } else {
471 for (j = 0; j < CELT_MAX_BANDS; j++)
472 block->prev_energy[0][j] = FFMIN(block->prev_energy[0][j], block->energy[j]);
473 }
474
475 for (j = 0; j < f->start_band; j++) {
476 block->prev_energy[0][j] = CELT_ENERGY_SILENCE;
477 block->energy[j] = 0.0;
478 }
479 for (j = f->end_band; j < CELT_MAX_BANDS; j++) {
480 block->prev_energy[0][j] = CELT_ENERGY_SILENCE;
481 block->energy[j] = 0.0;
482 }
483 }
484
485 f->seed = rc->range;
486
487 return 0;
488 }
489
ff_celt_flush(CeltFrame * f)490 void ff_celt_flush(CeltFrame *f)
491 {
492 int i, j;
493
494 if (f->flushed)
495 return;
496
497 for (i = 0; i < 2; i++) {
498 CeltBlock *block = &f->block[i];
499
500 for (j = 0; j < CELT_MAX_BANDS; j++)
501 block->prev_energy[0][j] = block->prev_energy[1][j] = CELT_ENERGY_SILENCE;
502
503 memset(block->energy, 0, sizeof(block->energy));
504 memset(block->buf, 0, sizeof(block->buf));
505
506 memset(block->pf_gains, 0, sizeof(block->pf_gains));
507 memset(block->pf_gains_old, 0, sizeof(block->pf_gains_old));
508 memset(block->pf_gains_new, 0, sizeof(block->pf_gains_new));
509
510 /* libopus uses CELT_EMPH_COEFF on init, but 0 is better since there's
511 * a lesser discontinuity when seeking.
512 * The deemphasis functions differ from libopus in that they require
513 * an initial state divided by the coefficient. */
514 block->emph_coeff = 0.0f / CELT_EMPH_COEFF;
515 }
516 f->seed = 0;
517
518 f->flushed = 1;
519 }
520
ff_celt_free(CeltFrame ** f)521 void ff_celt_free(CeltFrame **f)
522 {
523 CeltFrame *frm = *f;
524 int i;
525
526 if (!frm)
527 return;
528
529 for (i = 0; i < FF_ARRAY_ELEMS(frm->imdct); i++)
530 ff_mdct15_uninit(&frm->imdct[i]);
531
532 ff_celt_pvq_uninit(&frm->pvq);
533
534 av_freep(&frm->dsp);
535 av_freep(f);
536 }
537
ff_celt_init(AVCodecContext * avctx,CeltFrame ** f,int output_channels,int apply_phase_inv)538 int ff_celt_init(AVCodecContext *avctx, CeltFrame **f, int output_channels,
539 int apply_phase_inv)
540 {
541 CeltFrame *frm;
542 int i, ret;
543
544 if (output_channels != 1 && output_channels != 2) {
545 av_log(avctx, AV_LOG_ERROR, "Invalid number of output channels: %d\n",
546 output_channels);
547 return AVERROR(EINVAL);
548 }
549
550 frm = av_mallocz(sizeof(*frm));
551 if (!frm)
552 return AVERROR(ENOMEM);
553
554 frm->avctx = avctx;
555 frm->output_channels = output_channels;
556 frm->apply_phase_inv = apply_phase_inv;
557
558 for (i = 0; i < FF_ARRAY_ELEMS(frm->imdct); i++)
559 if ((ret = ff_mdct15_init(&frm->imdct[i], 1, i + 3, -1.0f/32768)) < 0)
560 goto fail;
561
562 if ((ret = ff_celt_pvq_init(&frm->pvq, 0)) < 0)
563 goto fail;
564
565 frm->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);
566 if (!frm->dsp) {
567 ret = AVERROR(ENOMEM);
568 goto fail;
569 }
570
571 ff_opus_dsp_init(&frm->opusdsp);
572 ff_celt_flush(frm);
573
574 *f = frm;
575
576 return 0;
577 fail:
578 ff_celt_free(&frm);
579 return ret;
580 }
581