2 * Copyright (c) 2012 Andrew D'Addesio
3 * Copyright (c) 2013-2014 Mozilla Corporation
4 * Copyright (c) 2016 Rostislav Pehlivanov <atomnuker@gmail.com>
6 * This file is part of FFmpeg.
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.
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.
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
28 #include "opus_celt.h"
32 /* Use the 2D z-transform to apply prediction in both the time domain (alpha)
33 * and the frequency domain (beta) */
34 static void celt_decode_coarse_energy(CeltFrame *f, OpusRangeCoder *rc)
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];
43 if (opus_rc_tell(rc) + 3 <= f->framebits && ff_opus_rc_dec_log(rc, 3)) {
45 beta = 1.0f - (4915.0f/32768.0f);
46 model = ff_celt_coarse_energy_dist[f->size][1];
49 for (i = 0; i < CELT_MAX_BANDS; i++) {
50 for (j = 0; j < f->channels; j++) {
51 CeltBlock *block = &f->block[j];
55 if (i < f->start_band || i >= f->end_band) {
56 block->energy[i] = 0.0;
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);
72 block->energy[i] = FFMAX(-9.0f, block->energy[i]) * alpha + prev[j] + value;
73 prev[j] += beta * value;
78 static void celt_decode_fine_energy(CeltFrame *f, OpusRangeCoder *rc)
81 for (i = f->start_band; i < f->end_band; i++) {
86 for (j = 0; j < f->channels; j++) {
87 CeltBlock *block = &f->block[j];
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;
97 static void celt_decode_final_energy(CeltFrame *f, OpusRangeCoder *rc)
100 int bits_left = f->framebits - opus_rc_tell(rc);
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)
107 for (j = 0; j < f->channels; j++) {
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;
119 static void celt_decode_tf_changes(CeltFrame *f, OpusRangeCoder *rc)
121 int i, diff = 0, tf_select = 0, tf_changed = 0, tf_select_bit;
122 int consumed, bits = f->transient ? 2 : 4;
124 consumed = opus_rc_tell(rc);
125 tf_select_bit = (f->size != 0 && consumed+bits+1 <= f->framebits);
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);
133 f->tf_change[i] = diff;
134 bits = f->transient ? 4 : 5;
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);
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]];
146 static void celt_denormalize(CeltFrame *f, CeltBlock *block, float *data)
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));
155 for (j = 0; j < ff_celt_freq_range[i] << f->size; j++)
160 static void celt_postfilter_apply_transition(CeltBlock *block, float *data)
162 const int T0 = block->pf_period_old;
163 const int T1 = block->pf_period;
168 float x0, x1, x2, x3, x4;
172 if (block->pf_gains[0] == 0.0 &&
173 block->pf_gains_old[0] == 0.0)
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];
188 for (i = 0; i < CELT_OVERLAP; i++) {
189 float w = ff_celt_window2[i];
190 x0 = data[i - T1 + 2];
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]) +
196 w * g11 * (x1 + x3) +
205 static void celt_postfilter(CeltFrame *f, CeltBlock *block)
207 int len = f->blocksize * f->blocks;
208 const int filter_len = len - 2 * CELT_OVERLAP;
210 celt_postfilter_apply_transition(block, block->buf + 1024);
212 block->pf_period_old = block->pf_period;
213 memcpy(block->pf_gains_old, block->pf_gains, sizeof(block->pf_gains));
215 block->pf_period = block->pf_period_new;
216 memcpy(block->pf_gains, block->pf_gains_new, sizeof(block->pf_gains));
218 if (len > CELT_OVERLAP) {
219 celt_postfilter_apply_transition(block, block->buf + 1024 + CELT_OVERLAP);
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,
226 block->pf_period_old = block->pf_period;
227 memcpy(block->pf_gains_old, block->pf_gains, sizeof(block->pf_gains));
230 memmove(block->buf, block->buf + len, (1024 + CELT_OVERLAP / 2) * sizeof(float));
233 static int parse_postfilter(CeltFrame *f, OpusRangeCoder *rc, int consumed)
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));
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) {
244 int tapset, octave, period;
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;
252 for (i = 0; i < 2; i++) {
253 CeltBlock *block = &f->block[i];
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];
262 consumed = opus_rc_tell(rc);
268 static void process_anticollapse(CeltFrame *f, CeltBlock *block, float *X)
272 for (i = f->start_band; i < f->end_band; i++) {
277 float thresh, sqrt_1;
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);
285 xptr = X + (ff_celt_freq_bands[i] << f->size);
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];
292 prev[0] = FFMAX(prev[0], block1->prev_energy[0][i]);
293 prev[1] = FFMAX(prev[1], block1->prev_energy[1][i]);
295 Ediff = block->energy[i] - FFMIN(prev[0], prev[1]);
296 Ediff = FFMAX(0, Ediff);
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);
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;
314 /* We just added some energy, so we need to renormalize */
316 celt_renormalize_vector(xptr, ff_celt_freq_range[i] << f->size, 1.0f);
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)
324 int i, j, downmix = 0;
325 int consumed; // bits of entropy consumed thus far for this frame
326 MDCT15Context *imdct;
328 if (channels != 1 && channels != 2) {
329 av_log(f->avctx, AV_LOG_ERROR, "Invalid number of coded channels: %d\n",
331 return AVERROR_INVALIDDATA;
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;
343 f->channels = channels;
344 f->start_band = start_band;
345 f->end_band = end_band;
346 f->framebits = rc->rb.bytes * 8;
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",
353 return AVERROR_INVALIDDATA;
356 if (!f->output_channels)
357 f->output_channels = channels;
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));
364 consumed = opus_rc_tell(rc);
366 /* obtain silence flag */
367 if (consumed >= f->framebits)
369 else if (consumed == 1)
370 f->silence = ff_opus_rc_dec_log(rc, 15);
374 consumed = f->framebits;
375 rc->total_bits += f->framebits - opus_rc_tell(rc);
378 /* obtain post-filter options */
379 consumed = parse_postfilter(f, rc, consumed);
381 /* obtain transient flag */
382 if (f->size != 0 && consumed+3 <= f->framebits)
383 f->transient = ff_opus_rc_dec_log(rc, 3);
385 f->blocks = f->transient ? 1 << f->size : 1;
386 f->blocksize = frame_size / f->blocks;
388 imdct = f->imdct[f->transient ? 0 : f->size];
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]);
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);
401 if (f->anticollapse_needed)
402 f->anticollapse = ff_opus_rc_get_raw(rc, 1);
404 celt_decode_final_energy(f, rc);
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];
412 process_anticollapse(f, block, f->block[i].coeffs);
414 celt_denormalize(f, block, f->block[i].coeffs);
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));
421 } else if (f->output_channels > f->channels)
422 memcpy(f->block[1].coeffs, f->block[0].coeffs, frame_size * sizeof(float));
425 for (i = 0; i < 2; i++) {
426 CeltBlock *block = &f->block[i];
428 for (j = 0; j < FF_ARRAY_ELEMS(block->energy); j++)
429 block->energy[j] = CELT_ENERGY_SILENCE;
431 memset(f->block[0].coeffs, 0, sizeof(f->block[0].coeffs));
432 memset(f->block[1].coeffs, 0, sizeof(f->block[1].coeffs));
435 /* transform and output for each output channel */
436 for (i = 0; i < f->output_channels; i++) {
437 CeltBlock *block = &f->block[i];
439 /* iMDCT and overlap-add */
440 for (j = 0; j < f->blocks; j++) {
441 float *dst = block->buf + 1024 + j * f->blocksize;
443 imdct->imdct_half(imdct, dst + CELT_OVERLAP / 2, f->block[i].coeffs + j,
445 f->dsp->vector_fmul_window(dst, dst, dst + CELT_OVERLAP / 2,
446 ff_celt_window, CELT_OVERLAP / 2);
450 f->dsp->vector_fmul_scalar(&block->buf[1024], &block->buf[1024], 0.5f, frame_size);
453 celt_postfilter(f, block);
456 block->emph_coeff = f->opusdsp.deemphasis(output[i],
457 &block->buf[1024 - frame_size],
458 block->emph_coeff, frame_size);
462 memcpy(f->block[1].energy, f->block[0].energy, sizeof(f->block[0].energy));
464 for (i = 0; i < 2; i++ ) {
465 CeltBlock *block = &f->block[i];
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]));
471 for (j = 0; j < CELT_MAX_BANDS; j++)
472 block->prev_energy[0][j] = FFMIN(block->prev_energy[0][j], block->energy[j]);
475 for (j = 0; j < f->start_band; j++) {
476 block->prev_energy[0][j] = CELT_ENERGY_SILENCE;
477 block->energy[j] = 0.0;
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;
490 void ff_celt_flush(CeltFrame *f)
497 for (i = 0; i < 2; i++) {
498 CeltBlock *block = &f->block[i];
500 for (j = 0; j < CELT_MAX_BANDS; j++)
501 block->prev_energy[0][j] = block->prev_energy[1][j] = CELT_ENERGY_SILENCE;
503 memset(block->energy, 0, sizeof(block->energy));
504 memset(block->buf, 0, sizeof(block->buf));
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));
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;
521 void ff_celt_free(CeltFrame **f)
529 for (i = 0; i < FF_ARRAY_ELEMS(frm->imdct); i++)
530 ff_mdct15_uninit(&frm->imdct[i]);
532 ff_celt_pvq_uninit(&frm->pvq);
538 int ff_celt_init(AVCodecContext *avctx, CeltFrame **f, int output_channels,
544 if (output_channels != 1 && output_channels != 2) {
545 av_log(avctx, AV_LOG_ERROR, "Invalid number of output channels: %d\n",
547 return AVERROR(EINVAL);
550 frm = av_mallocz(sizeof(*frm));
552 return AVERROR(ENOMEM);
555 frm->output_channels = output_channels;
556 frm->apply_phase_inv = apply_phase_inv;
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)
562 if ((ret = ff_celt_pvq_init(&frm->pvq, 0)) < 0)
565 frm->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);
567 ret = AVERROR(ENOMEM);
571 ff_opus_dsp_init(&frm->opusdsp);