3 * Copyright (c) 2017 Rostislav Pehlivanov <atomnuker@gmail.com>
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
24 #include "opusenc_psy.h"
27 #include "libavutil/float_dsp.h"
28 #include "libavutil/opt.h"
30 #include "bytestream.h"
31 #include "audio_frame_queue.h"
33 typedef struct OpusEncContext {
35 OpusEncOptions options;
36 OpusPsyContext psyctx;
37 AVCodecContext *avctx;
39 AVFloatDSPContext *dsp;
40 MDCT15Context *mdct[CELT_BLOCK_NB];
42 struct FFBufQueue bufqueue;
47 OpusPacketInfo packet;
54 /* Actual energy the decoder will have */
55 float last_quantized_energy[OPUS_MAX_CHANNELS][CELT_MAX_BANDS];
57 DECLARE_ALIGNED(32, float, scratch)[2048];
60 static void opus_write_extradata(AVCodecContext *avctx)
62 uint8_t *bs = avctx->extradata;
64 bytestream_put_buffer(&bs, "OpusHead", 8);
65 bytestream_put_byte (&bs, 0x1);
66 bytestream_put_byte (&bs, avctx->channels);
67 bytestream_put_le16 (&bs, avctx->initial_padding);
68 bytestream_put_le32 (&bs, avctx->sample_rate);
69 bytestream_put_le16 (&bs, 0x0);
70 bytestream_put_byte (&bs, 0x0); /* Default layout */
73 static int opus_gen_toc(OpusEncContext *s, uint8_t *toc, int *size, int *fsize_needed)
75 int i, tmp = 0x0, extended_toc = 0;
76 static const int toc_cfg[][OPUS_MODE_NB][OPUS_BANDWITH_NB] = {
77 /* Silk Hybrid Celt Layer */
78 /* NB MB WB SWB FB NB MB WB SWB FB NB MB WB SWB FB Bandwidth */
79 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 17, 0, 21, 25, 29 } }, /* 2.5 ms */
80 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 18, 0, 22, 26, 30 } }, /* 5 ms */
81 { { 1, 5, 9, 0, 0 }, { 0, 0, 0, 13, 15 }, { 19, 0, 23, 27, 31 } }, /* 10 ms */
82 { { 2, 6, 10, 0, 0 }, { 0, 0, 0, 14, 16 }, { 20, 0, 24, 28, 32 } }, /* 20 ms */
83 { { 3, 7, 11, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 40 ms */
84 { { 4, 8, 12, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 60 ms */
86 int cfg = toc_cfg[s->packet.framesize][s->packet.mode][s->packet.bandwidth];
90 if (s->packet.frames == 2) { /* 2 packets */
91 if (s->frame[0].framebits == s->frame[1].framebits) { /* same size */
93 } else { /* different size */
95 *fsize_needed = 1; /* put frame sizes in the packet */
97 } else if (s->packet.frames > 2) {
101 tmp |= (s->channels > 1) << 2; /* Stereo or mono */
102 tmp |= (cfg - 1) << 3; /* codec configuration */
105 for (i = 0; i < (s->packet.frames - 1); i++)
106 *fsize_needed |= (s->frame[i].framebits != s->frame[i + 1].framebits);
107 tmp = (*fsize_needed) << 7; /* vbr flag */
108 tmp |= (0) << 6; /* padding flag */
109 tmp |= s->packet.frames;
112 *size = 1 + extended_toc;
116 static void celt_frame_setup_input(OpusEncContext *s, CeltFrame *f)
120 const int subframesize = s->avctx->frame_size;
121 int subframes = OPUS_BLOCK_SIZE(s->packet.framesize) / subframesize;
123 cur = ff_bufqueue_get(&s->bufqueue);
125 for (ch = 0; ch < f->channels; ch++) {
126 CeltBlock *b = &f->block[ch];
127 const void *input = cur->extended_data[ch];
128 size_t bps = av_get_bytes_per_sample(cur->format);
129 memcpy(b->overlap, input, bps*cur->nb_samples);
134 for (sf = 0; sf < subframes; sf++) {
135 if (sf != (subframes - 1))
136 cur = ff_bufqueue_get(&s->bufqueue);
138 cur = ff_bufqueue_peek(&s->bufqueue, 0);
140 for (ch = 0; ch < f->channels; ch++) {
141 CeltBlock *b = &f->block[ch];
142 const void *input = cur->extended_data[ch];
143 const size_t bps = av_get_bytes_per_sample(cur->format);
144 const size_t left = (subframesize - cur->nb_samples)*bps;
145 const size_t len = FFMIN(subframesize, cur->nb_samples)*bps;
146 memcpy(&b->samples[sf*subframesize], input, len);
147 memset(&b->samples[cur->nb_samples], 0, left);
150 /* Last frame isn't popped off and freed yet - we need it for overlap */
151 if (sf != (subframes - 1))
156 /* Apply the pre emphasis filter */
157 static void celt_apply_preemph_filter(OpusEncContext *s, CeltFrame *f)
160 const int subframesize = s->avctx->frame_size;
161 const int subframes = OPUS_BLOCK_SIZE(s->packet.framesize) / subframesize;
164 for (ch = 0; ch < f->channels; ch++) {
165 CeltBlock *b = &f->block[ch];
166 float m = b->emph_coeff;
167 for (i = 0; i < CELT_OVERLAP; i++) {
168 float sample = b->overlap[i];
169 b->overlap[i] = sample - m;
170 m = sample * CELT_EMPH_COEFF;
175 /* Filter the samples but do not update the last subframe's coeff - overlap ^^^ */
176 for (sf = 0; sf < subframes; sf++) {
177 for (ch = 0; ch < f->channels; ch++) {
178 CeltBlock *b = &f->block[ch];
179 float m = b->emph_coeff;
180 for (i = 0; i < subframesize; i++) {
181 float sample = b->samples[sf*subframesize + i];
182 b->samples[sf*subframesize + i] = sample - m;
183 m = sample * CELT_EMPH_COEFF;
185 if (sf != (subframes - 1))
191 /* Create the window and do the mdct */
192 static void celt_frame_mdct(OpusEncContext *s, CeltFrame *f)
195 float *win = s->scratch, *temp = s->scratch + 1920;
198 for (ch = 0; ch < f->channels; ch++) {
199 CeltBlock *b = &f->block[ch];
200 float *src1 = b->overlap;
201 for (t = 0; t < f->blocks; t++) {
202 float *src2 = &b->samples[CELT_OVERLAP*t];
203 s->dsp->vector_fmul(win, src1, ff_celt_window, 128);
204 s->dsp->vector_fmul_reverse(&win[CELT_OVERLAP], src2,
205 ff_celt_window - 8, 128);
207 s->mdct[0]->mdct(s->mdct[0], b->coeffs + t, win, f->blocks);
211 int blk_len = OPUS_BLOCK_SIZE(f->size), wlen = OPUS_BLOCK_SIZE(f->size + 1);
212 int rwin = blk_len - CELT_OVERLAP, lap_dst = (wlen - blk_len - CELT_OVERLAP) >> 1;
213 memset(win, 0, wlen*sizeof(float));
214 for (ch = 0; ch < f->channels; ch++) {
215 CeltBlock *b = &f->block[ch];
218 s->dsp->vector_fmul(temp, b->overlap, ff_celt_window, 128);
219 memcpy(win + lap_dst, temp, CELT_OVERLAP*sizeof(float));
221 /* Samples, flat top window */
222 memcpy(&win[lap_dst + CELT_OVERLAP], b->samples, rwin*sizeof(float));
224 /* Samples, windowed */
225 s->dsp->vector_fmul_reverse(temp, b->samples + rwin,
226 ff_celt_window - 8, 128);
227 memcpy(win + lap_dst + blk_len, temp, CELT_OVERLAP*sizeof(float));
229 s->mdct[f->size]->mdct(s->mdct[f->size], b->coeffs, win, 1);
233 for (ch = 0; ch < f->channels; ch++) {
234 CeltBlock *block = &f->block[ch];
235 for (i = 0; i < CELT_MAX_BANDS; i++) {
237 int band_offset = ff_celt_freq_bands[i] << f->size;
238 int band_size = ff_celt_freq_range[i] << f->size;
239 float *coeffs = &block->coeffs[band_offset];
241 for (j = 0; j < band_size; j++)
242 ener += coeffs[j]*coeffs[j];
244 block->lin_energy[i] = sqrtf(ener) + FLT_EPSILON;
245 ener = 1.0f/block->lin_energy[i];
247 for (j = 0; j < band_size; j++)
250 block->energy[i] = log2f(block->lin_energy[i]) - ff_celt_mean_energy[i];
252 /* CELT_ENERGY_SILENCE is what the decoder uses and its not -infinity */
253 block->energy[i] = FFMAX(block->energy[i], CELT_ENERGY_SILENCE);
258 static void celt_enc_tf(OpusRangeCoder *rc, CeltFrame *f)
260 int i, tf_select = 0, diff = 0, tf_changed = 0, tf_select_needed;
261 int bits = f->transient ? 2 : 4;
263 tf_select_needed = ((f->size && (opus_rc_tell(rc) + bits + 1) <= f->framebits));
265 for (i = f->start_band; i < f->end_band; i++) {
266 if ((opus_rc_tell(rc) + bits + tf_select_needed) <= f->framebits) {
267 const int tbit = (diff ^ 1) == f->tf_change[i];
268 ff_opus_rc_enc_log(rc, tbit, bits);
272 bits = f->transient ? 4 : 5;
275 if (tf_select_needed && ff_celt_tf_select[f->size][f->transient][0][tf_changed] !=
276 ff_celt_tf_select[f->size][f->transient][1][tf_changed]) {
277 ff_opus_rc_enc_log(rc, f->tf_select, 1);
278 tf_select = f->tf_select;
281 for (i = f->start_band; i < f->end_band; i++)
282 f->tf_change[i] = ff_celt_tf_select[f->size][f->transient][tf_select][f->tf_change[i]];
285 void ff_celt_enc_bitalloc(OpusRangeCoder *rc, CeltFrame *f)
287 int i, j, low, high, total, done, bandbits, remaining, tbits_8ths;
288 int skip_startband = f->start_band;
290 int intensitystereo_bit = 0;
291 int dualstereo_bit = 0;
296 int boost[CELT_MAX_BANDS];
297 int trim_offset[CELT_MAX_BANDS];
298 int threshold[CELT_MAX_BANDS];
299 int bits1[CELT_MAX_BANDS];
300 int bits2[CELT_MAX_BANDS];
302 /* Tell the spread to the decoder */
303 if (opus_rc_tell(rc) + 4 <= f->framebits)
304 ff_opus_rc_enc_cdf(rc, f->spread, ff_celt_model_spread);
306 f->spread = CELT_SPREAD_NORMAL;
308 /* Generate static allocation caps */
309 for (i = 0; i < CELT_MAX_BANDS; i++) {
310 cap[i] = (ff_celt_static_caps[f->size][f->channels - 1][i] + 64)
311 * ff_celt_freq_range[i] << (f->channels - 1) << f->size >> 2;
315 tbits_8ths = f->framebits << 3;
316 for (i = f->start_band; i < f->end_band; i++) {
317 int quanta, b_dynalloc, boost_amount = f->alloc_boost[i];
321 quanta = ff_celt_freq_range[i] << (f->channels - 1) << f->size;
322 quanta = FFMIN(quanta << 3, FFMAX(6 << 3, quanta));
323 b_dynalloc = dynalloc;
325 while (opus_rc_tell_frac(rc) + (b_dynalloc << 3) < tbits_8ths && boost[i] < cap[i]) {
326 int is_boost = boost_amount--;
328 ff_opus_rc_enc_log(rc, is_boost, b_dynalloc);
333 tbits_8ths -= quanta;
339 dynalloc = FFMAX(2, dynalloc - 1);
342 /* Put allocation trim */
343 if (opus_rc_tell_frac(rc) + (6 << 3) <= tbits_8ths)
344 ff_opus_rc_enc_cdf(rc, f->alloc_trim, ff_celt_model_alloc_trim);
346 /* Anti-collapse bit reservation */
347 tbits_8ths = (f->framebits << 3) - opus_rc_tell_frac(rc) - 1;
348 f->anticollapse_needed = 0;
349 if (f->transient && f->size >= 2 && tbits_8ths >= ((f->size + 2) << 3))
350 f->anticollapse_needed = 1 << 3;
351 tbits_8ths -= f->anticollapse_needed;
353 /* Band skip bit reservation */
354 if (tbits_8ths >= 1 << 3)
356 tbits_8ths -= skip_bit;
358 /* Intensity/dual stereo bit reservation */
359 if (f->channels == 2) {
360 intensitystereo_bit = ff_celt_log2_frac[f->end_band - f->start_band];
361 if (intensitystereo_bit <= tbits_8ths) {
362 tbits_8ths -= intensitystereo_bit;
363 if (tbits_8ths >= 1 << 3) {
364 dualstereo_bit = 1 << 3;
365 tbits_8ths -= 1 << 3;
368 intensitystereo_bit = 0;
373 for (i = f->start_band; i < f->end_band; i++) {
374 int trim = f->alloc_trim - 5 - f->size;
375 int band = ff_celt_freq_range[i] * (f->end_band - i - 1);
376 int duration = f->size + 3;
377 int scale = duration + f->channels - 1;
379 /* PVQ minimum allocation threshold, below this value the band is
381 threshold[i] = FFMAX(3 * ff_celt_freq_range[i] << duration >> 4,
384 trim_offset[i] = trim * (band << scale) >> 6;
386 if (ff_celt_freq_range[i] << f->size == 1)
387 trim_offset[i] -= f->channels << 3;
392 high = CELT_VECTORS - 1;
393 while (low <= high) {
394 int center = (low + high) >> 1;
397 for (i = f->end_band - 1; i >= f->start_band; i--) {
398 bandbits = ff_celt_freq_range[i] * ff_celt_static_alloc[center][i]
399 << (f->channels - 1) << f->size >> 2;
402 bandbits = FFMAX(0, bandbits + trim_offset[i]);
403 bandbits += boost[i];
405 if (bandbits >= threshold[i] || done) {
407 total += FFMIN(bandbits, cap[i]);
408 } else if (bandbits >= f->channels << 3)
409 total += f->channels << 3;
412 if (total > tbits_8ths)
420 for (i = f->start_band; i < f->end_band; i++) {
421 bits1[i] = ff_celt_freq_range[i] * ff_celt_static_alloc[low][i]
422 << (f->channels - 1) << f->size >> 2;
423 bits2[i] = high >= CELT_VECTORS ? cap[i] :
424 ff_celt_freq_range[i] * ff_celt_static_alloc[high][i]
425 << (f->channels - 1) << f->size >> 2;
428 bits1[i] = FFMAX(0, bits1[i] + trim_offset[i]);
430 bits2[i] = FFMAX(0, bits2[i] + trim_offset[i]);
432 bits1[i] += boost[i];
433 bits2[i] += boost[i];
437 bits2[i] = FFMAX(0, bits2[i] - bits1[i]);
442 high = 1 << CELT_ALLOC_STEPS;
443 for (i = 0; i < CELT_ALLOC_STEPS; i++) {
444 int center = (low + high) >> 1;
447 for (j = f->end_band - 1; j >= f->start_band; j--) {
448 bandbits = bits1[j] + (center * bits2[j] >> CELT_ALLOC_STEPS);
450 if (bandbits >= threshold[j] || done) {
452 total += FFMIN(bandbits, cap[j]);
453 } else if (bandbits >= f->channels << 3)
454 total += f->channels << 3;
456 if (total > tbits_8ths)
464 for (i = f->end_band - 1; i >= f->start_band; i--) {
465 bandbits = bits1[i] + (low * bits2[i] >> CELT_ALLOC_STEPS);
467 if (bandbits >= threshold[i] || done)
470 bandbits = (bandbits >= f->channels << 3) ?
471 f->channels << 3 : 0;
473 bandbits = FFMIN(bandbits, cap[i]);
474 f->pulses[i] = bandbits;
479 for (f->coded_bands = f->end_band; ; f->coded_bands--) {
481 j = f->coded_bands - 1;
483 if (j == skip_startband) {
484 /* all remaining bands are not skipped */
485 tbits_8ths += skip_bit;
489 /* determine the number of bits available for coding "do not skip" markers */
490 remaining = tbits_8ths - total;
491 bandbits = remaining / (ff_celt_freq_bands[j+1] - ff_celt_freq_bands[f->start_band]);
492 remaining -= bandbits * (ff_celt_freq_bands[j+1] - ff_celt_freq_bands[f->start_band]);
493 allocation = f->pulses[j] + bandbits * ff_celt_freq_range[j]
494 + FFMAX(0, remaining - (ff_celt_freq_bands[j] - ff_celt_freq_bands[f->start_band]));
496 /* a "do not skip" marker is only coded if the allocation is
497 above the chosen threshold */
498 if (allocation >= FFMAX(threshold[j], (f->channels + 1) << 3)) {
499 const int do_not_skip = f->coded_bands <= f->skip_band_floor;
500 ff_opus_rc_enc_log(rc, do_not_skip, 1);
505 allocation -= 1 << 3;
508 /* the band is skipped, so reclaim its bits */
509 total -= f->pulses[j];
510 if (intensitystereo_bit) {
511 total -= intensitystereo_bit;
512 intensitystereo_bit = ff_celt_log2_frac[j - f->start_band];
513 total += intensitystereo_bit;
516 total += f->pulses[j] = (allocation >= f->channels << 3) ? f->channels << 3 : 0;
519 /* Encode stereo flags */
520 if (intensitystereo_bit) {
521 f->intensity_stereo = FFMIN(f->intensity_stereo, f->coded_bands);
522 ff_opus_rc_enc_uint(rc, f->intensity_stereo, f->coded_bands + 1 - f->start_band);
524 if (f->intensity_stereo <= f->start_band)
525 tbits_8ths += dualstereo_bit; /* no intensity stereo means no dual stereo */
526 else if (dualstereo_bit)
527 ff_opus_rc_enc_log(rc, f->dual_stereo, 1);
529 /* Supply the remaining bits in this frame to lower bands */
530 remaining = tbits_8ths - total;
531 bandbits = remaining / (ff_celt_freq_bands[f->coded_bands] - ff_celt_freq_bands[f->start_band]);
532 remaining -= bandbits * (ff_celt_freq_bands[f->coded_bands] - ff_celt_freq_bands[f->start_band]);
533 for (i = f->start_band; i < f->coded_bands; i++) {
534 int bits = FFMIN(remaining, ff_celt_freq_range[i]);
536 f->pulses[i] += bits + bandbits * ff_celt_freq_range[i];
540 /* Finally determine the allocation */
541 for (i = f->start_band; i < f->coded_bands; i++) {
542 int N = ff_celt_freq_range[i] << f->size;
543 int prev_extra = extrabits;
544 f->pulses[i] += extrabits;
547 int dof; // degrees of freedom
548 int temp; // dof * channels * log(dof)
549 int offset; // fine energy quantization offset, i.e.
550 // extra bits assigned over the standard
552 int fine_bits, max_bits;
554 extrabits = FFMAX(0, f->pulses[i] - cap[i]);
555 f->pulses[i] -= extrabits;
557 /* intensity stereo makes use of an extra degree of freedom */
558 dof = N * f->channels + (f->channels == 2 && N > 2 && !f->dual_stereo && i < f->intensity_stereo);
559 temp = dof * (ff_celt_log_freq_range[i] + (f->size << 3));
560 offset = (temp >> 1) - dof * CELT_FINE_OFFSET;
561 if (N == 2) /* dof=2 is the only case that doesn't fit the model */
564 /* grant an additional bias for the first and second pulses */
565 if (f->pulses[i] + offset < 2 * (dof << 3))
567 else if (f->pulses[i] + offset < 3 * (dof << 3))
570 fine_bits = (f->pulses[i] + offset + (dof << 2)) / (dof << 3);
571 max_bits = FFMIN((f->pulses[i] >> 3) >> (f->channels - 1), CELT_MAX_FINE_BITS);
573 max_bits = FFMAX(max_bits, 0);
575 f->fine_bits[i] = av_clip(fine_bits, 0, max_bits);
577 /* if fine_bits was rounded down or capped,
578 give priority for the final fine energy pass */
579 f->fine_priority[i] = (f->fine_bits[i] * (dof << 3) >= f->pulses[i] + offset);
581 /* the remaining bits are assigned to PVQ */
582 f->pulses[i] -= f->fine_bits[i] << (f->channels - 1) << 3;
584 /* all bits go to fine energy except for the sign bit */
585 extrabits = FFMAX(0, f->pulses[i] - (f->channels << 3));
586 f->pulses[i] -= extrabits;
588 f->fine_priority[i] = 1;
591 /* hand back a limited number of extra fine energy bits to this band */
593 int fineextra = FFMIN(extrabits >> (f->channels + 2),
594 CELT_MAX_FINE_BITS - f->fine_bits[i]);
595 f->fine_bits[i] += fineextra;
597 fineextra <<= f->channels + 2;
598 f->fine_priority[i] = (fineextra >= extrabits - prev_extra);
599 extrabits -= fineextra;
602 f->remaining = extrabits;
604 /* skipped bands dedicate all of their bits for fine energy */
605 for (; i < f->end_band; i++) {
606 f->fine_bits[i] = f->pulses[i] >> (f->channels - 1) >> 3;
608 f->fine_priority[i] = f->fine_bits[i] < 1;
612 static void celt_enc_quant_pfilter(OpusRangeCoder *rc, CeltFrame *f)
614 float gain = f->pf_gain;
615 int i, txval, octave = f->pf_octave, period = f->pf_period, tapset = f->pf_tapset;
617 ff_opus_rc_enc_log(rc, f->pfilter, 1);
622 txval = FFMIN(octave, 6);
623 ff_opus_rc_enc_uint(rc, txval, 6);
626 txval = av_clip(period - (16 << octave) + 1, 0, (1 << (4 + octave)) - 1);
627 ff_opus_rc_put_raw(rc, period, 4 + octave);
628 period = txval + (16 << octave) - 1;
630 txval = FFMIN(((int)(gain / 0.09375f)) - 1, 7);
631 ff_opus_rc_put_raw(rc, txval, 3);
632 gain = 0.09375f * (txval + 1);
634 if ((opus_rc_tell(rc) + 2) <= f->framebits)
635 ff_opus_rc_enc_cdf(rc, tapset, ff_celt_model_tapset);
638 /* Finally create the coeffs */
639 for (i = 0; i < 2; i++) {
640 CeltBlock *block = &f->block[i];
642 block->pf_period_new = FFMAX(period, CELT_POSTFILTER_MINPERIOD);
643 block->pf_gains_new[0] = gain * ff_celt_postfilter_taps[tapset][0];
644 block->pf_gains_new[1] = gain * ff_celt_postfilter_taps[tapset][1];
645 block->pf_gains_new[2] = gain * ff_celt_postfilter_taps[tapset][2];
649 static void exp_quant_coarse(OpusRangeCoder *rc, CeltFrame *f,
650 float last_energy[][CELT_MAX_BANDS], int intra)
653 float alpha, beta, prev[2] = { 0, 0 };
654 const uint8_t *pmod = ff_celt_coarse_energy_dist[f->size][intra];
656 /* Inter is really just differential coding */
657 if (opus_rc_tell(rc) + 3 <= f->framebits)
658 ff_opus_rc_enc_log(rc, intra, 3);
664 beta = 1.0f - (4915.0f/32768.0f);
666 alpha = ff_celt_alpha_coef[f->size];
667 beta = ff_celt_beta_coef[f->size];
670 for (i = f->start_band; i < f->end_band; i++) {
671 for (ch = 0; ch < f->channels; ch++) {
672 CeltBlock *block = &f->block[ch];
673 const int left = f->framebits - opus_rc_tell(rc);
674 const float last = FFMAX(-9.0f, last_energy[ch][i]);
675 float diff = block->energy[i] - prev[ch] - last*alpha;
676 int q_en = lrintf(diff);
678 ff_opus_rc_enc_laplace(rc, &q_en, pmod[i << 1] << 7, pmod[(i << 1) + 1] << 6);
679 } else if (left >= 2) {
680 q_en = av_clip(q_en, -1, 1);
681 ff_opus_rc_enc_cdf(rc, 2*q_en + 3*(q_en < 0), ff_celt_model_energy_small);
682 } else if (left >= 1) {
683 q_en = av_clip(q_en, -1, 0);
684 ff_opus_rc_enc_log(rc, (q_en & 1), 1);
687 block->error_energy[i] = q_en - diff;
688 prev[ch] += beta * q_en;
693 static void celt_quant_coarse(OpusRangeCoder *rc, CeltFrame *f,
694 float last_energy[][CELT_MAX_BANDS])
696 uint32_t inter, intra;
697 OPUS_RC_CHECKPOINT_SPAWN(rc);
699 exp_quant_coarse(rc, f, last_energy, 1);
700 intra = OPUS_RC_CHECKPOINT_BITS(rc);
702 OPUS_RC_CHECKPOINT_ROLLBACK(rc);
704 exp_quant_coarse(rc, f, last_energy, 0);
705 inter = OPUS_RC_CHECKPOINT_BITS(rc);
707 if (inter > intra) { /* Unlikely */
708 OPUS_RC_CHECKPOINT_ROLLBACK(rc);
709 exp_quant_coarse(rc, f, last_energy, 1);
713 static void celt_quant_fine(OpusRangeCoder *rc, CeltFrame *f)
716 for (i = f->start_band; i < f->end_band; i++) {
717 if (!f->fine_bits[i])
719 for (ch = 0; ch < f->channels; ch++) {
720 CeltBlock *block = &f->block[ch];
721 int quant, lim = (1 << f->fine_bits[i]);
722 float offset, diff = 0.5f - block->error_energy[i];
723 quant = av_clip(floor(diff*lim), 0, lim - 1);
724 ff_opus_rc_put_raw(rc, quant, f->fine_bits[i]);
725 offset = 0.5f - ((quant + 0.5f) * (1 << (14 - f->fine_bits[i])) / 16384.0f);
726 block->error_energy[i] -= offset;
731 static void celt_quant_final(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f)
734 for (priority = 0; priority < 2; priority++) {
735 for (i = f->start_band; i < f->end_band && (f->framebits - opus_rc_tell(rc)) >= f->channels; i++) {
736 if (f->fine_priority[i] != priority || f->fine_bits[i] >= CELT_MAX_FINE_BITS)
738 for (ch = 0; ch < f->channels; ch++) {
739 CeltBlock *block = &f->block[ch];
740 const float err = block->error_energy[i];
741 const float offset = 0.5f * (1 << (14 - f->fine_bits[i] - 1)) / 16384.0f;
742 const int sign = FFABS(err + offset) < FFABS(err - offset);
743 ff_opus_rc_put_raw(rc, sign, 1);
744 block->error_energy[i] -= offset*(1 - 2*sign);
750 static void celt_quant_bands(OpusRangeCoder *rc, CeltFrame *f)
752 float lowband_scratch[8 * 22];
753 float norm[2 * 8 * 100];
755 int totalbits = (f->framebits << 3) - f->anticollapse_needed;
757 int update_lowband = 1;
758 int lowband_offset = 0;
762 for (i = f->start_band; i < f->end_band; i++) {
763 uint32_t cm[2] = { (1 << f->blocks) - 1, (1 << f->blocks) - 1 };
764 int band_offset = ff_celt_freq_bands[i] << f->size;
765 int band_size = ff_celt_freq_range[i] << f->size;
766 float *X = f->block[0].coeffs + band_offset;
767 float *Y = (f->channels == 2) ? f->block[1].coeffs + band_offset : NULL;
769 int consumed = opus_rc_tell_frac(rc);
770 float *norm2 = norm + 8 * 100;
771 int effective_lowband = -1;
774 /* Compute how many bits we want to allocate to this band */
775 if (i != f->start_band)
776 f->remaining -= consumed;
777 f->remaining2 = totalbits - consumed - 1;
778 if (i <= f->coded_bands - 1) {
779 int curr_balance = f->remaining / FFMIN(3, f->coded_bands-i);
780 b = av_clip_uintp2(FFMIN(f->remaining2 + 1, f->pulses[i] + curr_balance), 14);
783 if (ff_celt_freq_bands[i] - ff_celt_freq_range[i] >= ff_celt_freq_bands[f->start_band] &&
784 (update_lowband || lowband_offset == 0))
787 /* Get a conservative estimate of the collapse_mask's for the bands we're
788 going to be folding from. */
789 if (lowband_offset != 0 && (f->spread != CELT_SPREAD_AGGRESSIVE ||
790 f->blocks > 1 || f->tf_change[i] < 0)) {
791 int foldstart, foldend;
793 /* This ensures we never repeat spectral content within one band */
794 effective_lowband = FFMAX(ff_celt_freq_bands[f->start_band],
795 ff_celt_freq_bands[lowband_offset] - ff_celt_freq_range[i]);
796 foldstart = lowband_offset;
797 while (ff_celt_freq_bands[--foldstart] > effective_lowband);
798 foldend = lowband_offset - 1;
799 while (ff_celt_freq_bands[++foldend] < effective_lowband + ff_celt_freq_range[i]);
802 for (j = foldstart; j < foldend; j++) {
803 cm[0] |= f->block[0].collapse_masks[j];
804 cm[1] |= f->block[f->channels - 1].collapse_masks[j];
808 if (f->dual_stereo && i == f->intensity_stereo) {
809 /* Switch off dual stereo to do intensity */
811 for (j = ff_celt_freq_bands[f->start_band] << f->size; j < band_offset; j++)
812 norm[j] = (norm[j] + norm2[j]) / 2;
815 if (f->dual_stereo) {
816 cm[0] = f->pvq->encode_band(f->pvq, f, rc, i, X, NULL, band_size, b / 2, f->blocks,
817 effective_lowband != -1 ? norm + (effective_lowband << f->size) : NULL, f->size,
818 norm + band_offset, 0, 1.0f, lowband_scratch, cm[0]);
820 cm[1] = f->pvq->encode_band(f->pvq, f, rc, i, Y, NULL, band_size, b / 2, f->blocks,
821 effective_lowband != -1 ? norm2 + (effective_lowband << f->size) : NULL, f->size,
822 norm2 + band_offset, 0, 1.0f, lowband_scratch, cm[1]);
824 cm[0] = f->pvq->encode_band(f->pvq, f, rc, i, X, Y, band_size, b, f->blocks,
825 effective_lowband != -1 ? norm + (effective_lowband << f->size) : NULL, f->size,
826 norm + band_offset, 0, 1.0f, lowband_scratch, cm[0] | cm[1]);
830 f->block[0].collapse_masks[i] = (uint8_t)cm[0];
831 f->block[f->channels - 1].collapse_masks[i] = (uint8_t)cm[1];
832 f->remaining += f->pulses[i] + consumed;
834 /* Update the folding position only as long as we have 1 bit/sample depth */
835 update_lowband = (b > band_size << 3);
839 static void celt_encode_frame(OpusEncContext *s, OpusRangeCoder *rc,
840 CeltFrame *f, int index)
844 ff_opus_rc_enc_init(rc);
846 ff_opus_psy_celt_frame_init(&s->psyctx, f, index);
848 celt_frame_setup_input(s, f);
851 if (f->framebits >= 16)
852 ff_opus_rc_enc_log(rc, 1, 15); /* Silence (if using explicit singalling) */
853 for (ch = 0; ch < s->channels; ch++)
854 memset(s->last_quantized_energy[ch], 0.0f, sizeof(float)*CELT_MAX_BANDS);
859 celt_apply_preemph_filter(s, f);
861 ff_opus_rc_enc_log(rc, 0, 15);
862 celt_enc_quant_pfilter(rc, f);
866 celt_frame_mdct(s, f);
868 /* Need to handle transient/non-transient switches at any point during analysis */
869 while (ff_opus_psy_celt_frame_process(&s->psyctx, f, index))
870 celt_frame_mdct(s, f);
872 ff_opus_rc_enc_init(rc);
875 ff_opus_rc_enc_log(rc, 0, 15);
878 if (!f->start_band && opus_rc_tell(rc) + 16 <= f->framebits)
879 celt_enc_quant_pfilter(rc, f);
882 if (f->size && opus_rc_tell(rc) + 3 <= f->framebits)
883 ff_opus_rc_enc_log(rc, f->transient, 3);
886 celt_quant_coarse(rc, f, s->last_quantized_energy);
888 ff_celt_enc_bitalloc(rc, f);
889 celt_quant_fine (rc, f);
890 celt_quant_bands (rc, f);
892 /* Anticollapse bit */
893 if (f->anticollapse_needed)
894 ff_opus_rc_put_raw(rc, f->anticollapse, 1);
896 /* Final per-band energy adjustments from leftover bits */
897 celt_quant_final(s, rc, f);
899 for (ch = 0; ch < f->channels; ch++) {
900 CeltBlock *block = &f->block[ch];
901 for (i = 0; i < CELT_MAX_BANDS; i++)
902 s->last_quantized_energy[ch][i] = block->energy[i] + block->error_energy[i];
906 static inline int write_opuslacing(uint8_t *dst, int v)
908 dst[0] = FFMIN(v - FFALIGN(v - 255, 4), v);
909 dst[1] = v - dst[0] >> 2;
910 return 1 + (v >= 252);
913 static void opus_packet_assembler(OpusEncContext *s, AVPacket *avpkt)
915 int i, offset, fsize_needed;
918 opus_gen_toc(s, avpkt->data, &offset, &fsize_needed);
920 /* Frame sizes if needed */
922 for (i = 0; i < s->packet.frames - 1; i++) {
923 offset += write_opuslacing(avpkt->data + offset,
924 s->frame[i].framebits >> 3);
929 for (i = 0; i < s->packet.frames; i++) {
930 ff_opus_rc_enc_end(&s->rc[i], avpkt->data + offset,
931 s->frame[i].framebits >> 3);
932 offset += s->frame[i].framebits >> 3;
935 avpkt->size = offset;
938 /* Used as overlap for the first frame and padding for the last encoded packet */
939 static AVFrame *spawn_empty_frame(OpusEncContext *s)
942 AVFrame *f = av_frame_alloc();
945 f->format = s->avctx->sample_fmt;
946 f->nb_samples = s->avctx->frame_size;
947 f->channel_layout = s->avctx->channel_layout;
948 if (av_frame_get_buffer(f, 4)) {
952 for (i = 0; i < s->channels; i++) {
953 size_t bps = av_get_bytes_per_sample(f->format);
954 memset(f->extended_data[i], 0, bps*f->nb_samples);
959 static int opus_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
960 const AVFrame *frame, int *got_packet_ptr)
962 OpusEncContext *s = avctx->priv_data;
963 int i, ret, frame_size, alloc_size = 0;
965 if (frame) { /* Add new frame to queue */
966 if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
968 ff_bufqueue_add(avctx, &s->bufqueue, av_frame_clone(frame));
970 ff_opus_psy_signal_eof(&s->psyctx);
971 if (!s->afq.remaining_samples)
972 return 0; /* We've been flushed and there's nothing left to encode */
975 /* Run the psychoacoustic system */
976 if (ff_opus_psy_process(&s->psyctx, &s->packet))
979 frame_size = OPUS_BLOCK_SIZE(s->packet.framesize);
982 /* This can go negative, that's not a problem, we only pad if positive */
983 int pad_empty = s->packet.frames*(frame_size/s->avctx->frame_size) - s->bufqueue.available + 1;
984 /* Pad with empty 2.5 ms frames to whatever framesize was decided,
985 * this should only happen at the very last flush frame. The frames
986 * allocated here will be freed (because they have no other references)
987 * after they get used by celt_frame_setup_input() */
988 for (i = 0; i < pad_empty; i++) {
989 AVFrame *empty = spawn_empty_frame(s);
991 return AVERROR(ENOMEM);
992 ff_bufqueue_add(avctx, &s->bufqueue, empty);
996 for (i = 0; i < s->packet.frames; i++) {
997 celt_encode_frame(s, &s->rc[i], &s->frame[i], i);
998 alloc_size += s->frame[i].framebits >> 3;
1001 /* Worst case toc + the frame lengths if needed */
1002 alloc_size += 2 + s->packet.frames*2;
1004 if ((ret = ff_alloc_packet2(avctx, avpkt, alloc_size, 0)) < 0)
1007 /* Assemble packet */
1008 opus_packet_assembler(s, avpkt);
1010 /* Update the psychoacoustic system */
1011 ff_opus_psy_postencode_update(&s->psyctx, s->frame, s->rc);
1013 /* Remove samples from queue and skip if needed */
1014 ff_af_queue_remove(&s->afq, s->packet.frames*frame_size, &avpkt->pts, &avpkt->duration);
1015 if (s->packet.frames*frame_size > avpkt->duration) {
1016 uint8_t *side = av_packet_new_side_data(avpkt, AV_PKT_DATA_SKIP_SAMPLES, 10);
1018 return AVERROR(ENOMEM);
1019 AV_WL32(&side[4], s->packet.frames*frame_size - avpkt->duration + 120);
1022 *got_packet_ptr = 1;
1027 static av_cold int opus_encode_end(AVCodecContext *avctx)
1030 OpusEncContext *s = avctx->priv_data;
1032 for (i = 0; i < CELT_BLOCK_NB; i++)
1033 ff_mdct15_uninit(&s->mdct[i]);
1035 ff_celt_pvq_uninit(&s->pvq);
1037 av_freep(&s->frame);
1039 ff_af_queue_close(&s->afq);
1040 ff_opus_psy_end(&s->psyctx);
1041 ff_bufqueue_discard_all(&s->bufqueue);
1042 av_freep(&avctx->extradata);
1047 static av_cold int opus_encode_init(AVCodecContext *avctx)
1049 int i, ch, ret, max_frames;
1050 OpusEncContext *s = avctx->priv_data;
1053 s->channels = avctx->channels;
1055 /* Opus allows us to change the framesize on each packet (and each packet may
1056 * have multiple frames in it) but we can't change the codec's frame size on
1057 * runtime, so fix it to the lowest possible number of samples and use a queue
1058 * to accumulate AVFrames until we have enough to encode whatever the encoder
1059 * decides is the best */
1060 avctx->frame_size = 120;
1061 /* Initial padding will change if SILK is ever supported */
1062 avctx->initial_padding = 120;
1064 if (!avctx->bit_rate) {
1065 int coupled = ff_opus_default_coupled_streams[s->channels - 1];
1066 avctx->bit_rate = coupled*(96000) + (s->channels - coupled*2)*(48000);
1067 } else if (avctx->bit_rate < 6000 || avctx->bit_rate > 255000 * s->channels) {
1068 int64_t clipped_rate = av_clip(avctx->bit_rate, 6000, 255000 * s->channels);
1069 av_log(avctx, AV_LOG_ERROR, "Unsupported bitrate %"PRId64" kbps, clipping to %"PRId64" kbps\n",
1070 avctx->bit_rate/1000, clipped_rate/1000);
1071 avctx->bit_rate = clipped_rate;
1075 avctx->extradata_size = 19;
1076 avctx->extradata = av_malloc(avctx->extradata_size + AV_INPUT_BUFFER_PADDING_SIZE);
1077 if (!avctx->extradata)
1078 return AVERROR(ENOMEM);
1079 opus_write_extradata(avctx);
1081 ff_af_queue_init(avctx, &s->afq);
1083 if ((ret = ff_celt_pvq_init(&s->pvq)) < 0)
1086 if (!(s->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT)))
1087 return AVERROR(ENOMEM);
1089 /* I have no idea why a base scaling factor of 68 works, could be the twiddles */
1090 for (i = 0; i < CELT_BLOCK_NB; i++)
1091 if ((ret = ff_mdct15_init(&s->mdct[i], 0, i + 3, 68 << (CELT_BLOCK_NB - 1 - i))))
1092 return AVERROR(ENOMEM);
1094 /* Zero out previous energy (matters for inter first frame) */
1095 for (ch = 0; ch < s->channels; ch++)
1096 memset(s->last_quantized_energy[ch], 0.0f, sizeof(float)*CELT_MAX_BANDS);
1098 /* Allocate an empty frame to use as overlap for the first frame of audio */
1099 ff_bufqueue_add(avctx, &s->bufqueue, spawn_empty_frame(s));
1100 if (!ff_bufqueue_peek(&s->bufqueue, 0))
1101 return AVERROR(ENOMEM);
1103 if ((ret = ff_opus_psy_init(&s->psyctx, s->avctx, &s->bufqueue, &s->options)))
1106 /* Frame structs and range coder buffers */
1107 max_frames = ceilf(FFMIN(s->options.max_delay_ms, 120.0f)/2.5f);
1108 s->frame = av_malloc(max_frames*sizeof(CeltFrame));
1110 return AVERROR(ENOMEM);
1111 s->rc = av_malloc(max_frames*sizeof(OpusRangeCoder));
1113 return AVERROR(ENOMEM);
1115 for (i = 0; i < max_frames; i++) {
1116 s->frame[i].dsp = s->dsp;
1117 s->frame[i].avctx = s->avctx;
1118 s->frame[i].seed = 0;
1119 s->frame[i].pvq = s->pvq;
1120 s->frame[i].block[0].emph_coeff = s->frame[i].block[1].emph_coeff = 0.0f;
1126 #define OPUSENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
1127 static const AVOption opusenc_options[] = {
1128 { "opus_delay", "Maximum delay in milliseconds", offsetof(OpusEncContext, options.max_delay_ms), AV_OPT_TYPE_FLOAT, { .dbl = OPUS_MAX_LOOKAHEAD }, 2.5f, OPUS_MAX_LOOKAHEAD, OPUSENC_FLAGS, "max_delay_ms" },
1132 static const AVClass opusenc_class = {
1133 .class_name = "Opus encoder",
1134 .item_name = av_default_item_name,
1135 .option = opusenc_options,
1136 .version = LIBAVUTIL_VERSION_INT,
1139 static const AVCodecDefault opusenc_defaults[] = {
1141 { "compression_level", "10" },
1145 AVCodec ff_opus_encoder = {
1147 .long_name = NULL_IF_CONFIG_SMALL("Opus"),
1148 .type = AVMEDIA_TYPE_AUDIO,
1149 .id = AV_CODEC_ID_OPUS,
1150 .defaults = opusenc_defaults,
1151 .priv_class = &opusenc_class,
1152 .priv_data_size = sizeof(OpusEncContext),
1153 .init = opus_encode_init,
1154 .encode2 = opus_encode_frame,
1155 .close = opus_encode_end,
1156 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE | FF_CODEC_CAP_INIT_CLEANUP,
1157 .capabilities = AV_CODEC_CAP_EXPERIMENTAL | AV_CODEC_CAP_SMALL_LAST_FRAME | AV_CODEC_CAP_DELAY,
1158 .supported_samplerates = (const int []){ 48000, 0 },
1159 .channel_layouts = (const uint64_t []){ AV_CH_LAYOUT_MONO,
1160 AV_CH_LAYOUT_STEREO, 0 },
1161 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
1162 AV_SAMPLE_FMT_NONE },