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
22 #include "opus_celt.h"
26 #include "libavutil/float_dsp.h"
27 #include "libavutil/opt.h"
29 #include "bytestream.h"
30 #include "audio_frame_queue.h"
32 /* Determines the maximum delay the psychoacoustic system will use for lookahead */
33 #define FF_BUFQUEUE_SIZE 145
34 #include "libavfilter/bufferqueue.h"
36 #define OPUS_MAX_LOOKAHEAD ((FF_BUFQUEUE_SIZE - 1)*2.5f)
38 #define OPUS_MAX_CHANNELS 2
40 /* 120 ms / 2.5 ms = 48 frames (extremely improbable, but the encoder'll work) */
41 #define OPUS_MAX_FRAMES_PER_PACKET 48
43 #define OPUS_BLOCK_SIZE(x) (2 * 15 * (1 << ((x) + 2)))
45 #define OPUS_SAMPLES_TO_BLOCK_SIZE(x) (ff_log2((x) / (2 * 15)) - 2)
47 typedef struct OpusEncOptions {
51 typedef struct OpusEncContext {
53 OpusEncOptions options;
54 AVCodecContext *avctx;
56 AVFloatDSPContext *dsp;
57 MDCT15Context *mdct[CELT_BLOCK_NB];
58 struct FFBufQueue bufqueue;
61 enum OpusBandwidth bandwidth;
70 /* Actual energy the decoder will have */
71 float last_quantized_energy[OPUS_MAX_CHANNELS][CELT_MAX_BANDS];
73 DECLARE_ALIGNED(32, float, scratch)[2048];
76 static void opus_write_extradata(AVCodecContext *avctx)
78 uint8_t *bs = avctx->extradata;
80 bytestream_put_buffer(&bs, "OpusHead", 8);
81 bytestream_put_byte (&bs, 0x1);
82 bytestream_put_byte (&bs, avctx->channels);
83 bytestream_put_le16 (&bs, avctx->initial_padding);
84 bytestream_put_le32 (&bs, avctx->sample_rate);
85 bytestream_put_le16 (&bs, 0x0);
86 bytestream_put_byte (&bs, 0x0); /* Default layout */
89 static int opus_gen_toc(OpusEncContext *s, uint8_t *toc, int *size, int *fsize_needed)
91 int i, tmp = 0x0, extended_toc = 0;
92 static const int toc_cfg[][OPUS_MODE_NB][OPUS_BANDWITH_NB] = {
93 /* Silk Hybrid Celt Layer */
94 /* NB MB WB SWB FB NB MB WB SWB FB NB MB WB SWB FB Bandwidth */
95 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 17, 0, 21, 25, 29 } }, /* 2.5 ms */
96 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 18, 0, 22, 26, 30 } }, /* 5 ms */
97 { { 1, 5, 9, 0, 0 }, { 0, 0, 0, 13, 15 }, { 19, 0, 23, 27, 31 } }, /* 10 ms */
98 { { 2, 6, 10, 0, 0 }, { 0, 0, 0, 14, 16 }, { 20, 0, 24, 28, 32 } }, /* 20 ms */
99 { { 3, 7, 11, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 40 ms */
100 { { 4, 8, 12, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 60 ms */
102 int cfg = toc_cfg[s->pkt_framesize][s->mode][s->bandwidth];
106 if (s->pkt_frames == 2) { /* 2 packets */
107 if (s->frame[0].framebits == s->frame[1].framebits) { /* same size */
109 } else { /* different size */
111 *fsize_needed = 1; /* put frame sizes in the packet */
113 } else if (s->pkt_frames > 2) {
117 tmp |= (s->channels > 1) << 2; /* Stereo or mono */
118 tmp |= (cfg - 1) << 3; /* codec configuration */
121 for (i = 0; i < (s->pkt_frames - 1); i++)
122 *fsize_needed |= (s->frame[i].framebits != s->frame[i + 1].framebits);
123 tmp = (*fsize_needed) << 7; /* vbr flag */
124 tmp |= s->pkt_frames; /* frame number - can be 0 as well */
127 *size = 1 + extended_toc;
131 static void celt_frame_setup_input(OpusEncContext *s, CeltFrame *f)
135 const int subframesize = s->avctx->frame_size;
136 int subframes = OPUS_BLOCK_SIZE(s->pkt_framesize) / subframesize;
138 cur = ff_bufqueue_get(&s->bufqueue);
140 for (ch = 0; ch < f->channels; ch++) {
141 CeltBlock *b = &f->block[ch];
142 const void *input = cur->extended_data[ch];
143 size_t bps = av_get_bytes_per_sample(cur->format);
144 memcpy(b->overlap, input, bps*cur->nb_samples);
149 for (sf = 0; sf < subframes; sf++) {
150 if (sf != (subframes - 1))
151 cur = ff_bufqueue_get(&s->bufqueue);
153 cur = ff_bufqueue_peek(&s->bufqueue, 0);
155 for (ch = 0; ch < f->channels; ch++) {
156 CeltBlock *b = &f->block[ch];
157 const void *input = cur->extended_data[ch];
158 const size_t bps = av_get_bytes_per_sample(cur->format);
159 const size_t left = (subframesize - cur->nb_samples)*bps;
160 const size_t len = FFMIN(subframesize, cur->nb_samples)*bps;
161 memcpy(&b->samples[sf*subframesize], input, len);
162 memset(&b->samples[cur->nb_samples], 0, left);
165 /* Last frame isn't popped off and freed yet - we need it for overlap */
166 if (sf != (subframes - 1))
171 /* Apply the pre emphasis filter */
172 static void celt_apply_preemph_filter(OpusEncContext *s, CeltFrame *f)
175 const int subframesize = s->avctx->frame_size;
176 const int subframes = OPUS_BLOCK_SIZE(s->pkt_framesize) / subframesize;
179 for (ch = 0; ch < f->channels; ch++) {
180 CeltBlock *b = &f->block[ch];
181 float m = b->emph_coeff;
182 for (i = 0; i < CELT_OVERLAP; i++) {
183 float sample = b->overlap[i];
184 b->overlap[i] = sample - m;
185 m = sample * CELT_EMPH_COEFF;
190 /* Filter the samples but do not update the last subframe's coeff - overlap ^^^ */
191 for (sf = 0; sf < subframes; sf++) {
192 for (ch = 0; ch < f->channels; ch++) {
193 CeltBlock *b = &f->block[ch];
194 float m = b->emph_coeff;
195 for (i = 0; i < subframesize; i++) {
196 float sample = b->samples[sf*subframesize + i];
197 b->samples[sf*subframesize + i] = sample - m;
198 m = sample * CELT_EMPH_COEFF;
200 if (sf != (subframes - 1))
206 /* Create the window and do the mdct */
207 static void celt_frame_mdct(OpusEncContext *s, CeltFrame *f)
210 float *win = s->scratch;
212 /* I think I can use s->dsp->vector_fmul_window for transients at least */
214 for (ch = 0; ch < f->channels; ch++) {
215 CeltBlock *b = &f->block[ch];
216 float *src1 = b->overlap;
217 for (t = 0; t < f->blocks; t++) {
218 float *src2 = &b->samples[CELT_OVERLAP*t];
219 for (i = 0; i < CELT_OVERLAP; i++) {
220 win[ i] = src1[i]*ff_celt_window[i];
221 win[CELT_OVERLAP + i] = src2[i]*ff_celt_window[CELT_OVERLAP - i - 1];
224 s->mdct[0]->mdct(s->mdct[0], b->coeffs + t, win, f->blocks);
228 int blk_len = OPUS_BLOCK_SIZE(f->size), wlen = OPUS_BLOCK_SIZE(f->size + 1);
229 int rwin = blk_len - CELT_OVERLAP, lap_dst = (wlen - blk_len - CELT_OVERLAP) >> 1;
230 for (ch = 0; ch < f->channels; ch++) {
231 CeltBlock *b = &f->block[ch];
233 memset(win, 0, wlen*sizeof(float));
235 memcpy(&win[lap_dst + CELT_OVERLAP], b->samples, rwin*sizeof(float));
237 /* Alignment fucks me over */
238 //s->dsp->vector_fmul(&dst[lap_dst], b->overlap, ff_celt_window, CELT_OVERLAP);
239 //s->dsp->vector_fmul_reverse(&dst[lap_dst + blk_len - CELT_OVERLAP], b->samples, ff_celt_window, CELT_OVERLAP);
241 for (i = 0; i < CELT_OVERLAP; i++) {
242 win[lap_dst + i] = b->overlap[i] *ff_celt_window[i];
243 win[lap_dst + blk_len + i] = b->samples[rwin + i]*ff_celt_window[CELT_OVERLAP - i - 1];
246 s->mdct[f->size]->mdct(s->mdct[f->size], b->coeffs, win, 1);
251 /* Fills the bands and normalizes them */
252 static int celt_frame_map_norm_bands(OpusEncContext *s, CeltFrame *f)
254 int i, j, ch, noise = 0;
256 for (ch = 0; ch < f->channels; ch++) {
257 CeltBlock *block = &f->block[ch];
258 float *start = block->coeffs;
259 for (i = 0; i < CELT_MAX_BANDS; i++) {
262 /* Calculate band bins */
263 block->band_bins[i] = ff_celt_freq_range[i] << f->size;
264 block->band_coeffs[i] = start;
265 start += block->band_bins[i];
267 /* Normalize band energy */
268 for (j = 0; j < block->band_bins[i]; j++)
269 ener += block->band_coeffs[i][j]*block->band_coeffs[i][j];
271 block->lin_energy[i] = sqrtf(ener) + FLT_EPSILON;
272 ener = 1.0f/block->lin_energy[i];
274 for (j = 0; j < block->band_bins[i]; j++)
275 block->band_coeffs[i][j] *= ener;
277 block->energy[i] = log2f(block->lin_energy[i]) - ff_celt_mean_energy[i];
279 /* CELT_ENERGY_SILENCE is what the decoder uses and its not -infinity */
280 block->energy[i] = FFMAX(block->energy[i], CELT_ENERGY_SILENCE);
281 noise |= block->energy[i] > CELT_ENERGY_SILENCE;
287 static void celt_enc_tf(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f)
289 int i, tf_select = 0, diff = 0, tf_changed = 0, tf_select_needed;
290 int bits = f->transient ? 2 : 4;
292 tf_select_needed = ((f->size && (opus_rc_tell(rc) + bits + 1) <= f->framebits));
294 for (i = f->start_band; i < f->end_band; i++) {
295 if ((opus_rc_tell(rc) + bits + tf_select_needed) <= f->framebits) {
296 const int tbit = (diff ^ 1) == f->tf_change[i];
297 ff_opus_rc_enc_log(rc, tbit, bits);
301 bits = f->transient ? 4 : 5;
304 if (tf_select_needed && ff_celt_tf_select[f->size][f->transient][0][tf_changed] !=
305 ff_celt_tf_select[f->size][f->transient][1][tf_changed]) {
306 ff_opus_rc_enc_log(rc, f->tf_select, 1);
307 tf_select = f->tf_select;
310 for (i = f->start_band; i < f->end_band; i++)
311 f->tf_change[i] = ff_celt_tf_select[f->size][f->transient][tf_select][f->tf_change[i]];
314 static void celt_bitalloc(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f)
316 int i, j, low, high, total, done, bandbits, remaining, tbits_8ths;
317 int skip_startband = f->start_band;
319 int intensitystereo_bit = 0;
320 int dualstereo_bit = 0;
325 int boost[CELT_MAX_BANDS];
326 int trim_offset[CELT_MAX_BANDS];
327 int threshold[CELT_MAX_BANDS];
328 int bits1[CELT_MAX_BANDS];
329 int bits2[CELT_MAX_BANDS];
331 /* Tell the spread to the decoder */
332 if (opus_rc_tell(rc) + 4 <= f->framebits)
333 ff_opus_rc_enc_cdf(rc, f->spread, ff_celt_model_spread);
335 /* Generate static allocation caps */
336 for (i = 0; i < CELT_MAX_BANDS; i++) {
337 cap[i] = (ff_celt_static_caps[f->size][f->channels - 1][i] + 64)
338 * ff_celt_freq_range[i] << (f->channels - 1) << f->size >> 2;
342 tbits_8ths = f->framebits << 3;
343 for (i = f->start_band; i < f->end_band; i++) {
344 int quanta, b_dynalloc, boost_amount = f->alloc_boost[i];
348 quanta = ff_celt_freq_range[i] << (f->channels - 1) << f->size;
349 quanta = FFMIN(quanta << 3, FFMAX(6 << 3, quanta));
350 b_dynalloc = dynalloc;
352 while (opus_rc_tell_frac(rc) + (b_dynalloc << 3) < tbits_8ths && boost[i] < cap[i]) {
353 int is_boost = boost_amount--;
355 ff_opus_rc_enc_log(rc, is_boost, b_dynalloc);
360 tbits_8ths -= quanta;
366 dynalloc = FFMAX(2, dynalloc - 1);
369 /* Put allocation trim */
370 if (opus_rc_tell_frac(rc) + (6 << 3) <= tbits_8ths)
371 ff_opus_rc_enc_cdf(rc, f->alloc_trim, ff_celt_model_alloc_trim);
373 /* Anti-collapse bit reservation */
374 tbits_8ths = (f->framebits << 3) - opus_rc_tell_frac(rc) - 1;
375 f->anticollapse_needed = 0;
376 if (f->transient && f->size >= 2 && tbits_8ths >= ((f->size + 2) << 3))
377 f->anticollapse_needed = 1 << 3;
378 tbits_8ths -= f->anticollapse_needed;
380 /* Band skip bit reservation */
381 if (tbits_8ths >= 1 << 3)
383 tbits_8ths -= skip_bit;
385 /* Intensity/dual stereo bit reservation */
386 if (f->channels == 2) {
387 intensitystereo_bit = ff_celt_log2_frac[f->end_band - f->start_band];
388 if (intensitystereo_bit <= tbits_8ths) {
389 tbits_8ths -= intensitystereo_bit;
390 if (tbits_8ths >= 1 << 3) {
391 dualstereo_bit = 1 << 3;
392 tbits_8ths -= 1 << 3;
395 intensitystereo_bit = 0;
400 for (i = f->start_band; i < f->end_band; i++) {
401 int trim = f->alloc_trim - 5 - f->size;
402 int band = ff_celt_freq_range[i] * (f->end_band - i - 1);
403 int duration = f->size + 3;
404 int scale = duration + f->channels - 1;
406 /* PVQ minimum allocation threshold, below this value the band is
408 threshold[i] = FFMAX(3 * ff_celt_freq_range[i] << duration >> 4,
411 trim_offset[i] = trim * (band << scale) >> 6;
413 if (ff_celt_freq_range[i] << f->size == 1)
414 trim_offset[i] -= f->channels << 3;
419 high = CELT_VECTORS - 1;
420 while (low <= high) {
421 int center = (low + high) >> 1;
424 for (i = f->end_band - 1; i >= f->start_band; i--) {
425 bandbits = ff_celt_freq_range[i] * ff_celt_static_alloc[center][i]
426 << (f->channels - 1) << f->size >> 2;
429 bandbits = FFMAX(0, bandbits + trim_offset[i]);
430 bandbits += boost[i];
432 if (bandbits >= threshold[i] || done) {
434 total += FFMIN(bandbits, cap[i]);
435 } else if (bandbits >= f->channels << 3)
436 total += f->channels << 3;
439 if (total > tbits_8ths)
447 for (i = f->start_band; i < f->end_band; i++) {
448 bits1[i] = ff_celt_freq_range[i] * ff_celt_static_alloc[low][i]
449 << (f->channels - 1) << f->size >> 2;
450 bits2[i] = high >= CELT_VECTORS ? cap[i] :
451 ff_celt_freq_range[i] * ff_celt_static_alloc[high][i]
452 << (f->channels - 1) << f->size >> 2;
455 bits1[i] = FFMAX(0, bits1[i] + trim_offset[i]);
457 bits2[i] = FFMAX(0, bits2[i] + trim_offset[i]);
459 bits1[i] += boost[i];
460 bits2[i] += boost[i];
464 bits2[i] = FFMAX(0, bits2[i] - bits1[i]);
469 high = 1 << CELT_ALLOC_STEPS;
470 for (i = 0; i < CELT_ALLOC_STEPS; i++) {
471 int center = (low + high) >> 1;
474 for (j = f->end_band - 1; j >= f->start_band; j--) {
475 bandbits = bits1[j] + (center * bits2[j] >> CELT_ALLOC_STEPS);
477 if (bandbits >= threshold[j] || done) {
479 total += FFMIN(bandbits, cap[j]);
480 } else if (bandbits >= f->channels << 3)
481 total += f->channels << 3;
483 if (total > tbits_8ths)
491 for (i = f->end_band - 1; i >= f->start_band; i--) {
492 bandbits = bits1[i] + (low * bits2[i] >> CELT_ALLOC_STEPS);
494 if (bandbits >= threshold[i] || done)
497 bandbits = (bandbits >= f->channels << 3) ?
498 f->channels << 3 : 0;
500 bandbits = FFMIN(bandbits, cap[i]);
501 f->pulses[i] = bandbits;
506 for (f->coded_bands = f->end_band; ; f->coded_bands--) {
508 j = f->coded_bands - 1;
510 if (j == skip_startband) {
511 /* all remaining bands are not skipped */
512 tbits_8ths += skip_bit;
516 /* determine the number of bits available for coding "do not skip" markers */
517 remaining = tbits_8ths - total;
518 bandbits = remaining / (ff_celt_freq_bands[j+1] - ff_celt_freq_bands[f->start_band]);
519 remaining -= bandbits * (ff_celt_freq_bands[j+1] - ff_celt_freq_bands[f->start_band]);
520 allocation = f->pulses[j] + bandbits * ff_celt_freq_range[j]
521 + FFMAX(0, remaining - (ff_celt_freq_bands[j] - ff_celt_freq_bands[f->start_band]));
523 /* a "do not skip" marker is only coded if the allocation is
524 above the chosen threshold */
525 if (allocation >= FFMAX(threshold[j], (f->channels + 1) << 3)) {
526 const int do_not_skip = f->coded_bands <= f->skip_band_floor;
527 ff_opus_rc_enc_log(rc, do_not_skip, 1);
532 allocation -= 1 << 3;
535 /* the band is skipped, so reclaim its bits */
536 total -= f->pulses[j];
537 if (intensitystereo_bit) {
538 total -= intensitystereo_bit;
539 intensitystereo_bit = ff_celt_log2_frac[j - f->start_band];
540 total += intensitystereo_bit;
543 total += f->pulses[j] = (allocation >= f->channels << 3) ? f->channels << 3 : 0;
546 /* Encode stereo flags */
547 if (intensitystereo_bit) {
548 f->intensity_stereo = FFMIN(f->intensity_stereo, f->coded_bands);
549 ff_opus_rc_enc_uint(rc, f->intensity_stereo, f->coded_bands + 1 - f->start_band);
551 if (f->intensity_stereo <= f->start_band)
552 tbits_8ths += dualstereo_bit; /* no intensity stereo means no dual stereo */
553 else if (dualstereo_bit)
554 ff_opus_rc_enc_log(rc, f->dual_stereo, 1);
556 /* Supply the remaining bits in this frame to lower bands */
557 remaining = tbits_8ths - total;
558 bandbits = remaining / (ff_celt_freq_bands[f->coded_bands] - ff_celt_freq_bands[f->start_band]);
559 remaining -= bandbits * (ff_celt_freq_bands[f->coded_bands] - ff_celt_freq_bands[f->start_band]);
560 for (i = f->start_band; i < f->coded_bands; i++) {
561 int bits = FFMIN(remaining, ff_celt_freq_range[i]);
563 f->pulses[i] += bits + bandbits * ff_celt_freq_range[i];
567 /* Finally determine the allocation */
568 for (i = f->start_band; i < f->coded_bands; i++) {
569 int N = ff_celt_freq_range[i] << f->size;
570 int prev_extra = extrabits;
571 f->pulses[i] += extrabits;
574 int dof; // degrees of freedom
575 int temp; // dof * channels * log(dof)
576 int offset; // fine energy quantization offset, i.e.
577 // extra bits assigned over the standard
579 int fine_bits, max_bits;
581 extrabits = FFMAX(0, f->pulses[i] - cap[i]);
582 f->pulses[i] -= extrabits;
584 /* intensity stereo makes use of an extra degree of freedom */
585 dof = N * f->channels + (f->channels == 2 && N > 2 && !f->dual_stereo && i < f->intensity_stereo);
586 temp = dof * (ff_celt_log_freq_range[i] + (f->size << 3));
587 offset = (temp >> 1) - dof * CELT_FINE_OFFSET;
588 if (N == 2) /* dof=2 is the only case that doesn't fit the model */
591 /* grant an additional bias for the first and second pulses */
592 if (f->pulses[i] + offset < 2 * (dof << 3))
594 else if (f->pulses[i] + offset < 3 * (dof << 3))
597 fine_bits = (f->pulses[i] + offset + (dof << 2)) / (dof << 3);
598 max_bits = FFMIN((f->pulses[i] >> 3) >> (f->channels - 1), CELT_MAX_FINE_BITS);
600 max_bits = FFMAX(max_bits, 0);
602 f->fine_bits[i] = av_clip(fine_bits, 0, max_bits);
604 /* if fine_bits was rounded down or capped,
605 give priority for the final fine energy pass */
606 f->fine_priority[i] = (f->fine_bits[i] * (dof << 3) >= f->pulses[i] + offset);
608 /* the remaining bits are assigned to PVQ */
609 f->pulses[i] -= f->fine_bits[i] << (f->channels - 1) << 3;
611 /* all bits go to fine energy except for the sign bit */
612 extrabits = FFMAX(0, f->pulses[i] - (f->channels << 3));
613 f->pulses[i] -= extrabits;
615 f->fine_priority[i] = 1;
618 /* hand back a limited number of extra fine energy bits to this band */
620 int fineextra = FFMIN(extrabits >> (f->channels + 2),
621 CELT_MAX_FINE_BITS - f->fine_bits[i]);
622 f->fine_bits[i] += fineextra;
624 fineextra <<= f->channels + 2;
625 f->fine_priority[i] = (fineextra >= extrabits - prev_extra);
626 extrabits -= fineextra;
629 f->remaining = extrabits;
631 /* skipped bands dedicate all of their bits for fine energy */
632 for (; i < f->end_band; i++) {
633 f->fine_bits[i] = f->pulses[i] >> (f->channels - 1) >> 3;
635 f->fine_priority[i] = f->fine_bits[i] < 1;
639 static void celt_quant_coarse(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f)
642 float alpha, beta, prev[2] = { 0, 0 };
643 const uint8_t *pmod = ff_celt_coarse_energy_dist[f->size][f->intra];
645 /* Inter is really just differential coding */
646 if (opus_rc_tell(rc) + 3 <= f->framebits)
647 ff_opus_rc_enc_log(rc, f->intra, 3);
653 beta = 1.0f - 4915.0f/32768.0f;
655 alpha = ff_celt_alpha_coef[f->size];
656 beta = 1.0f - ff_celt_beta_coef[f->size];
659 for (i = f->start_band; i < f->end_band; i++) {
660 for (ch = 0; ch < f->channels; ch++) {
661 CeltBlock *block = &f->block[ch];
662 const int left = f->framebits - opus_rc_tell(rc);
663 const float last = FFMAX(-9.0f, s->last_quantized_energy[ch][i]);
664 float diff = block->energy[i] - prev[ch] - last*alpha;
665 int q_en = lrintf(diff);
667 ff_opus_rc_enc_laplace(rc, &q_en, pmod[i << 1] << 7, pmod[(i << 1) + 1] << 6);
668 } else if (left >= 2) {
669 q_en = av_clip(q_en, -1, 1);
670 ff_opus_rc_enc_cdf(rc, 2*q_en + 3*(q_en < 0), ff_celt_model_energy_small);
671 } else if (left >= 1) {
672 q_en = av_clip(q_en, -1, 0);
673 ff_opus_rc_enc_log(rc, (q_en & 1), 1);
676 block->error_energy[i] = q_en - diff;
677 prev[ch] += beta * q_en;
682 static void celt_quant_fine(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f)
685 for (i = f->start_band; i < f->end_band; i++) {
686 if (!f->fine_bits[i])
688 for (ch = 0; ch < f->channels; ch++) {
689 CeltBlock *block = &f->block[ch];
690 int quant, lim = (1 << f->fine_bits[i]);
691 float offset, diff = 0.5f - block->error_energy[i];
692 quant = av_clip(floor(diff*lim), 0, lim - 1);
693 ff_opus_rc_put_raw(rc, quant, f->fine_bits[i]);
694 offset = 0.5f - ((quant + 0.5f) * (1 << (14 - f->fine_bits[i])) / 16384.0f);
695 block->error_energy[i] -= offset;
700 static void celt_quant_final(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f)
703 for (priority = 0; priority < 2; priority++) {
704 for (i = f->start_band; i < f->end_band && (f->framebits - opus_rc_tell(rc)) >= f->channels; i++) {
705 if (f->fine_priority[i] != priority || f->fine_bits[i] >= CELT_MAX_FINE_BITS)
707 for (ch = 0; ch < f->channels; ch++) {
708 CeltBlock *block = &f->block[ch];
709 const float err = block->error_energy[i];
710 const float offset = 0.5f * (1 << (14 - f->fine_bits[i] - 1)) / 16384.0f;
711 const int sign = FFABS(err + offset) < FFABS(err - offset);
712 ff_opus_rc_put_raw(rc, sign, 1);
713 block->error_energy[i] -= offset*(1 - 2*sign);
719 static void celt_quant_bands(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f)
721 float lowband_scratch[8 * 22];
722 float norm[2 * 8 * 100];
724 int totalbits = (f->framebits << 3) - f->anticollapse_needed;
726 int update_lowband = 1;
727 int lowband_offset = 0;
731 for (i = f->start_band; i < f->end_band; i++) {
732 int band_offset = ff_celt_freq_bands[i] << f->size;
733 int band_size = ff_celt_freq_range[i] << f->size;
734 float *X = f->block[0].coeffs + band_offset;
735 float *Y = (f->channels == 2) ? f->block[1].coeffs + band_offset : NULL;
737 int consumed = opus_rc_tell_frac(rc);
738 float *norm2 = norm + 8 * 100;
739 int effective_lowband = -1;
743 /* Compute how many bits we want to allocate to this band */
744 if (i != f->start_band)
745 f->remaining -= consumed;
746 f->remaining2 = totalbits - consumed - 1;
747 if (i <= f->coded_bands - 1) {
748 int curr_balance = f->remaining / FFMIN(3, f->coded_bands-i);
749 b = av_clip_uintp2(FFMIN(f->remaining2 + 1, f->pulses[i] + curr_balance), 14);
753 if (ff_celt_freq_bands[i] - ff_celt_freq_range[i] >= ff_celt_freq_bands[f->start_band] &&
754 (update_lowband || lowband_offset == 0))
757 /* Get a conservative estimate of the collapse_mask's for the bands we're
758 going to be folding from. */
759 if (lowband_offset != 0 && (f->spread != CELT_SPREAD_AGGRESSIVE ||
760 f->blocks > 1 || f->tf_change[i] < 0)) {
761 int foldstart, foldend;
763 /* This ensures we never repeat spectral content within one band */
764 effective_lowband = FFMAX(ff_celt_freq_bands[f->start_band],
765 ff_celt_freq_bands[lowband_offset] - ff_celt_freq_range[i]);
766 foldstart = lowband_offset;
767 while (ff_celt_freq_bands[--foldstart] > effective_lowband);
768 foldend = lowband_offset - 1;
769 while (ff_celt_freq_bands[++foldend] < effective_lowband + ff_celt_freq_range[i]);
772 for (j = foldstart; j < foldend; j++) {
773 cm[0] |= f->block[0].collapse_masks[j];
774 cm[1] |= f->block[f->channels - 1].collapse_masks[j];
777 /* Otherwise, we'll be using the LCG to fold, so all blocks will (almost
778 always) be non-zero.*/
779 cm[0] = cm[1] = (1 << f->blocks) - 1;
781 if (f->dual_stereo && i == f->intensity_stereo) {
782 /* Switch off dual stereo to do intensity */
784 for (j = ff_celt_freq_bands[f->start_band] << f->size; j < band_offset; j++)
785 norm[j] = (norm[j] + norm2[j]) / 2;
788 if (f->dual_stereo) {
789 cm[0] = ff_celt_encode_band(f, rc, i, X, NULL, band_size, b / 2, f->blocks,
790 effective_lowband != -1 ? norm + (effective_lowband << f->size) : NULL, f->size,
791 norm + band_offset, 0, 1.0f, lowband_scratch, cm[0]);
793 cm[1] = ff_celt_encode_band(f, rc, i, Y, NULL, band_size, b/2, f->blocks,
794 effective_lowband != -1 ? norm2 + (effective_lowband << f->size) : NULL, f->size,
795 norm2 + band_offset, 0, 1.0f, lowband_scratch, cm[1]);
797 cm[0] = ff_celt_encode_band(f, rc, i, X, Y, band_size, b, f->blocks,
798 effective_lowband != -1 ? norm + (effective_lowband << f->size) : NULL, f->size,
799 norm + band_offset, 0, 1.0f, lowband_scratch, cm[0]|cm[1]);
803 f->block[0].collapse_masks[i] = (uint8_t)cm[0];
804 f->block[f->channels - 1].collapse_masks[i] = (uint8_t)cm[1];
805 f->remaining += f->pulses[i] + consumed;
807 /* Update the folding position only as long as we have 1 bit/sample depth */
808 update_lowband = (b > band_size << 3);
812 static void celt_encode_frame(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f)
816 celt_frame_setup_input(s, f);
817 celt_apply_preemph_filter(s, f);
819 /* Not implemented */
821 celt_frame_mdct(s, f);
822 f->silence = celt_frame_map_norm_bands(s, f);
828 ff_opus_rc_enc_log(rc, f->silence, 15);
830 if (!f->start_band && opus_rc_tell(rc) + 16 <= f->framebits)
831 ff_opus_rc_enc_log(rc, f->pfilter, 1);
834 /* Not implemented */
837 if (f->size && opus_rc_tell(rc) + 3 <= f->framebits)
838 ff_opus_rc_enc_log(rc, f->transient, 3);
840 celt_quant_coarse (s, rc, f);
841 celt_enc_tf (s, rc, f);
842 celt_bitalloc (s, rc, f);
843 celt_quant_fine (s, rc, f);
844 celt_quant_bands (s, rc, f);
846 if (f->anticollapse_needed)
847 ff_opus_rc_put_raw(rc, f->anticollapse, 1);
849 celt_quant_final(s, rc, f);
851 for (ch = 0; ch < f->channels; ch++) {
852 CeltBlock *block = &f->block[ch];
853 for (i = 0; i < CELT_MAX_BANDS; i++)
854 s->last_quantized_energy[ch][i] = block->energy[i] + block->error_energy[i];
858 static void ff_opus_psy_process(OpusEncContext *s, int end, int *need_more)
860 int max_delay_samples = (s->options.max_delay_ms*s->avctx->sample_rate)/1000;
861 int max_bsize = FFMIN(OPUS_SAMPLES_TO_BLOCK_SIZE(max_delay_samples), CELT_BLOCK_960);
864 s->pkt_framesize = max_bsize;
865 s->mode = OPUS_MODE_CELT;
866 s->bandwidth = OPUS_BANDWIDTH_FULLBAND;
868 *need_more = s->bufqueue.available*s->avctx->frame_size < (max_delay_samples + CELT_OVERLAP);
869 /* Don't request more if we start being flushed with NULL frames */
870 *need_more = !end && *need_more;
873 static void ff_opus_psy_celt_frame_setup(OpusEncContext *s, CeltFrame *f, int index)
875 int frame_size = OPUS_BLOCK_SIZE(s->pkt_framesize);
879 f->start_band = (s->mode == OPUS_MODE_HYBRID) ? 17 : 0;
880 f->end_band = ff_celt_band_end[s->bandwidth];
881 f->channels = s->channels;
882 f->size = s->pkt_framesize;
892 f->skip_band_floor = f->end_band;
893 f->intensity_stereo = f->end_band;
895 f->spread = CELT_SPREAD_NORMAL;
896 memset(f->tf_change, 0, sizeof(int)*CELT_MAX_BANDS);
897 memset(f->alloc_boost, 0, sizeof(int)*CELT_MAX_BANDS);
899 f->blocks = f->transient ? frame_size/CELT_OVERLAP : 1;
900 f->framebits = FFALIGN(lrintf((double)s->avctx->bit_rate/(s->avctx->sample_rate/frame_size)), 8);
903 static void opus_packet_assembler(OpusEncContext *s, AVPacket *avpkt)
905 int i, offset, fsize_needed;
908 opus_gen_toc(s, avpkt->data, &offset, &fsize_needed);
910 for (i = 0; i < s->pkt_frames; i++) {
911 ff_opus_rc_enc_end(&s->rc[i], avpkt->data + offset, s->frame[i].framebits >> 3);
912 offset += s->frame[i].framebits >> 3;
915 avpkt->size = offset;
918 /* Used as overlap for the first frame and padding for the last encoded packet */
919 static AVFrame *spawn_empty_frame(OpusEncContext *s)
922 AVFrame *f = av_frame_alloc();
925 f->format = s->avctx->sample_fmt;
926 f->nb_samples = s->avctx->frame_size;
927 f->channel_layout = s->avctx->channel_layout;
928 if (av_frame_get_buffer(f, 4)) {
932 for (i = 0; i < s->channels; i++) {
933 size_t bps = av_get_bytes_per_sample(f->format);
934 memset(f->extended_data[i], 0, bps*f->nb_samples);
939 static int opus_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
940 const AVFrame *frame, int *got_packet_ptr)
942 OpusEncContext *s = avctx->priv_data;
943 int i, ret, frame_size, need_more, alloc_size = 0;
945 if (frame) { /* Add new frame to queue */
946 if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
948 ff_bufqueue_add(avctx, &s->bufqueue, av_frame_clone(frame));
950 if (!s->afq.remaining_samples)
951 return 0; /* We've been flushed and there's nothing left to encode */
954 /* Run the psychoacoustic system */
955 ff_opus_psy_process(s, !frame, &need_more);
957 /* Get more samples for lookahead/encoding */
961 frame_size = OPUS_BLOCK_SIZE(s->pkt_framesize);
964 /* This can go negative, that's not a problem, we only pad if positive */
965 int pad_empty = s->pkt_frames*(frame_size/s->avctx->frame_size) - s->bufqueue.available + 1;
966 /* Pad with empty 2.5 ms frames to whatever framesize was decided,
967 * this should only happen at the very last flush frame. The frames
968 * allocated here will be freed (because they have no other references)
969 * after they get used by celt_frame_setup_input() */
970 for (i = 0; i < pad_empty; i++) {
971 AVFrame *empty = spawn_empty_frame(s);
973 return AVERROR(ENOMEM);
974 ff_bufqueue_add(avctx, &s->bufqueue, empty);
978 for (i = 0; i < s->pkt_frames; i++) {
979 ff_opus_rc_enc_init(&s->rc[i]);
980 ff_opus_psy_celt_frame_setup(s, &s->frame[i], i);
981 celt_encode_frame(s, &s->rc[i], &s->frame[i]);
982 alloc_size += s->frame[i].framebits >> 3;
985 /* Worst case toc + the frame lengths if needed */
986 alloc_size += 2 + s->pkt_frames*2;
988 if ((ret = ff_alloc_packet2(avctx, avpkt, alloc_size, 0)) < 0)
991 /* Assemble packet */
992 opus_packet_assembler(s, avpkt);
994 /* Remove samples from queue and skip if needed */
995 ff_af_queue_remove(&s->afq, s->pkt_frames*frame_size, &avpkt->pts, &avpkt->duration);
996 if (s->pkt_frames*frame_size > avpkt->duration) {
997 uint8_t *side = av_packet_new_side_data(avpkt, AV_PKT_DATA_SKIP_SAMPLES, 10);
999 return AVERROR(ENOMEM);
1000 AV_WL32(&side[4], s->pkt_frames*frame_size - avpkt->duration + 120);
1003 *got_packet_ptr = 1;
1008 static av_cold int opus_encode_end(AVCodecContext *avctx)
1011 OpusEncContext *s = avctx->priv_data;
1013 for (i = 0; i < CELT_BLOCK_NB; i++)
1014 ff_mdct15_uninit(&s->mdct[i]);
1017 av_freep(&s->frame);
1019 ff_af_queue_close(&s->afq);
1020 ff_bufqueue_discard_all(&s->bufqueue);
1021 av_freep(&avctx->extradata);
1026 static av_cold int opus_encode_init(AVCodecContext *avctx)
1029 OpusEncContext *s = avctx->priv_data;
1032 s->channels = avctx->channels;
1034 /* Opus allows us to change the framesize on each packet (and each packet may
1035 * have multiple frames in it) but we can't change the codec's frame size on
1036 * runtime, so fix it to the lowest possible number of samples and use a queue
1037 * to accumulate AVFrames until we have enough to encode whatever the encoder
1038 * decides is the best */
1039 avctx->frame_size = 120;
1040 /* Initial padding will change if SILK is ever supported */
1041 avctx->initial_padding = 120;
1043 avctx->cutoff = !avctx->cutoff ? 20000 : avctx->cutoff;
1045 if (!avctx->bit_rate) {
1046 int coupled = ff_opus_default_coupled_streams[s->channels - 1];
1047 avctx->bit_rate = coupled*(96000) + (s->channels - coupled*2)*(48000);
1048 } else if (avctx->bit_rate < 6000 || avctx->bit_rate > 255000 * s->channels) {
1049 int64_t clipped_rate = av_clip(avctx->bit_rate, 6000, 255000 * s->channels);
1050 av_log(avctx, AV_LOG_ERROR, "Unsupported bitrate %"PRId64" kbps, clipping to %"PRId64" kbps\n",
1051 avctx->bit_rate/1000, clipped_rate/1000);
1052 avctx->bit_rate = clipped_rate;
1055 /* Frame structs and range coder buffers */
1056 s->frame = av_malloc(OPUS_MAX_FRAMES_PER_PACKET*sizeof(CeltFrame));
1058 return AVERROR(ENOMEM);
1059 s->rc = av_malloc(OPUS_MAX_FRAMES_PER_PACKET*sizeof(OpusRangeCoder));
1061 return AVERROR(ENOMEM);
1064 avctx->extradata_size = 19;
1065 avctx->extradata = av_malloc(avctx->extradata_size + AV_INPUT_BUFFER_PADDING_SIZE);
1066 if (!avctx->extradata)
1067 return AVERROR(ENOMEM);
1068 opus_write_extradata(avctx);
1070 ff_af_queue_init(avctx, &s->afq);
1072 if (!(s->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT)))
1073 return AVERROR(ENOMEM);
1075 /* I have no idea why a base scaling factor of 68 works, could be the twiddles */
1076 for (i = 0; i < CELT_BLOCK_NB; i++)
1077 if ((ret = ff_mdct15_init(&s->mdct[i], 0, i + 3, 68 << (CELT_BLOCK_NB - 1 - i))))
1078 return AVERROR(ENOMEM);
1080 for (i = 0; i < OPUS_MAX_FRAMES_PER_PACKET; i++)
1081 s->frame[i].block[0].emph_coeff = s->frame[i].block[1].emph_coeff = 0.0f;
1083 /* Zero out previous energy (matters for inter first frame) */
1084 for (ch = 0; ch < s->channels; ch++)
1085 for (i = 0; i < CELT_MAX_BANDS; i++)
1086 s->last_quantized_energy[ch][i] = 0.0f;
1088 /* Allocate an empty frame to use as overlap for the first frame of audio */
1089 ff_bufqueue_add(avctx, &s->bufqueue, spawn_empty_frame(s));
1090 if (!ff_bufqueue_peek(&s->bufqueue, 0))
1091 return AVERROR(ENOMEM);
1096 #define OPUSENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
1097 static const AVOption opusenc_options[] = {
1098 { "opus_delay", "Maximum delay (and lookahead) in milliseconds", offsetof(OpusEncContext, options.max_delay_ms), AV_OPT_TYPE_FLOAT, { .dbl = OPUS_MAX_LOOKAHEAD }, 2.5f, OPUS_MAX_LOOKAHEAD, OPUSENC_FLAGS },
1102 static const AVClass opusenc_class = {
1103 .class_name = "Opus encoder",
1104 .item_name = av_default_item_name,
1105 .option = opusenc_options,
1106 .version = LIBAVUTIL_VERSION_INT,
1109 static const AVCodecDefault opusenc_defaults[] = {
1111 { "compression_level", "10" },
1115 AVCodec ff_opus_encoder = {
1117 .long_name = NULL_IF_CONFIG_SMALL("Opus"),
1118 .type = AVMEDIA_TYPE_AUDIO,
1119 .id = AV_CODEC_ID_OPUS,
1120 .defaults = opusenc_defaults,
1121 .priv_class = &opusenc_class,
1122 .priv_data_size = sizeof(OpusEncContext),
1123 .init = opus_encode_init,
1124 .encode2 = opus_encode_frame,
1125 .close = opus_encode_end,
1126 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE | FF_CODEC_CAP_INIT_CLEANUP,
1127 .capabilities = AV_CODEC_CAP_EXPERIMENTAL | AV_CODEC_CAP_SMALL_LAST_FRAME | AV_CODEC_CAP_DELAY,
1128 .supported_samplerates = (const int []){ 48000, 0 },
1129 .channel_layouts = (const uint64_t []){ AV_CH_LAYOUT_MONO,
1130 AV_CH_LAYOUT_STEREO, 0 },
1131 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
1132 AV_SAMPLE_FMT_NONE },