FFmpeg coverage


Directory: ../../../ffmpeg/
File: src/libavcodec/opusenc.c
Date: 2024-07-14 13:34:57
Exec Total Coverage
Lines: 0 399 0.0%
Functions: 0 18 0.0%
Branches: 0 214 0.0%

Line Branch Exec Source
1 /*
2 * Opus encoder
3 * Copyright (c) 2017 Rostislav Pehlivanov <atomnuker@gmail.com>
4 *
5 * This file is part of FFmpeg.
6 *
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.
11 *
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.
16 *
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
20 */
21
22 #include <float.h>
23
24 #include "encode.h"
25 #include "opusenc.h"
26 #include "opus_pvq.h"
27 #include "opusenc_psy.h"
28 #include "opustab.h"
29
30 #include "libavutil/channel_layout.h"
31 #include "libavutil/float_dsp.h"
32 #include "libavutil/mem.h"
33 #include "libavutil/mem_internal.h"
34 #include "libavutil/opt.h"
35 #include "bytestream.h"
36 #include "audio_frame_queue.h"
37 #include "codec_internal.h"
38
39 typedef struct OpusEncContext {
40 AVClass *av_class;
41 OpusEncOptions options;
42 OpusPsyContext psyctx;
43 AVCodecContext *avctx;
44 AudioFrameQueue afq;
45 AVFloatDSPContext *dsp;
46 AVTXContext *tx[CELT_BLOCK_NB];
47 av_tx_fn tx_fn[CELT_BLOCK_NB];
48 CeltPVQ *pvq;
49 struct FFBufQueue bufqueue;
50
51 uint8_t enc_id[64];
52 int enc_id_bits;
53
54 OpusPacketInfo packet;
55
56 int channels;
57
58 CeltFrame *frame;
59 OpusRangeCoder *rc;
60
61 /* Actual energy the decoder will have */
62 float last_quantized_energy[OPUS_MAX_CHANNELS][CELT_MAX_BANDS];
63
64 DECLARE_ALIGNED(32, float, scratch)[2048];
65 } OpusEncContext;
66
67 static void opus_write_extradata(AVCodecContext *avctx)
68 {
69 uint8_t *bs = avctx->extradata;
70
71 bytestream_put_buffer(&bs, "OpusHead", 8);
72 bytestream_put_byte (&bs, 0x1);
73 bytestream_put_byte (&bs, avctx->ch_layout.nb_channels);
74 bytestream_put_le16 (&bs, avctx->initial_padding);
75 bytestream_put_le32 (&bs, avctx->sample_rate);
76 bytestream_put_le16 (&bs, 0x0);
77 bytestream_put_byte (&bs, 0x0); /* Default layout */
78 }
79
80 static int opus_gen_toc(OpusEncContext *s, uint8_t *toc, int *size, int *fsize_needed)
81 {
82 int tmp = 0x0, extended_toc = 0;
83 static const int toc_cfg[][OPUS_MODE_NB][OPUS_BANDWITH_NB] = {
84 /* Silk Hybrid Celt Layer */
85 /* NB MB WB SWB FB NB MB WB SWB FB NB MB WB SWB FB Bandwidth */
86 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 17, 0, 21, 25, 29 } }, /* 2.5 ms */
87 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 18, 0, 22, 26, 30 } }, /* 5 ms */
88 { { 1, 5, 9, 0, 0 }, { 0, 0, 0, 13, 15 }, { 19, 0, 23, 27, 31 } }, /* 10 ms */
89 { { 2, 6, 10, 0, 0 }, { 0, 0, 0, 14, 16 }, { 20, 0, 24, 28, 32 } }, /* 20 ms */
90 { { 3, 7, 11, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 40 ms */
91 { { 4, 8, 12, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 60 ms */
92 };
93 int cfg = toc_cfg[s->packet.framesize][s->packet.mode][s->packet.bandwidth];
94 *fsize_needed = 0;
95 if (!cfg)
96 return 1;
97 if (s->packet.frames == 2) { /* 2 packets */
98 if (s->frame[0].framebits == s->frame[1].framebits) { /* same size */
99 tmp = 0x1;
100 } else { /* different size */
101 tmp = 0x2;
102 *fsize_needed = 1; /* put frame sizes in the packet */
103 }
104 } else if (s->packet.frames > 2) {
105 tmp = 0x3;
106 extended_toc = 1;
107 }
108 tmp |= (s->channels > 1) << 2; /* Stereo or mono */
109 tmp |= (cfg - 1) << 3; /* codec configuration */
110 *toc++ = tmp;
111 if (extended_toc) {
112 for (int i = 0; i < (s->packet.frames - 1); i++)
113 *fsize_needed |= (s->frame[i].framebits != s->frame[i + 1].framebits);
114 tmp = (*fsize_needed) << 7; /* vbr flag */
115 tmp |= (0) << 6; /* padding flag */
116 tmp |= s->packet.frames;
117 *toc++ = tmp;
118 }
119 *size = 1 + extended_toc;
120 return 0;
121 }
122
123 static void celt_frame_setup_input(OpusEncContext *s, CeltFrame *f)
124 {
125 AVFrame *cur = NULL;
126 const int subframesize = s->avctx->frame_size;
127 int subframes = OPUS_BLOCK_SIZE(s->packet.framesize) / subframesize;
128
129 cur = ff_bufqueue_get(&s->bufqueue);
130
131 for (int ch = 0; ch < f->channels; ch++) {
132 CeltBlock *b = &f->block[ch];
133 const void *input = cur->extended_data[ch];
134 size_t bps = av_get_bytes_per_sample(cur->format);
135 memcpy(b->overlap, input, bps*cur->nb_samples);
136 }
137
138 av_frame_free(&cur);
139
140 for (int sf = 0; sf < subframes; sf++) {
141 if (sf != (subframes - 1))
142 cur = ff_bufqueue_get(&s->bufqueue);
143 else
144 cur = ff_bufqueue_peek(&s->bufqueue, 0);
145
146 for (int ch = 0; ch < f->channels; ch++) {
147 CeltBlock *b = &f->block[ch];
148 const void *input = cur->extended_data[ch];
149 const size_t bps = av_get_bytes_per_sample(cur->format);
150 const size_t left = (subframesize - cur->nb_samples)*bps;
151 const size_t len = FFMIN(subframesize, cur->nb_samples)*bps;
152 memcpy(&b->samples[sf*subframesize], input, len);
153 memset(&b->samples[cur->nb_samples], 0, left);
154 }
155
156 /* Last frame isn't popped off and freed yet - we need it for overlap */
157 if (sf != (subframes - 1))
158 av_frame_free(&cur);
159 }
160 }
161
162 /* Apply the pre emphasis filter */
163 static void celt_apply_preemph_filter(OpusEncContext *s, CeltFrame *f)
164 {
165 const int subframesize = s->avctx->frame_size;
166 const int subframes = OPUS_BLOCK_SIZE(s->packet.framesize) / subframesize;
167 const float c = ff_opus_deemph_weights[0];
168
169 /* Filter overlap */
170 for (int ch = 0; ch < f->channels; ch++) {
171 CeltBlock *b = &f->block[ch];
172 float m = b->emph_coeff;
173 for (int i = 0; i < CELT_OVERLAP; i++) {
174 float sample = b->overlap[i];
175 b->overlap[i] = sample - m;
176 m = sample * c;
177 }
178 b->emph_coeff = m;
179 }
180
181 /* Filter the samples but do not update the last subframe's coeff - overlap ^^^ */
182 for (int sf = 0; sf < subframes; sf++) {
183 for (int ch = 0; ch < f->channels; ch++) {
184 CeltBlock *b = &f->block[ch];
185 float m = b->emph_coeff;
186 for (int i = 0; i < subframesize; i++) {
187 float sample = b->samples[sf*subframesize + i];
188 b->samples[sf*subframesize + i] = sample - m;
189 m = sample * c;
190 }
191 if (sf != (subframes - 1))
192 b->emph_coeff = m;
193 }
194 }
195 }
196
197 /* Create the window and do the mdct */
198 static void celt_frame_mdct(OpusEncContext *s, CeltFrame *f)
199 {
200 float *win = s->scratch, *temp = s->scratch + 1920;
201
202 if (f->transient) {
203 for (int ch = 0; ch < f->channels; ch++) {
204 CeltBlock *b = &f->block[ch];
205 float *src1 = b->overlap;
206 for (int t = 0; t < f->blocks; t++) {
207 float *src2 = &b->samples[CELT_OVERLAP*t];
208 s->dsp->vector_fmul(win, src1, ff_celt_window, 128);
209 s->dsp->vector_fmul_reverse(&win[CELT_OVERLAP], src2,
210 ff_celt_window_padded, 128);
211 src1 = src2;
212 s->tx_fn[0](s->tx[0], b->coeffs + t, win, sizeof(float)*f->blocks);
213 }
214 }
215 } else {
216 int blk_len = OPUS_BLOCK_SIZE(f->size), wlen = OPUS_BLOCK_SIZE(f->size + 1);
217 int rwin = blk_len - CELT_OVERLAP, lap_dst = (wlen - blk_len - CELT_OVERLAP) >> 1;
218 memset(win, 0, wlen*sizeof(float));
219 for (int ch = 0; ch < f->channels; ch++) {
220 CeltBlock *b = &f->block[ch];
221
222 /* Overlap */
223 s->dsp->vector_fmul(temp, b->overlap, ff_celt_window, 128);
224 memcpy(win + lap_dst, temp, CELT_OVERLAP*sizeof(float));
225
226 /* Samples, flat top window */
227 memcpy(&win[lap_dst + CELT_OVERLAP], b->samples, rwin*sizeof(float));
228
229 /* Samples, windowed */
230 s->dsp->vector_fmul_reverse(temp, b->samples + rwin,
231 ff_celt_window_padded, 128);
232 memcpy(win + lap_dst + blk_len, temp, CELT_OVERLAP*sizeof(float));
233
234 s->tx_fn[f->size](s->tx[f->size], b->coeffs, win, sizeof(float));
235 }
236 }
237
238 for (int ch = 0; ch < f->channels; ch++) {
239 CeltBlock *block = &f->block[ch];
240 for (int i = 0; i < CELT_MAX_BANDS; i++) {
241 float ener = 0.0f;
242 int band_offset = ff_celt_freq_bands[i] << f->size;
243 int band_size = ff_celt_freq_range[i] << f->size;
244 float *coeffs = &block->coeffs[band_offset];
245
246 for (int j = 0; j < band_size; j++)
247 ener += coeffs[j]*coeffs[j];
248
249 block->lin_energy[i] = sqrtf(ener) + FLT_EPSILON;
250 ener = 1.0f/block->lin_energy[i];
251
252 for (int j = 0; j < band_size; j++)
253 coeffs[j] *= ener;
254
255 block->energy[i] = log2f(block->lin_energy[i]) - ff_celt_mean_energy[i];
256
257 /* CELT_ENERGY_SILENCE is what the decoder uses and its not -infinity */
258 block->energy[i] = FFMAX(block->energy[i], CELT_ENERGY_SILENCE);
259 }
260 }
261 }
262
263 static void celt_enc_tf(CeltFrame *f, OpusRangeCoder *rc)
264 {
265 int tf_select = 0, diff = 0, tf_changed = 0, tf_select_needed;
266 int bits = f->transient ? 2 : 4;
267
268 tf_select_needed = ((f->size && (opus_rc_tell(rc) + bits + 1) <= f->framebits));
269
270 for (int i = f->start_band; i < f->end_band; i++) {
271 if ((opus_rc_tell(rc) + bits + tf_select_needed) <= f->framebits) {
272 const int tbit = (diff ^ 1) == f->tf_change[i];
273 ff_opus_rc_enc_log(rc, tbit, bits);
274 diff ^= tbit;
275 tf_changed |= diff;
276 }
277 bits = f->transient ? 4 : 5;
278 }
279
280 if (tf_select_needed && ff_celt_tf_select[f->size][f->transient][0][tf_changed] !=
281 ff_celt_tf_select[f->size][f->transient][1][tf_changed]) {
282 ff_opus_rc_enc_log(rc, f->tf_select, 1);
283 tf_select = f->tf_select;
284 }
285
286 for (int i = f->start_band; i < f->end_band; i++)
287 f->tf_change[i] = ff_celt_tf_select[f->size][f->transient][tf_select][f->tf_change[i]];
288 }
289
290 static void celt_enc_quant_pfilter(OpusRangeCoder *rc, CeltFrame *f)
291 {
292 float gain = f->pf_gain;
293 int txval, octave = f->pf_octave, period = f->pf_period, tapset = f->pf_tapset;
294
295 ff_opus_rc_enc_log(rc, f->pfilter, 1);
296 if (!f->pfilter)
297 return;
298
299 /* Octave */
300 txval = FFMIN(octave, 6);
301 ff_opus_rc_enc_uint(rc, txval, 6);
302 octave = txval;
303 /* Period */
304 txval = av_clip(period - (16 << octave) + 1, 0, (1 << (4 + octave)) - 1);
305 ff_opus_rc_put_raw(rc, period, 4 + octave);
306 period = txval + (16 << octave) - 1;
307 /* Gain */
308 txval = FFMIN(((int)(gain / 0.09375f)) - 1, 7);
309 ff_opus_rc_put_raw(rc, txval, 3);
310 gain = 0.09375f * (txval + 1);
311 /* Tapset */
312 if ((opus_rc_tell(rc) + 2) <= f->framebits)
313 ff_opus_rc_enc_cdf(rc, tapset, ff_celt_model_tapset);
314 else
315 tapset = 0;
316 /* Finally create the coeffs */
317 for (int i = 0; i < 2; i++) {
318 CeltBlock *block = &f->block[i];
319
320 block->pf_period_new = FFMAX(period, CELT_POSTFILTER_MINPERIOD);
321 block->pf_gains_new[0] = gain * ff_celt_postfilter_taps[tapset][0];
322 block->pf_gains_new[1] = gain * ff_celt_postfilter_taps[tapset][1];
323 block->pf_gains_new[2] = gain * ff_celt_postfilter_taps[tapset][2];
324 }
325 }
326
327 static void exp_quant_coarse(OpusRangeCoder *rc, CeltFrame *f,
328 float last_energy[][CELT_MAX_BANDS], int intra)
329 {
330 float alpha, beta, prev[2] = { 0, 0 };
331 const uint8_t *pmod = ff_celt_coarse_energy_dist[f->size][intra];
332
333 /* Inter is really just differential coding */
334 if (opus_rc_tell(rc) + 3 <= f->framebits)
335 ff_opus_rc_enc_log(rc, intra, 3);
336 else
337 intra = 0;
338
339 if (intra) {
340 alpha = 0.0f;
341 beta = 1.0f - (4915.0f/32768.0f);
342 } else {
343 alpha = ff_celt_alpha_coef[f->size];
344 beta = ff_celt_beta_coef[f->size];
345 }
346
347 for (int i = f->start_band; i < f->end_band; i++) {
348 for (int ch = 0; ch < f->channels; ch++) {
349 CeltBlock *block = &f->block[ch];
350 const int left = f->framebits - opus_rc_tell(rc);
351 const float last = FFMAX(-9.0f, last_energy[ch][i]);
352 float diff = block->energy[i] - prev[ch] - last*alpha;
353 int q_en = lrintf(diff);
354 if (left >= 15) {
355 ff_opus_rc_enc_laplace(rc, &q_en, pmod[i << 1] << 7, pmod[(i << 1) + 1] << 6);
356 } else if (left >= 2) {
357 q_en = av_clip(q_en, -1, 1);
358 ff_opus_rc_enc_cdf(rc, 2*q_en + 3*(q_en < 0), ff_celt_model_energy_small);
359 } else if (left >= 1) {
360 q_en = av_clip(q_en, -1, 0);
361 ff_opus_rc_enc_log(rc, (q_en & 1), 1);
362 } else q_en = -1;
363
364 block->error_energy[i] = q_en - diff;
365 prev[ch] += beta * q_en;
366 }
367 }
368 }
369
370 static void celt_quant_coarse(CeltFrame *f, OpusRangeCoder *rc,
371 float last_energy[][CELT_MAX_BANDS])
372 {
373 uint32_t inter, intra;
374 OPUS_RC_CHECKPOINT_SPAWN(rc);
375
376 exp_quant_coarse(rc, f, last_energy, 1);
377 intra = OPUS_RC_CHECKPOINT_BITS(rc);
378
379 OPUS_RC_CHECKPOINT_ROLLBACK(rc);
380
381 exp_quant_coarse(rc, f, last_energy, 0);
382 inter = OPUS_RC_CHECKPOINT_BITS(rc);
383
384 if (inter > intra) { /* Unlikely */
385 OPUS_RC_CHECKPOINT_ROLLBACK(rc);
386 exp_quant_coarse(rc, f, last_energy, 1);
387 }
388 }
389
390 static void celt_quant_fine(CeltFrame *f, OpusRangeCoder *rc)
391 {
392 for (int i = f->start_band; i < f->end_band; i++) {
393 if (!f->fine_bits[i])
394 continue;
395 for (int ch = 0; ch < f->channels; ch++) {
396 CeltBlock *block = &f->block[ch];
397 int quant, lim = (1 << f->fine_bits[i]);
398 float offset, diff = 0.5f - block->error_energy[i];
399 quant = av_clip(floor(diff*lim), 0, lim - 1);
400 ff_opus_rc_put_raw(rc, quant, f->fine_bits[i]);
401 offset = 0.5f - ((quant + 0.5f) * (1 << (14 - f->fine_bits[i])) / 16384.0f);
402 block->error_energy[i] -= offset;
403 }
404 }
405 }
406
407 static void celt_quant_final(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f)
408 {
409 for (int priority = 0; priority < 2; priority++) {
410 for (int i = f->start_band; i < f->end_band && (f->framebits - opus_rc_tell(rc)) >= f->channels; i++) {
411 if (f->fine_priority[i] != priority || f->fine_bits[i] >= CELT_MAX_FINE_BITS)
412 continue;
413 for (int ch = 0; ch < f->channels; ch++) {
414 CeltBlock *block = &f->block[ch];
415 const float err = block->error_energy[i];
416 const float offset = 0.5f * (1 << (14 - f->fine_bits[i] - 1)) / 16384.0f;
417 const int sign = FFABS(err + offset) < FFABS(err - offset);
418 ff_opus_rc_put_raw(rc, sign, 1);
419 block->error_energy[i] -= offset*(1 - 2*sign);
420 }
421 }
422 }
423 }
424
425 static void celt_encode_frame(OpusEncContext *s, OpusRangeCoder *rc,
426 CeltFrame *f, int index)
427 {
428 ff_opus_rc_enc_init(rc);
429
430 ff_opus_psy_celt_frame_init(&s->psyctx, f, index);
431
432 celt_frame_setup_input(s, f);
433
434 if (f->silence) {
435 if (f->framebits >= 16)
436 ff_opus_rc_enc_log(rc, 1, 15); /* Silence (if using explicit singalling) */
437 for (int ch = 0; ch < s->channels; ch++)
438 memset(s->last_quantized_energy[ch], 0.0f, sizeof(float)*CELT_MAX_BANDS);
439 return;
440 }
441
442 /* Filters */
443 celt_apply_preemph_filter(s, f);
444 if (f->pfilter) {
445 ff_opus_rc_enc_log(rc, 0, 15);
446 celt_enc_quant_pfilter(rc, f);
447 }
448
449 /* Transform */
450 celt_frame_mdct(s, f);
451
452 /* Need to handle transient/non-transient switches at any point during analysis */
453 while (ff_opus_psy_celt_frame_process(&s->psyctx, f, index))
454 celt_frame_mdct(s, f);
455
456 ff_opus_rc_enc_init(rc);
457
458 /* Silence */
459 ff_opus_rc_enc_log(rc, 0, 15);
460
461 /* Pitch filter */
462 if (!f->start_band && opus_rc_tell(rc) + 16 <= f->framebits)
463 celt_enc_quant_pfilter(rc, f);
464
465 /* Transient flag */
466 if (f->size && opus_rc_tell(rc) + 3 <= f->framebits)
467 ff_opus_rc_enc_log(rc, f->transient, 3);
468
469 /* Main encoding */
470 celt_quant_coarse (f, rc, s->last_quantized_energy);
471 celt_enc_tf (f, rc);
472 ff_celt_bitalloc (f, rc, 1);
473 celt_quant_fine (f, rc);
474 ff_celt_quant_bands(f, rc);
475
476 /* Anticollapse bit */
477 if (f->anticollapse_needed)
478 ff_opus_rc_put_raw(rc, f->anticollapse, 1);
479
480 /* Final per-band energy adjustments from leftover bits */
481 celt_quant_final(s, rc, f);
482
483 for (int ch = 0; ch < f->channels; ch++) {
484 CeltBlock *block = &f->block[ch];
485 for (int i = 0; i < CELT_MAX_BANDS; i++)
486 s->last_quantized_energy[ch][i] = block->energy[i] + block->error_energy[i];
487 }
488 }
489
490 static inline int write_opuslacing(uint8_t *dst, int v)
491 {
492 dst[0] = FFMIN(v - FFALIGN(v - 255, 4), v);
493 dst[1] = v - dst[0] >> 2;
494 return 1 + (v >= 252);
495 }
496
497 static void opus_packet_assembler(OpusEncContext *s, AVPacket *avpkt)
498 {
499 int offset, fsize_needed;
500
501 /* Write toc */
502 opus_gen_toc(s, avpkt->data, &offset, &fsize_needed);
503
504 /* Frame sizes if needed */
505 if (fsize_needed) {
506 for (int i = 0; i < s->packet.frames - 1; i++) {
507 offset += write_opuslacing(avpkt->data + offset,
508 s->frame[i].framebits >> 3);
509 }
510 }
511
512 /* Packets */
513 for (int i = 0; i < s->packet.frames; i++) {
514 ff_opus_rc_enc_end(&s->rc[i], avpkt->data + offset,
515 s->frame[i].framebits >> 3);
516 offset += s->frame[i].framebits >> 3;
517 }
518
519 avpkt->size = offset;
520 }
521
522 /* Used as overlap for the first frame and padding for the last encoded packet */
523 static AVFrame *spawn_empty_frame(OpusEncContext *s)
524 {
525 AVFrame *f = av_frame_alloc();
526 int ret;
527 if (!f)
528 return NULL;
529 f->format = s->avctx->sample_fmt;
530 f->nb_samples = s->avctx->frame_size;
531 ret = av_channel_layout_copy(&f->ch_layout, &s->avctx->ch_layout);
532 if (ret < 0) {
533 av_frame_free(&f);
534 return NULL;
535 }
536 if (av_frame_get_buffer(f, 4)) {
537 av_frame_free(&f);
538 return NULL;
539 }
540 for (int i = 0; i < s->channels; i++) {
541 size_t bps = av_get_bytes_per_sample(f->format);
542 memset(f->extended_data[i], 0, bps*f->nb_samples);
543 }
544 return f;
545 }
546
547 static int opus_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
548 const AVFrame *frame, int *got_packet_ptr)
549 {
550 OpusEncContext *s = avctx->priv_data;
551 int ret, frame_size, alloc_size = 0;
552
553 if (frame) { /* Add new frame to queue */
554 if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
555 return ret;
556 ff_bufqueue_add(avctx, &s->bufqueue, av_frame_clone(frame));
557 } else {
558 ff_opus_psy_signal_eof(&s->psyctx);
559 if (!s->afq.remaining_samples || !avctx->frame_num)
560 return 0; /* We've been flushed and there's nothing left to encode */
561 }
562
563 /* Run the psychoacoustic system */
564 if (ff_opus_psy_process(&s->psyctx, &s->packet))
565 return 0;
566
567 frame_size = OPUS_BLOCK_SIZE(s->packet.framesize);
568
569 if (!frame) {
570 /* This can go negative, that's not a problem, we only pad if positive */
571 int pad_empty = s->packet.frames*(frame_size/s->avctx->frame_size) - s->bufqueue.available + 1;
572 /* Pad with empty 2.5 ms frames to whatever framesize was decided,
573 * this should only happen at the very last flush frame. The frames
574 * allocated here will be freed (because they have no other references)
575 * after they get used by celt_frame_setup_input() */
576 for (int i = 0; i < pad_empty; i++) {
577 AVFrame *empty = spawn_empty_frame(s);
578 if (!empty)
579 return AVERROR(ENOMEM);
580 ff_bufqueue_add(avctx, &s->bufqueue, empty);
581 }
582 }
583
584 for (int i = 0; i < s->packet.frames; i++) {
585 celt_encode_frame(s, &s->rc[i], &s->frame[i], i);
586 alloc_size += s->frame[i].framebits >> 3;
587 }
588
589 /* Worst case toc + the frame lengths if needed */
590 alloc_size += 2 + s->packet.frames*2;
591
592 if ((ret = ff_alloc_packet(avctx, avpkt, alloc_size)) < 0)
593 return ret;
594
595 /* Assemble packet */
596 opus_packet_assembler(s, avpkt);
597
598 /* Update the psychoacoustic system */
599 ff_opus_psy_postencode_update(&s->psyctx, s->frame);
600
601 /* Remove samples from queue and skip if needed */
602 ff_af_queue_remove(&s->afq, s->packet.frames*frame_size, &avpkt->pts, &avpkt->duration);
603 if (s->packet.frames*frame_size > avpkt->duration) {
604 uint8_t *side = av_packet_new_side_data(avpkt, AV_PKT_DATA_SKIP_SAMPLES, 10);
605 if (!side)
606 return AVERROR(ENOMEM);
607 AV_WL32(&side[4], s->packet.frames*frame_size - avpkt->duration + 120);
608 }
609
610 *got_packet_ptr = 1;
611
612 return 0;
613 }
614
615 static av_cold int opus_encode_end(AVCodecContext *avctx)
616 {
617 OpusEncContext *s = avctx->priv_data;
618
619 for (int i = 0; i < CELT_BLOCK_NB; i++)
620 av_tx_uninit(&s->tx[i]);
621
622 ff_celt_pvq_uninit(&s->pvq);
623 av_freep(&s->dsp);
624 av_freep(&s->frame);
625 av_freep(&s->rc);
626 ff_af_queue_close(&s->afq);
627 ff_opus_psy_end(&s->psyctx);
628 ff_bufqueue_discard_all(&s->bufqueue);
629
630 return 0;
631 }
632
633 static av_cold int opus_encode_init(AVCodecContext *avctx)
634 {
635 int ret, max_frames;
636 OpusEncContext *s = avctx->priv_data;
637
638 s->avctx = avctx;
639 s->channels = avctx->ch_layout.nb_channels;
640
641 /* Opus allows us to change the framesize on each packet (and each packet may
642 * have multiple frames in it) but we can't change the codec's frame size on
643 * runtime, so fix it to the lowest possible number of samples and use a queue
644 * to accumulate AVFrames until we have enough to encode whatever the encoder
645 * decides is the best */
646 avctx->frame_size = 120;
647 /* Initial padding will change if SILK is ever supported */
648 avctx->initial_padding = 120;
649
650 if (!avctx->bit_rate) {
651 int coupled = ff_opus_default_coupled_streams[s->channels - 1];
652 avctx->bit_rate = coupled*(96000) + (s->channels - coupled*2)*(48000);
653 } else if (avctx->bit_rate < 6000 || avctx->bit_rate > 255000 * s->channels) {
654 int64_t clipped_rate = av_clip(avctx->bit_rate, 6000, 255000 * s->channels);
655 av_log(avctx, AV_LOG_ERROR, "Unsupported bitrate %"PRId64" kbps, clipping to %"PRId64" kbps\n",
656 avctx->bit_rate/1000, clipped_rate/1000);
657 avctx->bit_rate = clipped_rate;
658 }
659
660 /* Extradata */
661 avctx->extradata_size = 19;
662 avctx->extradata = av_malloc(avctx->extradata_size + AV_INPUT_BUFFER_PADDING_SIZE);
663 if (!avctx->extradata)
664 return AVERROR(ENOMEM);
665 opus_write_extradata(avctx);
666
667 ff_af_queue_init(avctx, &s->afq);
668
669 if ((ret = ff_celt_pvq_init(&s->pvq, 1)) < 0)
670 return ret;
671
672 if (!(s->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT)))
673 return AVERROR(ENOMEM);
674
675 /* I have no idea why a base scaling factor of 68 works, could be the twiddles */
676 for (int i = 0; i < CELT_BLOCK_NB; i++) {
677 const float scale = 68 << (CELT_BLOCK_NB - 1 - i);
678 if ((ret = av_tx_init(&s->tx[i], &s->tx_fn[i], AV_TX_FLOAT_MDCT, 0, 15 << (i + 3), &scale, 0)))
679 return AVERROR(ENOMEM);
680 }
681
682 /* Zero out previous energy (matters for inter first frame) */
683 for (int ch = 0; ch < s->channels; ch++)
684 memset(s->last_quantized_energy[ch], 0.0f, sizeof(float)*CELT_MAX_BANDS);
685
686 /* Allocate an empty frame to use as overlap for the first frame of audio */
687 ff_bufqueue_add(avctx, &s->bufqueue, spawn_empty_frame(s));
688 if (!ff_bufqueue_peek(&s->bufqueue, 0))
689 return AVERROR(ENOMEM);
690
691 if ((ret = ff_opus_psy_init(&s->psyctx, s->avctx, &s->bufqueue, &s->options)))
692 return ret;
693
694 /* Frame structs and range coder buffers */
695 max_frames = ceilf(FFMIN(s->options.max_delay_ms, 120.0f)/2.5f);
696 s->frame = av_malloc(max_frames*sizeof(CeltFrame));
697 if (!s->frame)
698 return AVERROR(ENOMEM);
699 s->rc = av_malloc(max_frames*sizeof(OpusRangeCoder));
700 if (!s->rc)
701 return AVERROR(ENOMEM);
702
703 for (int i = 0; i < max_frames; i++) {
704 s->frame[i].dsp = s->dsp;
705 s->frame[i].avctx = s->avctx;
706 s->frame[i].seed = 0;
707 s->frame[i].pvq = s->pvq;
708 s->frame[i].apply_phase_inv = s->options.apply_phase_inv;
709 s->frame[i].block[0].emph_coeff = s->frame[i].block[1].emph_coeff = 0.0f;
710 }
711
712 return 0;
713 }
714
715 #define OPUSENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
716 static const AVOption opusenc_options[] = {
717 { "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, .unit = "max_delay_ms" },
718 { "apply_phase_inv", "Apply intensity stereo phase inversion", offsetof(OpusEncContext, options.apply_phase_inv), AV_OPT_TYPE_BOOL, { .i64 = 1 }, 0, 1, OPUSENC_FLAGS, .unit = "apply_phase_inv" },
719 { NULL },
720 };
721
722 static const AVClass opusenc_class = {
723 .class_name = "Opus encoder",
724 .item_name = av_default_item_name,
725 .option = opusenc_options,
726 .version = LIBAVUTIL_VERSION_INT,
727 };
728
729 static const FFCodecDefault opusenc_defaults[] = {
730 { "b", "0" },
731 { "compression_level", "10" },
732 { NULL },
733 };
734
735 const FFCodec ff_opus_encoder = {
736 .p.name = "opus",
737 CODEC_LONG_NAME("Opus"),
738 .p.type = AVMEDIA_TYPE_AUDIO,
739 .p.id = AV_CODEC_ID_OPUS,
740 .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_DELAY |
741 AV_CODEC_CAP_SMALL_LAST_FRAME | AV_CODEC_CAP_EXPERIMENTAL,
742 .defaults = opusenc_defaults,
743 .p.priv_class = &opusenc_class,
744 .priv_data_size = sizeof(OpusEncContext),
745 .init = opus_encode_init,
746 FF_CODEC_ENCODE_CB(opus_encode_frame),
747 .close = opus_encode_end,
748 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
749 .p.supported_samplerates = (const int []){ 48000, 0 },
750 .p.ch_layouts = (const AVChannelLayout []){ AV_CHANNEL_LAYOUT_MONO,
751 AV_CHANNEL_LAYOUT_STEREO, { 0 } },
752 .p.sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
753 AV_SAMPLE_FMT_NONE },
754 };
755