FFmpeg coverage


Directory: ../../../ffmpeg/
File: src/libavcodec/opusenc.c
Date: 2022-12-09 07:38:14
Exec Total Coverage
Lines: 0 388 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_internal.h"
33 #include "libavutil/opt.h"
34 #include "bytestream.h"
35 #include "audio_frame_queue.h"
36 #include "codec_internal.h"
37
38 typedef struct OpusEncContext {
39 AVClass *av_class;
40 OpusEncOptions options;
41 OpusPsyContext psyctx;
42 AVCodecContext *avctx;
43 AudioFrameQueue afq;
44 AVFloatDSPContext *dsp;
45 AVTXContext *tx[CELT_BLOCK_NB];
46 av_tx_fn tx_fn[CELT_BLOCK_NB];
47 CeltPVQ *pvq;
48 struct FFBufQueue bufqueue;
49
50 uint8_t enc_id[64];
51 int enc_id_bits;
52
53 OpusPacketInfo packet;
54
55 int channels;
56
57 CeltFrame *frame;
58 OpusRangeCoder *rc;
59
60 /* Actual energy the decoder will have */
61 float last_quantized_energy[OPUS_MAX_CHANNELS][CELT_MAX_BANDS];
62
63 DECLARE_ALIGNED(32, float, scratch)[2048];
64 } OpusEncContext;
65
66 static void opus_write_extradata(AVCodecContext *avctx)
67 {
68 uint8_t *bs = avctx->extradata;
69
70 bytestream_put_buffer(&bs, "OpusHead", 8);
71 bytestream_put_byte (&bs, 0x1);
72 bytestream_put_byte (&bs, avctx->ch_layout.nb_channels);
73 bytestream_put_le16 (&bs, avctx->initial_padding);
74 bytestream_put_le32 (&bs, avctx->sample_rate);
75 bytestream_put_le16 (&bs, 0x0);
76 bytestream_put_byte (&bs, 0x0); /* Default layout */
77 }
78
79 static int opus_gen_toc(OpusEncContext *s, uint8_t *toc, int *size, int *fsize_needed)
80 {
81 int tmp = 0x0, extended_toc = 0;
82 static const int toc_cfg[][OPUS_MODE_NB][OPUS_BANDWITH_NB] = {
83 /* Silk Hybrid Celt Layer */
84 /* NB MB WB SWB FB NB MB WB SWB FB NB MB WB SWB FB Bandwidth */
85 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 17, 0, 21, 25, 29 } }, /* 2.5 ms */
86 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 18, 0, 22, 26, 30 } }, /* 5 ms */
87 { { 1, 5, 9, 0, 0 }, { 0, 0, 0, 13, 15 }, { 19, 0, 23, 27, 31 } }, /* 10 ms */
88 { { 2, 6, 10, 0, 0 }, { 0, 0, 0, 14, 16 }, { 20, 0, 24, 28, 32 } }, /* 20 ms */
89 { { 3, 7, 11, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 40 ms */
90 { { 4, 8, 12, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 60 ms */
91 };
92 int cfg = toc_cfg[s->packet.framesize][s->packet.mode][s->packet.bandwidth];
93 *fsize_needed = 0;
94 if (!cfg)
95 return 1;
96 if (s->packet.frames == 2) { /* 2 packets */
97 if (s->frame[0].framebits == s->frame[1].framebits) { /* same size */
98 tmp = 0x1;
99 } else { /* different size */
100 tmp = 0x2;
101 *fsize_needed = 1; /* put frame sizes in the packet */
102 }
103 } else if (s->packet.frames > 2) {
104 tmp = 0x3;
105 extended_toc = 1;
106 }
107 tmp |= (s->channels > 1) << 2; /* Stereo or mono */
108 tmp |= (cfg - 1) << 3; /* codec configuration */
109 *toc++ = tmp;
110 if (extended_toc) {
111 for (int i = 0; i < (s->packet.frames - 1); i++)
112 *fsize_needed |= (s->frame[i].framebits != s->frame[i + 1].framebits);
113 tmp = (*fsize_needed) << 7; /* vbr flag */
114 tmp |= (0) << 6; /* padding flag */
115 tmp |= s->packet.frames;
116 *toc++ = tmp;
117 }
118 *size = 1 + extended_toc;
119 return 0;
120 }
121
122 static void celt_frame_setup_input(OpusEncContext *s, CeltFrame *f)
123 {
124 AVFrame *cur = NULL;
125 const int subframesize = s->avctx->frame_size;
126 int subframes = OPUS_BLOCK_SIZE(s->packet.framesize) / subframesize;
127
128 cur = ff_bufqueue_get(&s->bufqueue);
129
130 for (int ch = 0; ch < f->channels; ch++) {
131 CeltBlock *b = &f->block[ch];
132 const void *input = cur->extended_data[ch];
133 size_t bps = av_get_bytes_per_sample(cur->format);
134 memcpy(b->overlap, input, bps*cur->nb_samples);
135 }
136
137 av_frame_free(&cur);
138
139 for (int sf = 0; sf < subframes; sf++) {
140 if (sf != (subframes - 1))
141 cur = ff_bufqueue_get(&s->bufqueue);
142 else
143 cur = ff_bufqueue_peek(&s->bufqueue, 0);
144
145 for (int ch = 0; ch < f->channels; ch++) {
146 CeltBlock *b = &f->block[ch];
147 const void *input = cur->extended_data[ch];
148 const size_t bps = av_get_bytes_per_sample(cur->format);
149 const size_t left = (subframesize - cur->nb_samples)*bps;
150 const size_t len = FFMIN(subframesize, cur->nb_samples)*bps;
151 memcpy(&b->samples[sf*subframesize], input, len);
152 memset(&b->samples[cur->nb_samples], 0, left);
153 }
154
155 /* Last frame isn't popped off and freed yet - we need it for overlap */
156 if (sf != (subframes - 1))
157 av_frame_free(&cur);
158 }
159 }
160
161 /* Apply the pre emphasis filter */
162 static void celt_apply_preemph_filter(OpusEncContext *s, CeltFrame *f)
163 {
164 const int subframesize = s->avctx->frame_size;
165 const int subframes = OPUS_BLOCK_SIZE(s->packet.framesize) / subframesize;
166
167 /* Filter overlap */
168 for (int ch = 0; ch < f->channels; ch++) {
169 CeltBlock *b = &f->block[ch];
170 float m = b->emph_coeff;
171 for (int i = 0; i < CELT_OVERLAP; i++) {
172 float sample = b->overlap[i];
173 b->overlap[i] = sample - m;
174 m = sample * CELT_EMPH_COEFF;
175 }
176 b->emph_coeff = m;
177 }
178
179 /* Filter the samples but do not update the last subframe's coeff - overlap ^^^ */
180 for (int sf = 0; sf < subframes; sf++) {
181 for (int ch = 0; ch < f->channels; ch++) {
182 CeltBlock *b = &f->block[ch];
183 float m = b->emph_coeff;
184 for (int i = 0; i < subframesize; i++) {
185 float sample = b->samples[sf*subframesize + i];
186 b->samples[sf*subframesize + i] = sample - m;
187 m = sample * CELT_EMPH_COEFF;
188 }
189 if (sf != (subframes - 1))
190 b->emph_coeff = m;
191 }
192 }
193 }
194
195 /* Create the window and do the mdct */
196 static void celt_frame_mdct(OpusEncContext *s, CeltFrame *f)
197 {
198 float *win = s->scratch, *temp = s->scratch + 1920;
199
200 if (f->transient) {
201 for (int ch = 0; ch < f->channels; ch++) {
202 CeltBlock *b = &f->block[ch];
203 float *src1 = b->overlap;
204 for (int t = 0; t < f->blocks; t++) {
205 float *src2 = &b->samples[CELT_OVERLAP*t];
206 s->dsp->vector_fmul(win, src1, ff_celt_window, 128);
207 s->dsp->vector_fmul_reverse(&win[CELT_OVERLAP], src2,
208 ff_celt_window_padded, 128);
209 src1 = src2;
210 s->tx_fn[0](s->tx[0], b->coeffs + t, win, sizeof(float)*f->blocks);
211 }
212 }
213 } else {
214 int blk_len = OPUS_BLOCK_SIZE(f->size), wlen = OPUS_BLOCK_SIZE(f->size + 1);
215 int rwin = blk_len - CELT_OVERLAP, lap_dst = (wlen - blk_len - CELT_OVERLAP) >> 1;
216 memset(win, 0, wlen*sizeof(float));
217 for (int ch = 0; ch < f->channels; ch++) {
218 CeltBlock *b = &f->block[ch];
219
220 /* Overlap */
221 s->dsp->vector_fmul(temp, b->overlap, ff_celt_window, 128);
222 memcpy(win + lap_dst, temp, CELT_OVERLAP*sizeof(float));
223
224 /* Samples, flat top window */
225 memcpy(&win[lap_dst + CELT_OVERLAP], b->samples, rwin*sizeof(float));
226
227 /* Samples, windowed */
228 s->dsp->vector_fmul_reverse(temp, b->samples + rwin,
229 ff_celt_window_padded, 128);
230 memcpy(win + lap_dst + blk_len, temp, CELT_OVERLAP*sizeof(float));
231
232 s->tx_fn[f->size](s->tx[f->size], b->coeffs, win, sizeof(float));
233 }
234 }
235
236 for (int ch = 0; ch < f->channels; ch++) {
237 CeltBlock *block = &f->block[ch];
238 for (int i = 0; i < CELT_MAX_BANDS; i++) {
239 float ener = 0.0f;
240 int band_offset = ff_celt_freq_bands[i] << f->size;
241 int band_size = ff_celt_freq_range[i] << f->size;
242 float *coeffs = &block->coeffs[band_offset];
243
244 for (int j = 0; j < band_size; j++)
245 ener += coeffs[j]*coeffs[j];
246
247 block->lin_energy[i] = sqrtf(ener) + FLT_EPSILON;
248 ener = 1.0f/block->lin_energy[i];
249
250 for (int j = 0; j < band_size; j++)
251 coeffs[j] *= ener;
252
253 block->energy[i] = log2f(block->lin_energy[i]) - ff_celt_mean_energy[i];
254
255 /* CELT_ENERGY_SILENCE is what the decoder uses and its not -infinity */
256 block->energy[i] = FFMAX(block->energy[i], CELT_ENERGY_SILENCE);
257 }
258 }
259 }
260
261 static void celt_enc_tf(CeltFrame *f, OpusRangeCoder *rc)
262 {
263 int tf_select = 0, diff = 0, tf_changed = 0, tf_select_needed;
264 int bits = f->transient ? 2 : 4;
265
266 tf_select_needed = ((f->size && (opus_rc_tell(rc) + bits + 1) <= f->framebits));
267
268 for (int i = f->start_band; i < f->end_band; i++) {
269 if ((opus_rc_tell(rc) + bits + tf_select_needed) <= f->framebits) {
270 const int tbit = (diff ^ 1) == f->tf_change[i];
271 ff_opus_rc_enc_log(rc, tbit, bits);
272 diff ^= tbit;
273 tf_changed |= diff;
274 }
275 bits = f->transient ? 4 : 5;
276 }
277
278 if (tf_select_needed && ff_celt_tf_select[f->size][f->transient][0][tf_changed] !=
279 ff_celt_tf_select[f->size][f->transient][1][tf_changed]) {
280 ff_opus_rc_enc_log(rc, f->tf_select, 1);
281 tf_select = f->tf_select;
282 }
283
284 for (int i = f->start_band; i < f->end_band; i++)
285 f->tf_change[i] = ff_celt_tf_select[f->size][f->transient][tf_select][f->tf_change[i]];
286 }
287
288 static void celt_enc_quant_pfilter(OpusRangeCoder *rc, CeltFrame *f)
289 {
290 float gain = f->pf_gain;
291 int txval, octave = f->pf_octave, period = f->pf_period, tapset = f->pf_tapset;
292
293 ff_opus_rc_enc_log(rc, f->pfilter, 1);
294 if (!f->pfilter)
295 return;
296
297 /* Octave */
298 txval = FFMIN(octave, 6);
299 ff_opus_rc_enc_uint(rc, txval, 6);
300 octave = txval;
301 /* Period */
302 txval = av_clip(period - (16 << octave) + 1, 0, (1 << (4 + octave)) - 1);
303 ff_opus_rc_put_raw(rc, period, 4 + octave);
304 period = txval + (16 << octave) - 1;
305 /* Gain */
306 txval = FFMIN(((int)(gain / 0.09375f)) - 1, 7);
307 ff_opus_rc_put_raw(rc, txval, 3);
308 gain = 0.09375f * (txval + 1);
309 /* Tapset */
310 if ((opus_rc_tell(rc) + 2) <= f->framebits)
311 ff_opus_rc_enc_cdf(rc, tapset, ff_celt_model_tapset);
312 else
313 tapset = 0;
314 /* Finally create the coeffs */
315 for (int i = 0; i < 2; i++) {
316 CeltBlock *block = &f->block[i];
317
318 block->pf_period_new = FFMAX(period, CELT_POSTFILTER_MINPERIOD);
319 block->pf_gains_new[0] = gain * ff_celt_postfilter_taps[tapset][0];
320 block->pf_gains_new[1] = gain * ff_celt_postfilter_taps[tapset][1];
321 block->pf_gains_new[2] = gain * ff_celt_postfilter_taps[tapset][2];
322 }
323 }
324
325 static void exp_quant_coarse(OpusRangeCoder *rc, CeltFrame *f,
326 float last_energy[][CELT_MAX_BANDS], int intra)
327 {
328 float alpha, beta, prev[2] = { 0, 0 };
329 const uint8_t *pmod = ff_celt_coarse_energy_dist[f->size][intra];
330
331 /* Inter is really just differential coding */
332 if (opus_rc_tell(rc) + 3 <= f->framebits)
333 ff_opus_rc_enc_log(rc, intra, 3);
334 else
335 intra = 0;
336
337 if (intra) {
338 alpha = 0.0f;
339 beta = 1.0f - (4915.0f/32768.0f);
340 } else {
341 alpha = ff_celt_alpha_coef[f->size];
342 beta = ff_celt_beta_coef[f->size];
343 }
344
345 for (int i = f->start_band; i < f->end_band; i++) {
346 for (int ch = 0; ch < f->channels; ch++) {
347 CeltBlock *block = &f->block[ch];
348 const int left = f->framebits - opus_rc_tell(rc);
349 const float last = FFMAX(-9.0f, last_energy[ch][i]);
350 float diff = block->energy[i] - prev[ch] - last*alpha;
351 int q_en = lrintf(diff);
352 if (left >= 15) {
353 ff_opus_rc_enc_laplace(rc, &q_en, pmod[i << 1] << 7, pmod[(i << 1) + 1] << 6);
354 } else if (left >= 2) {
355 q_en = av_clip(q_en, -1, 1);
356 ff_opus_rc_enc_cdf(rc, 2*q_en + 3*(q_en < 0), ff_celt_model_energy_small);
357 } else if (left >= 1) {
358 q_en = av_clip(q_en, -1, 0);
359 ff_opus_rc_enc_log(rc, (q_en & 1), 1);
360 } else q_en = -1;
361
362 block->error_energy[i] = q_en - diff;
363 prev[ch] += beta * q_en;
364 }
365 }
366 }
367
368 static void celt_quant_coarse(CeltFrame *f, OpusRangeCoder *rc,
369 float last_energy[][CELT_MAX_BANDS])
370 {
371 uint32_t inter, intra;
372 OPUS_RC_CHECKPOINT_SPAWN(rc);
373
374 exp_quant_coarse(rc, f, last_energy, 1);
375 intra = OPUS_RC_CHECKPOINT_BITS(rc);
376
377 OPUS_RC_CHECKPOINT_ROLLBACK(rc);
378
379 exp_quant_coarse(rc, f, last_energy, 0);
380 inter = OPUS_RC_CHECKPOINT_BITS(rc);
381
382 if (inter > intra) { /* Unlikely */
383 OPUS_RC_CHECKPOINT_ROLLBACK(rc);
384 exp_quant_coarse(rc, f, last_energy, 1);
385 }
386 }
387
388 static void celt_quant_fine(CeltFrame *f, OpusRangeCoder *rc)
389 {
390 for (int i = f->start_band; i < f->end_band; i++) {
391 if (!f->fine_bits[i])
392 continue;
393 for (int ch = 0; ch < f->channels; ch++) {
394 CeltBlock *block = &f->block[ch];
395 int quant, lim = (1 << f->fine_bits[i]);
396 float offset, diff = 0.5f - block->error_energy[i];
397 quant = av_clip(floor(diff*lim), 0, lim - 1);
398 ff_opus_rc_put_raw(rc, quant, f->fine_bits[i]);
399 offset = 0.5f - ((quant + 0.5f) * (1 << (14 - f->fine_bits[i])) / 16384.0f);
400 block->error_energy[i] -= offset;
401 }
402 }
403 }
404
405 static void celt_quant_final(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f)
406 {
407 for (int priority = 0; priority < 2; priority++) {
408 for (int i = f->start_band; i < f->end_band && (f->framebits - opus_rc_tell(rc)) >= f->channels; i++) {
409 if (f->fine_priority[i] != priority || f->fine_bits[i] >= CELT_MAX_FINE_BITS)
410 continue;
411 for (int ch = 0; ch < f->channels; ch++) {
412 CeltBlock *block = &f->block[ch];
413 const float err = block->error_energy[i];
414 const float offset = 0.5f * (1 << (14 - f->fine_bits[i] - 1)) / 16384.0f;
415 const int sign = FFABS(err + offset) < FFABS(err - offset);
416 ff_opus_rc_put_raw(rc, sign, 1);
417 block->error_energy[i] -= offset*(1 - 2*sign);
418 }
419 }
420 }
421 }
422
423 static void celt_encode_frame(OpusEncContext *s, OpusRangeCoder *rc,
424 CeltFrame *f, int index)
425 {
426 ff_opus_rc_enc_init(rc);
427
428 ff_opus_psy_celt_frame_init(&s->psyctx, f, index);
429
430 celt_frame_setup_input(s, f);
431
432 if (f->silence) {
433 if (f->framebits >= 16)
434 ff_opus_rc_enc_log(rc, 1, 15); /* Silence (if using explicit singalling) */
435 for (int ch = 0; ch < s->channels; ch++)
436 memset(s->last_quantized_energy[ch], 0.0f, sizeof(float)*CELT_MAX_BANDS);
437 return;
438 }
439
440 /* Filters */
441 celt_apply_preemph_filter(s, f);
442 if (f->pfilter) {
443 ff_opus_rc_enc_log(rc, 0, 15);
444 celt_enc_quant_pfilter(rc, f);
445 }
446
447 /* Transform */
448 celt_frame_mdct(s, f);
449
450 /* Need to handle transient/non-transient switches at any point during analysis */
451 while (ff_opus_psy_celt_frame_process(&s->psyctx, f, index))
452 celt_frame_mdct(s, f);
453
454 ff_opus_rc_enc_init(rc);
455
456 /* Silence */
457 ff_opus_rc_enc_log(rc, 0, 15);
458
459 /* Pitch filter */
460 if (!f->start_band && opus_rc_tell(rc) + 16 <= f->framebits)
461 celt_enc_quant_pfilter(rc, f);
462
463 /* Transient flag */
464 if (f->size && opus_rc_tell(rc) + 3 <= f->framebits)
465 ff_opus_rc_enc_log(rc, f->transient, 3);
466
467 /* Main encoding */
468 celt_quant_coarse (f, rc, s->last_quantized_energy);
469 celt_enc_tf (f, rc);
470 ff_celt_bitalloc (f, rc, 1);
471 celt_quant_fine (f, rc);
472 ff_celt_quant_bands(f, rc);
473
474 /* Anticollapse bit */
475 if (f->anticollapse_needed)
476 ff_opus_rc_put_raw(rc, f->anticollapse, 1);
477
478 /* Final per-band energy adjustments from leftover bits */
479 celt_quant_final(s, rc, f);
480
481 for (int ch = 0; ch < f->channels; ch++) {
482 CeltBlock *block = &f->block[ch];
483 for (int i = 0; i < CELT_MAX_BANDS; i++)
484 s->last_quantized_energy[ch][i] = block->energy[i] + block->error_energy[i];
485 }
486 }
487
488 static inline int write_opuslacing(uint8_t *dst, int v)
489 {
490 dst[0] = FFMIN(v - FFALIGN(v - 255, 4), v);
491 dst[1] = v - dst[0] >> 2;
492 return 1 + (v >= 252);
493 }
494
495 static void opus_packet_assembler(OpusEncContext *s, AVPacket *avpkt)
496 {
497 int offset, fsize_needed;
498
499 /* Write toc */
500 opus_gen_toc(s, avpkt->data, &offset, &fsize_needed);
501
502 /* Frame sizes if needed */
503 if (fsize_needed) {
504 for (int i = 0; i < s->packet.frames - 1; i++) {
505 offset += write_opuslacing(avpkt->data + offset,
506 s->frame[i].framebits >> 3);
507 }
508 }
509
510 /* Packets */
511 for (int i = 0; i < s->packet.frames; i++) {
512 ff_opus_rc_enc_end(&s->rc[i], avpkt->data + offset,
513 s->frame[i].framebits >> 3);
514 offset += s->frame[i].framebits >> 3;
515 }
516
517 avpkt->size = offset;
518 }
519
520 /* Used as overlap for the first frame and padding for the last encoded packet */
521 static AVFrame *spawn_empty_frame(OpusEncContext *s)
522 {
523 AVFrame *f = av_frame_alloc();
524 int ret;
525 if (!f)
526 return NULL;
527 f->format = s->avctx->sample_fmt;
528 f->nb_samples = s->avctx->frame_size;
529 ret = av_channel_layout_copy(&f->ch_layout, &s->avctx->ch_layout);
530 if (ret < 0) {
531 av_frame_free(&f);
532 return NULL;
533 }
534 if (av_frame_get_buffer(f, 4)) {
535 av_frame_free(&f);
536 return NULL;
537 }
538 for (int i = 0; i < s->channels; i++) {
539 size_t bps = av_get_bytes_per_sample(f->format);
540 memset(f->extended_data[i], 0, bps*f->nb_samples);
541 }
542 return f;
543 }
544
545 static int opus_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
546 const AVFrame *frame, int *got_packet_ptr)
547 {
548 OpusEncContext *s = avctx->priv_data;
549 int ret, frame_size, alloc_size = 0;
550
551 if (frame) { /* Add new frame to queue */
552 if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
553 return ret;
554 ff_bufqueue_add(avctx, &s->bufqueue, av_frame_clone(frame));
555 } else {
556 ff_opus_psy_signal_eof(&s->psyctx);
557 if (!s->afq.remaining_samples || !avctx->frame_number)
558 return 0; /* We've been flushed and there's nothing left to encode */
559 }
560
561 /* Run the psychoacoustic system */
562 if (ff_opus_psy_process(&s->psyctx, &s->packet))
563 return 0;
564
565 frame_size = OPUS_BLOCK_SIZE(s->packet.framesize);
566
567 if (!frame) {
568 /* This can go negative, that's not a problem, we only pad if positive */
569 int pad_empty = s->packet.frames*(frame_size/s->avctx->frame_size) - s->bufqueue.available + 1;
570 /* Pad with empty 2.5 ms frames to whatever framesize was decided,
571 * this should only happen at the very last flush frame. The frames
572 * allocated here will be freed (because they have no other references)
573 * after they get used by celt_frame_setup_input() */
574 for (int i = 0; i < pad_empty; i++) {
575 AVFrame *empty = spawn_empty_frame(s);
576 if (!empty)
577 return AVERROR(ENOMEM);
578 ff_bufqueue_add(avctx, &s->bufqueue, empty);
579 }
580 }
581
582 for (int i = 0; i < s->packet.frames; i++) {
583 celt_encode_frame(s, &s->rc[i], &s->frame[i], i);
584 alloc_size += s->frame[i].framebits >> 3;
585 }
586
587 /* Worst case toc + the frame lengths if needed */
588 alloc_size += 2 + s->packet.frames*2;
589
590 if ((ret = ff_alloc_packet(avctx, avpkt, alloc_size)) < 0)
591 return ret;
592
593 /* Assemble packet */
594 opus_packet_assembler(s, avpkt);
595
596 /* Update the psychoacoustic system */
597 ff_opus_psy_postencode_update(&s->psyctx, s->frame);
598
599 /* Remove samples from queue and skip if needed */
600 ff_af_queue_remove(&s->afq, s->packet.frames*frame_size, &avpkt->pts, &avpkt->duration);
601 if (s->packet.frames*frame_size > avpkt->duration) {
602 uint8_t *side = av_packet_new_side_data(avpkt, AV_PKT_DATA_SKIP_SAMPLES, 10);
603 if (!side)
604 return AVERROR(ENOMEM);
605 AV_WL32(&side[4], s->packet.frames*frame_size - avpkt->duration + 120);
606 }
607
608 *got_packet_ptr = 1;
609
610 return 0;
611 }
612
613 static av_cold int opus_encode_end(AVCodecContext *avctx)
614 {
615 OpusEncContext *s = avctx->priv_data;
616
617 for (int i = 0; i < CELT_BLOCK_NB; i++)
618 av_tx_uninit(&s->tx[i]);
619
620 ff_celt_pvq_uninit(&s->pvq);
621 av_freep(&s->dsp);
622 av_freep(&s->frame);
623 av_freep(&s->rc);
624 ff_af_queue_close(&s->afq);
625 ff_opus_psy_end(&s->psyctx);
626 ff_bufqueue_discard_all(&s->bufqueue);
627
628 return 0;
629 }
630
631 static av_cold int opus_encode_init(AVCodecContext *avctx)
632 {
633 int ret, max_frames;
634 OpusEncContext *s = avctx->priv_data;
635
636 s->avctx = avctx;
637 s->channels = avctx->ch_layout.nb_channels;
638
639 /* Opus allows us to change the framesize on each packet (and each packet may
640 * have multiple frames in it) but we can't change the codec's frame size on
641 * runtime, so fix it to the lowest possible number of samples and use a queue
642 * to accumulate AVFrames until we have enough to encode whatever the encoder
643 * decides is the best */
644 avctx->frame_size = 120;
645 /* Initial padding will change if SILK is ever supported */
646 avctx->initial_padding = 120;
647
648 if (!avctx->bit_rate) {
649 int coupled = ff_opus_default_coupled_streams[s->channels - 1];
650 avctx->bit_rate = coupled*(96000) + (s->channels - coupled*2)*(48000);
651 } else if (avctx->bit_rate < 6000 || avctx->bit_rate > 255000 * s->channels) {
652 int64_t clipped_rate = av_clip(avctx->bit_rate, 6000, 255000 * s->channels);
653 av_log(avctx, AV_LOG_ERROR, "Unsupported bitrate %"PRId64" kbps, clipping to %"PRId64" kbps\n",
654 avctx->bit_rate/1000, clipped_rate/1000);
655 avctx->bit_rate = clipped_rate;
656 }
657
658 /* Extradata */
659 avctx->extradata_size = 19;
660 avctx->extradata = av_malloc(avctx->extradata_size + AV_INPUT_BUFFER_PADDING_SIZE);
661 if (!avctx->extradata)
662 return AVERROR(ENOMEM);
663 opus_write_extradata(avctx);
664
665 ff_af_queue_init(avctx, &s->afq);
666
667 if ((ret = ff_celt_pvq_init(&s->pvq, 1)) < 0)
668 return ret;
669
670 if (!(s->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT)))
671 return AVERROR(ENOMEM);
672
673 /* I have no idea why a base scaling factor of 68 works, could be the twiddles */
674 for (int i = 0; i < CELT_BLOCK_NB; i++) {
675 const float scale = 68 << (CELT_BLOCK_NB - 1 - i);
676 if ((ret = av_tx_init(&s->tx[i], &s->tx_fn[i], AV_TX_FLOAT_MDCT, 0, 15 << (i + 3), &scale, 0)))
677 return AVERROR(ENOMEM);
678 }
679
680 /* Zero out previous energy (matters for inter first frame) */
681 for (int ch = 0; ch < s->channels; ch++)
682 memset(s->last_quantized_energy[ch], 0.0f, sizeof(float)*CELT_MAX_BANDS);
683
684 /* Allocate an empty frame to use as overlap for the first frame of audio */
685 ff_bufqueue_add(avctx, &s->bufqueue, spawn_empty_frame(s));
686 if (!ff_bufqueue_peek(&s->bufqueue, 0))
687 return AVERROR(ENOMEM);
688
689 if ((ret = ff_opus_psy_init(&s->psyctx, s->avctx, &s->bufqueue, &s->options)))
690 return ret;
691
692 /* Frame structs and range coder buffers */
693 max_frames = ceilf(FFMIN(s->options.max_delay_ms, 120.0f)/2.5f);
694 s->frame = av_malloc(max_frames*sizeof(CeltFrame));
695 if (!s->frame)
696 return AVERROR(ENOMEM);
697 s->rc = av_malloc(max_frames*sizeof(OpusRangeCoder));
698 if (!s->rc)
699 return AVERROR(ENOMEM);
700
701 for (int i = 0; i < max_frames; i++) {
702 s->frame[i].dsp = s->dsp;
703 s->frame[i].avctx = s->avctx;
704 s->frame[i].seed = 0;
705 s->frame[i].pvq = s->pvq;
706 s->frame[i].apply_phase_inv = s->options.apply_phase_inv;
707 s->frame[i].block[0].emph_coeff = s->frame[i].block[1].emph_coeff = 0.0f;
708 }
709
710 return 0;
711 }
712
713 #define OPUSENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
714 static const AVOption opusenc_options[] = {
715 { "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" },
716 { "apply_phase_inv", "Apply intensity stereo phase inversion", offsetof(OpusEncContext, options.apply_phase_inv), AV_OPT_TYPE_BOOL, { .i64 = 1 }, 0, 1, OPUSENC_FLAGS, "apply_phase_inv" },
717 { NULL },
718 };
719
720 static const AVClass opusenc_class = {
721 .class_name = "Opus encoder",
722 .item_name = av_default_item_name,
723 .option = opusenc_options,
724 .version = LIBAVUTIL_VERSION_INT,
725 };
726
727 static const FFCodecDefault opusenc_defaults[] = {
728 { "b", "0" },
729 { "compression_level", "10" },
730 { NULL },
731 };
732
733 const FFCodec ff_opus_encoder = {
734 .p.name = "opus",
735 CODEC_LONG_NAME("Opus"),
736 .p.type = AVMEDIA_TYPE_AUDIO,
737 .p.id = AV_CODEC_ID_OPUS,
738 .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_DELAY |
739 AV_CODEC_CAP_SMALL_LAST_FRAME | AV_CODEC_CAP_EXPERIMENTAL,
740 .defaults = opusenc_defaults,
741 .p.priv_class = &opusenc_class,
742 .priv_data_size = sizeof(OpusEncContext),
743 .init = opus_encode_init,
744 FF_CODEC_ENCODE_CB(opus_encode_frame),
745 .close = opus_encode_end,
746 .caps_internal = FF_CODEC_CAP_INIT_CLEANUP,
747 .p.supported_samplerates = (const int []){ 48000, 0 },
748 CODEC_OLD_CHANNEL_LAYOUTS(AV_CH_LAYOUT_MONO, AV_CH_LAYOUT_STEREO)
749 .p.ch_layouts = (const AVChannelLayout []){ AV_CHANNEL_LAYOUT_MONO,
750 AV_CHANNEL_LAYOUT_STEREO, { 0 } },
751 .p.sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
752 AV_SAMPLE_FMT_NONE },
753 };
754