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 "opusenc.h"
23 #include "opus_pvq.h"
24 #include "opusenc_psy.h"
25 #include "opustab.h"
26
27 #include "libavutil/float_dsp.h"
28 #include "libavutil/mem_internal.h"
29 #include "libavutil/opt.h"
30 #include "internal.h"
31 #include "bytestream.h"
32 #include "audio_frame_queue.h"
33
34 typedef struct OpusEncContext {
35 AVClass *av_class;
36 OpusEncOptions options;
37 OpusPsyContext psyctx;
38 AVCodecContext *avctx;
39 AudioFrameQueue afq;
40 AVFloatDSPContext *dsp;
41 MDCT15Context *mdct[CELT_BLOCK_NB];
42 CeltPVQ *pvq;
43 struct FFBufQueue bufqueue;
44
45 uint8_t enc_id[64];
46 int enc_id_bits;
47
48 OpusPacketInfo packet;
49
50 int channels;
51
52 CeltFrame *frame;
53 OpusRangeCoder *rc;
54
55 /* Actual energy the decoder will have */
56 float last_quantized_energy[OPUS_MAX_CHANNELS][CELT_MAX_BANDS];
57
58 DECLARE_ALIGNED(32, float, scratch)[2048];
59 } OpusEncContext;
60
opus_write_extradata(AVCodecContext * avctx)61 static void opus_write_extradata(AVCodecContext *avctx)
62 {
63 uint8_t *bs = avctx->extradata;
64
65 bytestream_put_buffer(&bs, "OpusHead", 8);
66 bytestream_put_byte (&bs, 0x1);
67 bytestream_put_byte (&bs, avctx->channels);
68 bytestream_put_le16 (&bs, avctx->initial_padding);
69 bytestream_put_le32 (&bs, avctx->sample_rate);
70 bytestream_put_le16 (&bs, 0x0);
71 bytestream_put_byte (&bs, 0x0); /* Default layout */
72 }
73
opus_gen_toc(OpusEncContext * s,uint8_t * toc,int * size,int * fsize_needed)74 static int opus_gen_toc(OpusEncContext *s, uint8_t *toc, int *size, int *fsize_needed)
75 {
76 int tmp = 0x0, extended_toc = 0;
77 static const int toc_cfg[][OPUS_MODE_NB][OPUS_BANDWITH_NB] = {
78 /* Silk Hybrid Celt Layer */
79 /* NB MB WB SWB FB NB MB WB SWB FB NB MB WB SWB FB Bandwidth */
80 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 17, 0, 21, 25, 29 } }, /* 2.5 ms */
81 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 18, 0, 22, 26, 30 } }, /* 5 ms */
82 { { 1, 5, 9, 0, 0 }, { 0, 0, 0, 13, 15 }, { 19, 0, 23, 27, 31 } }, /* 10 ms */
83 { { 2, 6, 10, 0, 0 }, { 0, 0, 0, 14, 16 }, { 20, 0, 24, 28, 32 } }, /* 20 ms */
84 { { 3, 7, 11, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 40 ms */
85 { { 4, 8, 12, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 60 ms */
86 };
87 int cfg = toc_cfg[s->packet.framesize][s->packet.mode][s->packet.bandwidth];
88 *fsize_needed = 0;
89 if (!cfg)
90 return 1;
91 if (s->packet.frames == 2) { /* 2 packets */
92 if (s->frame[0].framebits == s->frame[1].framebits) { /* same size */
93 tmp = 0x1;
94 } else { /* different size */
95 tmp = 0x2;
96 *fsize_needed = 1; /* put frame sizes in the packet */
97 }
98 } else if (s->packet.frames > 2) {
99 tmp = 0x3;
100 extended_toc = 1;
101 }
102 tmp |= (s->channels > 1) << 2; /* Stereo or mono */
103 tmp |= (cfg - 1) << 3; /* codec configuration */
104 *toc++ = tmp;
105 if (extended_toc) {
106 for (int i = 0; i < (s->packet.frames - 1); i++)
107 *fsize_needed |= (s->frame[i].framebits != s->frame[i + 1].framebits);
108 tmp = (*fsize_needed) << 7; /* vbr flag */
109 tmp |= (0) << 6; /* padding flag */
110 tmp |= s->packet.frames;
111 *toc++ = tmp;
112 }
113 *size = 1 + extended_toc;
114 return 0;
115 }
116
celt_frame_setup_input(OpusEncContext * s,CeltFrame * f)117 static void celt_frame_setup_input(OpusEncContext *s, CeltFrame *f)
118 {
119 AVFrame *cur = NULL;
120 const int subframesize = s->avctx->frame_size;
121 int subframes = OPUS_BLOCK_SIZE(s->packet.framesize) / subframesize;
122
123 cur = ff_bufqueue_get(&s->bufqueue);
124
125 for (int 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);
130 }
131
132 av_frame_free(&cur);
133
134 for (int sf = 0; sf < subframes; sf++) {
135 if (sf != (subframes - 1))
136 cur = ff_bufqueue_get(&s->bufqueue);
137 else
138 cur = ff_bufqueue_peek(&s->bufqueue, 0);
139
140 for (int 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);
148 }
149
150 /* Last frame isn't popped off and freed yet - we need it for overlap */
151 if (sf != (subframes - 1))
152 av_frame_free(&cur);
153 }
154 }
155
156 /* Apply the pre emphasis filter */
celt_apply_preemph_filter(OpusEncContext * s,CeltFrame * f)157 static void celt_apply_preemph_filter(OpusEncContext *s, CeltFrame *f)
158 {
159 const int subframesize = s->avctx->frame_size;
160 const int subframes = OPUS_BLOCK_SIZE(s->packet.framesize) / subframesize;
161
162 /* Filter overlap */
163 for (int ch = 0; ch < f->channels; ch++) {
164 CeltBlock *b = &f->block[ch];
165 float m = b->emph_coeff;
166 for (int i = 0; i < CELT_OVERLAP; i++) {
167 float sample = b->overlap[i];
168 b->overlap[i] = sample - m;
169 m = sample * CELT_EMPH_COEFF;
170 }
171 b->emph_coeff = m;
172 }
173
174 /* Filter the samples but do not update the last subframe's coeff - overlap ^^^ */
175 for (int sf = 0; sf < subframes; sf++) {
176 for (int ch = 0; ch < f->channels; ch++) {
177 CeltBlock *b = &f->block[ch];
178 float m = b->emph_coeff;
179 for (int i = 0; i < subframesize; i++) {
180 float sample = b->samples[sf*subframesize + i];
181 b->samples[sf*subframesize + i] = sample - m;
182 m = sample * CELT_EMPH_COEFF;
183 }
184 if (sf != (subframes - 1))
185 b->emph_coeff = m;
186 }
187 }
188 }
189
190 /* Create the window and do the mdct */
celt_frame_mdct(OpusEncContext * s,CeltFrame * f)191 static void celt_frame_mdct(OpusEncContext *s, CeltFrame *f)
192 {
193 float *win = s->scratch, *temp = s->scratch + 1920;
194
195 if (f->transient) {
196 for (int ch = 0; ch < f->channels; ch++) {
197 CeltBlock *b = &f->block[ch];
198 float *src1 = b->overlap;
199 for (int t = 0; t < f->blocks; t++) {
200 float *src2 = &b->samples[CELT_OVERLAP*t];
201 s->dsp->vector_fmul(win, src1, ff_celt_window, 128);
202 s->dsp->vector_fmul_reverse(&win[CELT_OVERLAP], src2,
203 ff_celt_window - 8, 128);
204 src1 = src2;
205 s->mdct[0]->mdct(s->mdct[0], b->coeffs + t, win, f->blocks);
206 }
207 }
208 } else {
209 int blk_len = OPUS_BLOCK_SIZE(f->size), wlen = OPUS_BLOCK_SIZE(f->size + 1);
210 int rwin = blk_len - CELT_OVERLAP, lap_dst = (wlen - blk_len - CELT_OVERLAP) >> 1;
211 memset(win, 0, wlen*sizeof(float));
212 for (int ch = 0; ch < f->channels; ch++) {
213 CeltBlock *b = &f->block[ch];
214
215 /* Overlap */
216 s->dsp->vector_fmul(temp, b->overlap, ff_celt_window, 128);
217 memcpy(win + lap_dst, temp, CELT_OVERLAP*sizeof(float));
218
219 /* Samples, flat top window */
220 memcpy(&win[lap_dst + CELT_OVERLAP], b->samples, rwin*sizeof(float));
221
222 /* Samples, windowed */
223 s->dsp->vector_fmul_reverse(temp, b->samples + rwin,
224 ff_celt_window - 8, 128);
225 memcpy(win + lap_dst + blk_len, temp, CELT_OVERLAP*sizeof(float));
226
227 s->mdct[f->size]->mdct(s->mdct[f->size], b->coeffs, win, 1);
228 }
229 }
230
231 for (int ch = 0; ch < f->channels; ch++) {
232 CeltBlock *block = &f->block[ch];
233 for (int i = 0; i < CELT_MAX_BANDS; i++) {
234 float ener = 0.0f;
235 int band_offset = ff_celt_freq_bands[i] << f->size;
236 int band_size = ff_celt_freq_range[i] << f->size;
237 float *coeffs = &block->coeffs[band_offset];
238
239 for (int j = 0; j < band_size; j++)
240 ener += coeffs[j]*coeffs[j];
241
242 block->lin_energy[i] = sqrtf(ener) + FLT_EPSILON;
243 ener = 1.0f/block->lin_energy[i];
244
245 for (int j = 0; j < band_size; j++)
246 coeffs[j] *= ener;
247
248 block->energy[i] = log2f(block->lin_energy[i]) - ff_celt_mean_energy[i];
249
250 /* CELT_ENERGY_SILENCE is what the decoder uses and its not -infinity */
251 block->energy[i] = FFMAX(block->energy[i], CELT_ENERGY_SILENCE);
252 }
253 }
254 }
255
celt_enc_tf(CeltFrame * f,OpusRangeCoder * rc)256 static void celt_enc_tf(CeltFrame *f, OpusRangeCoder *rc)
257 {
258 int tf_select = 0, diff = 0, tf_changed = 0, tf_select_needed;
259 int bits = f->transient ? 2 : 4;
260
261 tf_select_needed = ((f->size && (opus_rc_tell(rc) + bits + 1) <= f->framebits));
262
263 for (int i = f->start_band; i < f->end_band; i++) {
264 if ((opus_rc_tell(rc) + bits + tf_select_needed) <= f->framebits) {
265 const int tbit = (diff ^ 1) == f->tf_change[i];
266 ff_opus_rc_enc_log(rc, tbit, bits);
267 diff ^= tbit;
268 tf_changed |= diff;
269 }
270 bits = f->transient ? 4 : 5;
271 }
272
273 if (tf_select_needed && ff_celt_tf_select[f->size][f->transient][0][tf_changed] !=
274 ff_celt_tf_select[f->size][f->transient][1][tf_changed]) {
275 ff_opus_rc_enc_log(rc, f->tf_select, 1);
276 tf_select = f->tf_select;
277 }
278
279 for (int i = f->start_band; i < f->end_band; i++)
280 f->tf_change[i] = ff_celt_tf_select[f->size][f->transient][tf_select][f->tf_change[i]];
281 }
282
celt_enc_quant_pfilter(OpusRangeCoder * rc,CeltFrame * f)283 static void celt_enc_quant_pfilter(OpusRangeCoder *rc, CeltFrame *f)
284 {
285 float gain = f->pf_gain;
286 int txval, octave = f->pf_octave, period = f->pf_period, tapset = f->pf_tapset;
287
288 ff_opus_rc_enc_log(rc, f->pfilter, 1);
289 if (!f->pfilter)
290 return;
291
292 /* Octave */
293 txval = FFMIN(octave, 6);
294 ff_opus_rc_enc_uint(rc, txval, 6);
295 octave = txval;
296 /* Period */
297 txval = av_clip(period - (16 << octave) + 1, 0, (1 << (4 + octave)) - 1);
298 ff_opus_rc_put_raw(rc, period, 4 + octave);
299 period = txval + (16 << octave) - 1;
300 /* Gain */
301 txval = FFMIN(((int)(gain / 0.09375f)) - 1, 7);
302 ff_opus_rc_put_raw(rc, txval, 3);
303 gain = 0.09375f * (txval + 1);
304 /* Tapset */
305 if ((opus_rc_tell(rc) + 2) <= f->framebits)
306 ff_opus_rc_enc_cdf(rc, tapset, ff_celt_model_tapset);
307 else
308 tapset = 0;
309 /* Finally create the coeffs */
310 for (int i = 0; i < 2; i++) {
311 CeltBlock *block = &f->block[i];
312
313 block->pf_period_new = FFMAX(period, CELT_POSTFILTER_MINPERIOD);
314 block->pf_gains_new[0] = gain * ff_celt_postfilter_taps[tapset][0];
315 block->pf_gains_new[1] = gain * ff_celt_postfilter_taps[tapset][1];
316 block->pf_gains_new[2] = gain * ff_celt_postfilter_taps[tapset][2];
317 }
318 }
319
exp_quant_coarse(OpusRangeCoder * rc,CeltFrame * f,float last_energy[][CELT_MAX_BANDS],int intra)320 static void exp_quant_coarse(OpusRangeCoder *rc, CeltFrame *f,
321 float last_energy[][CELT_MAX_BANDS], int intra)
322 {
323 float alpha, beta, prev[2] = { 0, 0 };
324 const uint8_t *pmod = ff_celt_coarse_energy_dist[f->size][intra];
325
326 /* Inter is really just differential coding */
327 if (opus_rc_tell(rc) + 3 <= f->framebits)
328 ff_opus_rc_enc_log(rc, intra, 3);
329 else
330 intra = 0;
331
332 if (intra) {
333 alpha = 0.0f;
334 beta = 1.0f - (4915.0f/32768.0f);
335 } else {
336 alpha = ff_celt_alpha_coef[f->size];
337 beta = ff_celt_beta_coef[f->size];
338 }
339
340 for (int i = f->start_band; i < f->end_band; i++) {
341 for (int ch = 0; ch < f->channels; ch++) {
342 CeltBlock *block = &f->block[ch];
343 const int left = f->framebits - opus_rc_tell(rc);
344 const float last = FFMAX(-9.0f, last_energy[ch][i]);
345 float diff = block->energy[i] - prev[ch] - last*alpha;
346 int q_en = lrintf(diff);
347 if (left >= 15) {
348 ff_opus_rc_enc_laplace(rc, &q_en, pmod[i << 1] << 7, pmod[(i << 1) + 1] << 6);
349 } else if (left >= 2) {
350 q_en = av_clip(q_en, -1, 1);
351 ff_opus_rc_enc_cdf(rc, 2*q_en + 3*(q_en < 0), ff_celt_model_energy_small);
352 } else if (left >= 1) {
353 q_en = av_clip(q_en, -1, 0);
354 ff_opus_rc_enc_log(rc, (q_en & 1), 1);
355 } else q_en = -1;
356
357 block->error_energy[i] = q_en - diff;
358 prev[ch] += beta * q_en;
359 }
360 }
361 }
362
celt_quant_coarse(CeltFrame * f,OpusRangeCoder * rc,float last_energy[][CELT_MAX_BANDS])363 static void celt_quant_coarse(CeltFrame *f, OpusRangeCoder *rc,
364 float last_energy[][CELT_MAX_BANDS])
365 {
366 uint32_t inter, intra;
367 OPUS_RC_CHECKPOINT_SPAWN(rc);
368
369 exp_quant_coarse(rc, f, last_energy, 1);
370 intra = OPUS_RC_CHECKPOINT_BITS(rc);
371
372 OPUS_RC_CHECKPOINT_ROLLBACK(rc);
373
374 exp_quant_coarse(rc, f, last_energy, 0);
375 inter = OPUS_RC_CHECKPOINT_BITS(rc);
376
377 if (inter > intra) { /* Unlikely */
378 OPUS_RC_CHECKPOINT_ROLLBACK(rc);
379 exp_quant_coarse(rc, f, last_energy, 1);
380 }
381 }
382
celt_quant_fine(CeltFrame * f,OpusRangeCoder * rc)383 static void celt_quant_fine(CeltFrame *f, OpusRangeCoder *rc)
384 {
385 for (int i = f->start_band; i < f->end_band; i++) {
386 if (!f->fine_bits[i])
387 continue;
388 for (int ch = 0; ch < f->channels; ch++) {
389 CeltBlock *block = &f->block[ch];
390 int quant, lim = (1 << f->fine_bits[i]);
391 float offset, diff = 0.5f - block->error_energy[i];
392 quant = av_clip(floor(diff*lim), 0, lim - 1);
393 ff_opus_rc_put_raw(rc, quant, f->fine_bits[i]);
394 offset = 0.5f - ((quant + 0.5f) * (1 << (14 - f->fine_bits[i])) / 16384.0f);
395 block->error_energy[i] -= offset;
396 }
397 }
398 }
399
celt_quant_final(OpusEncContext * s,OpusRangeCoder * rc,CeltFrame * f)400 static void celt_quant_final(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f)
401 {
402 for (int priority = 0; priority < 2; priority++) {
403 for (int i = f->start_band; i < f->end_band && (f->framebits - opus_rc_tell(rc)) >= f->channels; i++) {
404 if (f->fine_priority[i] != priority || f->fine_bits[i] >= CELT_MAX_FINE_BITS)
405 continue;
406 for (int ch = 0; ch < f->channels; ch++) {
407 CeltBlock *block = &f->block[ch];
408 const float err = block->error_energy[i];
409 const float offset = 0.5f * (1 << (14 - f->fine_bits[i] - 1)) / 16384.0f;
410 const int sign = FFABS(err + offset) < FFABS(err - offset);
411 ff_opus_rc_put_raw(rc, sign, 1);
412 block->error_energy[i] -= offset*(1 - 2*sign);
413 }
414 }
415 }
416 }
417
celt_encode_frame(OpusEncContext * s,OpusRangeCoder * rc,CeltFrame * f,int index)418 static void celt_encode_frame(OpusEncContext *s, OpusRangeCoder *rc,
419 CeltFrame *f, int index)
420 {
421 ff_opus_rc_enc_init(rc);
422
423 ff_opus_psy_celt_frame_init(&s->psyctx, f, index);
424
425 celt_frame_setup_input(s, f);
426
427 if (f->silence) {
428 if (f->framebits >= 16)
429 ff_opus_rc_enc_log(rc, 1, 15); /* Silence (if using explicit singalling) */
430 for (int ch = 0; ch < s->channels; ch++)
431 memset(s->last_quantized_energy[ch], 0.0f, sizeof(float)*CELT_MAX_BANDS);
432 return;
433 }
434
435 /* Filters */
436 celt_apply_preemph_filter(s, f);
437 if (f->pfilter) {
438 ff_opus_rc_enc_log(rc, 0, 15);
439 celt_enc_quant_pfilter(rc, f);
440 }
441
442 /* Transform */
443 celt_frame_mdct(s, f);
444
445 /* Need to handle transient/non-transient switches at any point during analysis */
446 while (ff_opus_psy_celt_frame_process(&s->psyctx, f, index))
447 celt_frame_mdct(s, f);
448
449 ff_opus_rc_enc_init(rc);
450
451 /* Silence */
452 ff_opus_rc_enc_log(rc, 0, 15);
453
454 /* Pitch filter */
455 if (!f->start_band && opus_rc_tell(rc) + 16 <= f->framebits)
456 celt_enc_quant_pfilter(rc, f);
457
458 /* Transient flag */
459 if (f->size && opus_rc_tell(rc) + 3 <= f->framebits)
460 ff_opus_rc_enc_log(rc, f->transient, 3);
461
462 /* Main encoding */
463 celt_quant_coarse (f, rc, s->last_quantized_energy);
464 celt_enc_tf (f, rc);
465 ff_celt_bitalloc (f, rc, 1);
466 celt_quant_fine (f, rc);
467 ff_celt_quant_bands(f, rc);
468
469 /* Anticollapse bit */
470 if (f->anticollapse_needed)
471 ff_opus_rc_put_raw(rc, f->anticollapse, 1);
472
473 /* Final per-band energy adjustments from leftover bits */
474 celt_quant_final(s, rc, f);
475
476 for (int ch = 0; ch < f->channels; ch++) {
477 CeltBlock *block = &f->block[ch];
478 for (int i = 0; i < CELT_MAX_BANDS; i++)
479 s->last_quantized_energy[ch][i] = block->energy[i] + block->error_energy[i];
480 }
481 }
482
write_opuslacing(uint8_t * dst,int v)483 static inline int write_opuslacing(uint8_t *dst, int v)
484 {
485 dst[0] = FFMIN(v - FFALIGN(v - 255, 4), v);
486 dst[1] = v - dst[0] >> 2;
487 return 1 + (v >= 252);
488 }
489
opus_packet_assembler(OpusEncContext * s,AVPacket * avpkt)490 static void opus_packet_assembler(OpusEncContext *s, AVPacket *avpkt)
491 {
492 int offset, fsize_needed;
493
494 /* Write toc */
495 opus_gen_toc(s, avpkt->data, &offset, &fsize_needed);
496
497 /* Frame sizes if needed */
498 if (fsize_needed) {
499 for (int i = 0; i < s->packet.frames - 1; i++) {
500 offset += write_opuslacing(avpkt->data + offset,
501 s->frame[i].framebits >> 3);
502 }
503 }
504
505 /* Packets */
506 for (int i = 0; i < s->packet.frames; i++) {
507 ff_opus_rc_enc_end(&s->rc[i], avpkt->data + offset,
508 s->frame[i].framebits >> 3);
509 offset += s->frame[i].framebits >> 3;
510 }
511
512 avpkt->size = offset;
513 }
514
515 /* Used as overlap for the first frame and padding for the last encoded packet */
spawn_empty_frame(OpusEncContext * s)516 static AVFrame *spawn_empty_frame(OpusEncContext *s)
517 {
518 AVFrame *f = av_frame_alloc();
519 if (!f)
520 return NULL;
521 f->format = s->avctx->sample_fmt;
522 f->nb_samples = s->avctx->frame_size;
523 f->channel_layout = s->avctx->channel_layout;
524 if (av_frame_get_buffer(f, 4)) {
525 av_frame_free(&f);
526 return NULL;
527 }
528 for (int i = 0; i < s->channels; i++) {
529 size_t bps = av_get_bytes_per_sample(f->format);
530 memset(f->extended_data[i], 0, bps*f->nb_samples);
531 }
532 return f;
533 }
534
opus_encode_frame(AVCodecContext * avctx,AVPacket * avpkt,const AVFrame * frame,int * got_packet_ptr)535 static int opus_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
536 const AVFrame *frame, int *got_packet_ptr)
537 {
538 OpusEncContext *s = avctx->priv_data;
539 int ret, frame_size, alloc_size = 0;
540
541 if (frame) { /* Add new frame to queue */
542 if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
543 return ret;
544 ff_bufqueue_add(avctx, &s->bufqueue, av_frame_clone(frame));
545 } else {
546 ff_opus_psy_signal_eof(&s->psyctx);
547 if (!s->afq.remaining_samples || !avctx->frame_number)
548 return 0; /* We've been flushed and there's nothing left to encode */
549 }
550
551 /* Run the psychoacoustic system */
552 if (ff_opus_psy_process(&s->psyctx, &s->packet))
553 return 0;
554
555 frame_size = OPUS_BLOCK_SIZE(s->packet.framesize);
556
557 if (!frame) {
558 /* This can go negative, that's not a problem, we only pad if positive */
559 int pad_empty = s->packet.frames*(frame_size/s->avctx->frame_size) - s->bufqueue.available + 1;
560 /* Pad with empty 2.5 ms frames to whatever framesize was decided,
561 * this should only happen at the very last flush frame. The frames
562 * allocated here will be freed (because they have no other references)
563 * after they get used by celt_frame_setup_input() */
564 for (int i = 0; i < pad_empty; i++) {
565 AVFrame *empty = spawn_empty_frame(s);
566 if (!empty)
567 return AVERROR(ENOMEM);
568 ff_bufqueue_add(avctx, &s->bufqueue, empty);
569 }
570 }
571
572 for (int i = 0; i < s->packet.frames; i++) {
573 celt_encode_frame(s, &s->rc[i], &s->frame[i], i);
574 alloc_size += s->frame[i].framebits >> 3;
575 }
576
577 /* Worst case toc + the frame lengths if needed */
578 alloc_size += 2 + s->packet.frames*2;
579
580 if ((ret = ff_alloc_packet2(avctx, avpkt, alloc_size, 0)) < 0)
581 return ret;
582
583 /* Assemble packet */
584 opus_packet_assembler(s, avpkt);
585
586 /* Update the psychoacoustic system */
587 ff_opus_psy_postencode_update(&s->psyctx, s->frame, s->rc);
588
589 /* Remove samples from queue and skip if needed */
590 ff_af_queue_remove(&s->afq, s->packet.frames*frame_size, &avpkt->pts, &avpkt->duration);
591 if (s->packet.frames*frame_size > avpkt->duration) {
592 uint8_t *side = av_packet_new_side_data(avpkt, AV_PKT_DATA_SKIP_SAMPLES, 10);
593 if (!side)
594 return AVERROR(ENOMEM);
595 AV_WL32(&side[4], s->packet.frames*frame_size - avpkt->duration + 120);
596 }
597
598 *got_packet_ptr = 1;
599
600 return 0;
601 }
602
opus_encode_end(AVCodecContext * avctx)603 static av_cold int opus_encode_end(AVCodecContext *avctx)
604 {
605 OpusEncContext *s = avctx->priv_data;
606
607 for (int i = 0; i < CELT_BLOCK_NB; i++)
608 ff_mdct15_uninit(&s->mdct[i]);
609
610 ff_celt_pvq_uninit(&s->pvq);
611 av_freep(&s->dsp);
612 av_freep(&s->frame);
613 av_freep(&s->rc);
614 ff_af_queue_close(&s->afq);
615 ff_opus_psy_end(&s->psyctx);
616 ff_bufqueue_discard_all(&s->bufqueue);
617 av_freep(&avctx->extradata);
618
619 return 0;
620 }
621
opus_encode_init(AVCodecContext * avctx)622 static av_cold int opus_encode_init(AVCodecContext *avctx)
623 {
624 int ret, max_frames;
625 OpusEncContext *s = avctx->priv_data;
626
627 s->avctx = avctx;
628 s->channels = avctx->channels;
629
630 /* Opus allows us to change the framesize on each packet (and each packet may
631 * have multiple frames in it) but we can't change the codec's frame size on
632 * runtime, so fix it to the lowest possible number of samples and use a queue
633 * to accumulate AVFrames until we have enough to encode whatever the encoder
634 * decides is the best */
635 avctx->frame_size = 120;
636 /* Initial padding will change if SILK is ever supported */
637 avctx->initial_padding = 120;
638
639 if (!avctx->bit_rate) {
640 int coupled = ff_opus_default_coupled_streams[s->channels - 1];
641 avctx->bit_rate = coupled*(96000) + (s->channels - coupled*2)*(48000);
642 } else if (avctx->bit_rate < 6000 || avctx->bit_rate > 255000 * s->channels) {
643 int64_t clipped_rate = av_clip(avctx->bit_rate, 6000, 255000 * s->channels);
644 av_log(avctx, AV_LOG_ERROR, "Unsupported bitrate %"PRId64" kbps, clipping to %"PRId64" kbps\n",
645 avctx->bit_rate/1000, clipped_rate/1000);
646 avctx->bit_rate = clipped_rate;
647 }
648
649 /* Extradata */
650 avctx->extradata_size = 19;
651 avctx->extradata = av_malloc(avctx->extradata_size + AV_INPUT_BUFFER_PADDING_SIZE);
652 if (!avctx->extradata)
653 return AVERROR(ENOMEM);
654 opus_write_extradata(avctx);
655
656 ff_af_queue_init(avctx, &s->afq);
657
658 if ((ret = ff_celt_pvq_init(&s->pvq, 1)) < 0)
659 return ret;
660
661 if (!(s->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT)))
662 return AVERROR(ENOMEM);
663
664 /* I have no idea why a base scaling factor of 68 works, could be the twiddles */
665 for (int i = 0; i < CELT_BLOCK_NB; i++)
666 if ((ret = ff_mdct15_init(&s->mdct[i], 0, i + 3, 68 << (CELT_BLOCK_NB - 1 - i))))
667 return AVERROR(ENOMEM);
668
669 /* Zero out previous energy (matters for inter first frame) */
670 for (int ch = 0; ch < s->channels; ch++)
671 memset(s->last_quantized_energy[ch], 0.0f, sizeof(float)*CELT_MAX_BANDS);
672
673 /* Allocate an empty frame to use as overlap for the first frame of audio */
674 ff_bufqueue_add(avctx, &s->bufqueue, spawn_empty_frame(s));
675 if (!ff_bufqueue_peek(&s->bufqueue, 0))
676 return AVERROR(ENOMEM);
677
678 if ((ret = ff_opus_psy_init(&s->psyctx, s->avctx, &s->bufqueue, &s->options)))
679 return ret;
680
681 /* Frame structs and range coder buffers */
682 max_frames = ceilf(FFMIN(s->options.max_delay_ms, 120.0f)/2.5f);
683 s->frame = av_malloc(max_frames*sizeof(CeltFrame));
684 if (!s->frame)
685 return AVERROR(ENOMEM);
686 s->rc = av_malloc(max_frames*sizeof(OpusRangeCoder));
687 if (!s->rc)
688 return AVERROR(ENOMEM);
689
690 for (int i = 0; i < max_frames; i++) {
691 s->frame[i].dsp = s->dsp;
692 s->frame[i].avctx = s->avctx;
693 s->frame[i].seed = 0;
694 s->frame[i].pvq = s->pvq;
695 s->frame[i].apply_phase_inv = s->options.apply_phase_inv;
696 s->frame[i].block[0].emph_coeff = s->frame[i].block[1].emph_coeff = 0.0f;
697 }
698
699 return 0;
700 }
701
702 #define OPUSENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
703 static const AVOption opusenc_options[] = {
704 { "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" },
705 { "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" },
706 { NULL },
707 };
708
709 static const AVClass opusenc_class = {
710 .class_name = "Opus encoder",
711 .item_name = av_default_item_name,
712 .option = opusenc_options,
713 .version = LIBAVUTIL_VERSION_INT,
714 };
715
716 static const AVCodecDefault opusenc_defaults[] = {
717 { "b", "0" },
718 { "compression_level", "10" },
719 { NULL },
720 };
721
722 AVCodec ff_opus_encoder = {
723 .name = "opus",
724 .long_name = NULL_IF_CONFIG_SMALL("Opus"),
725 .type = AVMEDIA_TYPE_AUDIO,
726 .id = AV_CODEC_ID_OPUS,
727 .defaults = opusenc_defaults,
728 .priv_class = &opusenc_class,
729 .priv_data_size = sizeof(OpusEncContext),
730 .init = opus_encode_init,
731 .encode2 = opus_encode_frame,
732 .close = opus_encode_end,
733 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE | FF_CODEC_CAP_INIT_CLEANUP,
734 .capabilities = AV_CODEC_CAP_SMALL_LAST_FRAME | AV_CODEC_CAP_DELAY,
735 .supported_samplerates = (const int []){ 48000, 0 },
736 .channel_layouts = (const uint64_t []){ AV_CH_LAYOUT_MONO,
737 AV_CH_LAYOUT_STEREO, 0 },
738 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
739 AV_SAMPLE_FMT_NONE },
740 };
741