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 "encode.h"
23 #include "opusenc.h"
24 #include "opus_pvq.h"
25 #include "opusenc_psy.h"
26 #include "opustab.h"
27
28 #include "libavutil/channel_layout.h"
29 #include "libavutil/float_dsp.h"
30 #include "libavutil/mem_internal.h"
31 #include "libavutil/opt.h"
32 #include "bytestream.h"
33 #include "audio_frame_queue.h"
34 #include "codec_internal.h"
35
36 typedef struct OpusEncContext {
37 AVClass *av_class;
38 OpusEncOptions options;
39 OpusPsyContext psyctx;
40 AVCodecContext *avctx;
41 AudioFrameQueue afq;
42 AVFloatDSPContext *dsp;
43 MDCT15Context *mdct[CELT_BLOCK_NB];
44 CeltPVQ *pvq;
45 struct FFBufQueue bufqueue;
46
47 uint8_t enc_id[64];
48 int enc_id_bits;
49
50 OpusPacketInfo packet;
51
52 int channels;
53
54 CeltFrame *frame;
55 OpusRangeCoder *rc;
56
57 /* Actual energy the decoder will have */
58 float last_quantized_energy[OPUS_MAX_CHANNELS][CELT_MAX_BANDS];
59
60 DECLARE_ALIGNED(32, float, scratch)[2048];
61 } OpusEncContext;
62
opus_write_extradata(AVCodecContext * avctx)63 static void opus_write_extradata(AVCodecContext *avctx)
64 {
65 uint8_t *bs = avctx->extradata;
66
67 bytestream_put_buffer(&bs, "OpusHead", 8);
68 bytestream_put_byte (&bs, 0x1);
69 bytestream_put_byte (&bs, avctx->ch_layout.nb_channels);
70 bytestream_put_le16 (&bs, avctx->initial_padding);
71 bytestream_put_le32 (&bs, avctx->sample_rate);
72 bytestream_put_le16 (&bs, 0x0);
73 bytestream_put_byte (&bs, 0x0); /* Default layout */
74 }
75
opus_gen_toc(OpusEncContext * s,uint8_t * toc,int * size,int * fsize_needed)76 static int opus_gen_toc(OpusEncContext *s, uint8_t *toc, int *size, int *fsize_needed)
77 {
78 int tmp = 0x0, extended_toc = 0;
79 static const int toc_cfg[][OPUS_MODE_NB][OPUS_BANDWITH_NB] = {
80 /* Silk Hybrid Celt Layer */
81 /* NB MB WB SWB FB NB MB WB SWB FB NB MB WB SWB FB Bandwidth */
82 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 17, 0, 21, 25, 29 } }, /* 2.5 ms */
83 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 18, 0, 22, 26, 30 } }, /* 5 ms */
84 { { 1, 5, 9, 0, 0 }, { 0, 0, 0, 13, 15 }, { 19, 0, 23, 27, 31 } }, /* 10 ms */
85 { { 2, 6, 10, 0, 0 }, { 0, 0, 0, 14, 16 }, { 20, 0, 24, 28, 32 } }, /* 20 ms */
86 { { 3, 7, 11, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 40 ms */
87 { { 4, 8, 12, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 60 ms */
88 };
89 int cfg = toc_cfg[s->packet.framesize][s->packet.mode][s->packet.bandwidth];
90 *fsize_needed = 0;
91 if (!cfg)
92 return 1;
93 if (s->packet.frames == 2) { /* 2 packets */
94 if (s->frame[0].framebits == s->frame[1].framebits) { /* same size */
95 tmp = 0x1;
96 } else { /* different size */
97 tmp = 0x2;
98 *fsize_needed = 1; /* put frame sizes in the packet */
99 }
100 } else if (s->packet.frames > 2) {
101 tmp = 0x3;
102 extended_toc = 1;
103 }
104 tmp |= (s->channels > 1) << 2; /* Stereo or mono */
105 tmp |= (cfg - 1) << 3; /* codec configuration */
106 *toc++ = tmp;
107 if (extended_toc) {
108 for (int i = 0; i < (s->packet.frames - 1); i++)
109 *fsize_needed |= (s->frame[i].framebits != s->frame[i + 1].framebits);
110 tmp = (*fsize_needed) << 7; /* vbr flag */
111 tmp |= (0) << 6; /* padding flag */
112 tmp |= s->packet.frames;
113 *toc++ = tmp;
114 }
115 *size = 1 + extended_toc;
116 return 0;
117 }
118
celt_frame_setup_input(OpusEncContext * s,CeltFrame * f)119 static void celt_frame_setup_input(OpusEncContext *s, CeltFrame *f)
120 {
121 AVFrame *cur = NULL;
122 const int subframesize = s->avctx->frame_size;
123 int subframes = OPUS_BLOCK_SIZE(s->packet.framesize) / subframesize;
124
125 cur = ff_bufqueue_get(&s->bufqueue);
126
127 for (int ch = 0; ch < f->channels; ch++) {
128 CeltBlock *b = &f->block[ch];
129 const void *input = cur->extended_data[ch];
130 size_t bps = av_get_bytes_per_sample(cur->format);
131 memcpy(b->overlap, input, bps*cur->nb_samples);
132 }
133
134 av_frame_free(&cur);
135
136 for (int sf = 0; sf < subframes; sf++) {
137 if (sf != (subframes - 1))
138 cur = ff_bufqueue_get(&s->bufqueue);
139 else
140 cur = ff_bufqueue_peek(&s->bufqueue, 0);
141
142 for (int ch = 0; ch < f->channels; ch++) {
143 CeltBlock *b = &f->block[ch];
144 const void *input = cur->extended_data[ch];
145 const size_t bps = av_get_bytes_per_sample(cur->format);
146 const size_t left = (subframesize - cur->nb_samples)*bps;
147 const size_t len = FFMIN(subframesize, cur->nb_samples)*bps;
148 memcpy(&b->samples[sf*subframesize], input, len);
149 memset(&b->samples[cur->nb_samples], 0, left);
150 }
151
152 /* Last frame isn't popped off and freed yet - we need it for overlap */
153 if (sf != (subframes - 1))
154 av_frame_free(&cur);
155 }
156 }
157
158 /* Apply the pre emphasis filter */
celt_apply_preemph_filter(OpusEncContext * s,CeltFrame * f)159 static void celt_apply_preemph_filter(OpusEncContext *s, CeltFrame *f)
160 {
161 const int subframesize = s->avctx->frame_size;
162 const int subframes = OPUS_BLOCK_SIZE(s->packet.framesize) / subframesize;
163
164 /* Filter overlap */
165 for (int ch = 0; ch < f->channels; ch++) {
166 CeltBlock *b = &f->block[ch];
167 float m = b->emph_coeff;
168 for (int i = 0; i < CELT_OVERLAP; i++) {
169 float sample = b->overlap[i];
170 b->overlap[i] = sample - m;
171 m = sample * CELT_EMPH_COEFF;
172 }
173 b->emph_coeff = m;
174 }
175
176 /* Filter the samples but do not update the last subframe's coeff - overlap ^^^ */
177 for (int sf = 0; sf < subframes; sf++) {
178 for (int ch = 0; ch < f->channels; ch++) {
179 CeltBlock *b = &f->block[ch];
180 float m = b->emph_coeff;
181 for (int i = 0; i < subframesize; i++) {
182 float sample = b->samples[sf*subframesize + i];
183 b->samples[sf*subframesize + i] = sample - m;
184 m = sample * CELT_EMPH_COEFF;
185 }
186 if (sf != (subframes - 1))
187 b->emph_coeff = m;
188 }
189 }
190 }
191
192 /* Create the window and do the mdct */
celt_frame_mdct(OpusEncContext * s,CeltFrame * f)193 static void celt_frame_mdct(OpusEncContext *s, CeltFrame *f)
194 {
195 float *win = s->scratch, *temp = s->scratch + 1920;
196
197 if (f->transient) {
198 for (int ch = 0; ch < f->channels; ch++) {
199 CeltBlock *b = &f->block[ch];
200 float *src1 = b->overlap;
201 for (int t = 0; t < f->blocks; t++) {
202 float *src2 = &b->samples[CELT_OVERLAP*t];
203 s->dsp->vector_fmul(win, src1, ff_celt_window, 128);
204 s->dsp->vector_fmul_reverse(&win[CELT_OVERLAP], src2,
205 ff_celt_window - 8, 128);
206 src1 = src2;
207 s->mdct[0]->mdct(s->mdct[0], b->coeffs + t, win, f->blocks);
208 }
209 }
210 } else {
211 int blk_len = OPUS_BLOCK_SIZE(f->size), wlen = OPUS_BLOCK_SIZE(f->size + 1);
212 int rwin = blk_len - CELT_OVERLAP, lap_dst = (wlen - blk_len - CELT_OVERLAP) >> 1;
213 memset(win, 0, wlen*sizeof(float));
214 for (int ch = 0; ch < f->channels; ch++) {
215 CeltBlock *b = &f->block[ch];
216
217 /* Overlap */
218 s->dsp->vector_fmul(temp, b->overlap, ff_celt_window, 128);
219 memcpy(win + lap_dst, temp, CELT_OVERLAP*sizeof(float));
220
221 /* Samples, flat top window */
222 memcpy(&win[lap_dst + CELT_OVERLAP], b->samples, rwin*sizeof(float));
223
224 /* Samples, windowed */
225 s->dsp->vector_fmul_reverse(temp, b->samples + rwin,
226 ff_celt_window - 8, 128);
227 memcpy(win + lap_dst + blk_len, temp, CELT_OVERLAP*sizeof(float));
228
229 s->mdct[f->size]->mdct(s->mdct[f->size], b->coeffs, win, 1);
230 }
231 }
232
233 for (int ch = 0; ch < f->channels; ch++) {
234 CeltBlock *block = &f->block[ch];
235 for (int i = 0; i < CELT_MAX_BANDS; i++) {
236 float ener = 0.0f;
237 int band_offset = ff_celt_freq_bands[i] << f->size;
238 int band_size = ff_celt_freq_range[i] << f->size;
239 float *coeffs = &block->coeffs[band_offset];
240
241 for (int j = 0; j < band_size; j++)
242 ener += coeffs[j]*coeffs[j];
243
244 block->lin_energy[i] = sqrtf(ener) + FLT_EPSILON;
245 ener = 1.0f/block->lin_energy[i];
246
247 for (int j = 0; j < band_size; j++)
248 coeffs[j] *= ener;
249
250 block->energy[i] = log2f(block->lin_energy[i]) - ff_celt_mean_energy[i];
251
252 /* CELT_ENERGY_SILENCE is what the decoder uses and its not -infinity */
253 block->energy[i] = FFMAX(block->energy[i], CELT_ENERGY_SILENCE);
254 }
255 }
256 }
257
celt_enc_tf(CeltFrame * f,OpusRangeCoder * rc)258 static void celt_enc_tf(CeltFrame *f, OpusRangeCoder *rc)
259 {
260 int tf_select = 0, diff = 0, tf_changed = 0, tf_select_needed;
261 int bits = f->transient ? 2 : 4;
262
263 tf_select_needed = ((f->size && (opus_rc_tell(rc) + bits + 1) <= f->framebits));
264
265 for (int i = f->start_band; i < f->end_band; i++) {
266 if ((opus_rc_tell(rc) + bits + tf_select_needed) <= f->framebits) {
267 const int tbit = (diff ^ 1) == f->tf_change[i];
268 ff_opus_rc_enc_log(rc, tbit, bits);
269 diff ^= tbit;
270 tf_changed |= diff;
271 }
272 bits = f->transient ? 4 : 5;
273 }
274
275 if (tf_select_needed && ff_celt_tf_select[f->size][f->transient][0][tf_changed] !=
276 ff_celt_tf_select[f->size][f->transient][1][tf_changed]) {
277 ff_opus_rc_enc_log(rc, f->tf_select, 1);
278 tf_select = f->tf_select;
279 }
280
281 for (int i = f->start_band; i < f->end_band; i++)
282 f->tf_change[i] = ff_celt_tf_select[f->size][f->transient][tf_select][f->tf_change[i]];
283 }
284
celt_enc_quant_pfilter(OpusRangeCoder * rc,CeltFrame * f)285 static void celt_enc_quant_pfilter(OpusRangeCoder *rc, CeltFrame *f)
286 {
287 float gain = f->pf_gain;
288 int txval, octave = f->pf_octave, period = f->pf_period, tapset = f->pf_tapset;
289
290 ff_opus_rc_enc_log(rc, f->pfilter, 1);
291 if (!f->pfilter)
292 return;
293
294 /* Octave */
295 txval = FFMIN(octave, 6);
296 ff_opus_rc_enc_uint(rc, txval, 6);
297 octave = txval;
298 /* Period */
299 txval = av_clip(period - (16 << octave) + 1, 0, (1 << (4 + octave)) - 1);
300 ff_opus_rc_put_raw(rc, period, 4 + octave);
301 period = txval + (16 << octave) - 1;
302 /* Gain */
303 txval = FFMIN(((int)(gain / 0.09375f)) - 1, 7);
304 ff_opus_rc_put_raw(rc, txval, 3);
305 gain = 0.09375f * (txval + 1);
306 /* Tapset */
307 if ((opus_rc_tell(rc) + 2) <= f->framebits)
308 ff_opus_rc_enc_cdf(rc, tapset, ff_celt_model_tapset);
309 else
310 tapset = 0;
311 /* Finally create the coeffs */
312 for (int i = 0; i < 2; i++) {
313 CeltBlock *block = &f->block[i];
314
315 block->pf_period_new = FFMAX(period, CELT_POSTFILTER_MINPERIOD);
316 block->pf_gains_new[0] = gain * ff_celt_postfilter_taps[tapset][0];
317 block->pf_gains_new[1] = gain * ff_celt_postfilter_taps[tapset][1];
318 block->pf_gains_new[2] = gain * ff_celt_postfilter_taps[tapset][2];
319 }
320 }
321
exp_quant_coarse(OpusRangeCoder * rc,CeltFrame * f,float last_energy[][CELT_MAX_BANDS],int intra)322 static void exp_quant_coarse(OpusRangeCoder *rc, CeltFrame *f,
323 float last_energy[][CELT_MAX_BANDS], int intra)
324 {
325 float alpha, beta, prev[2] = { 0, 0 };
326 const uint8_t *pmod = ff_celt_coarse_energy_dist[f->size][intra];
327
328 /* Inter is really just differential coding */
329 if (opus_rc_tell(rc) + 3 <= f->framebits)
330 ff_opus_rc_enc_log(rc, intra, 3);
331 else
332 intra = 0;
333
334 if (intra) {
335 alpha = 0.0f;
336 beta = 1.0f - (4915.0f/32768.0f);
337 } else {
338 alpha = ff_celt_alpha_coef[f->size];
339 beta = ff_celt_beta_coef[f->size];
340 }
341
342 for (int i = f->start_band; i < f->end_band; i++) {
343 for (int ch = 0; ch < f->channels; ch++) {
344 CeltBlock *block = &f->block[ch];
345 const int left = f->framebits - opus_rc_tell(rc);
346 const float last = FFMAX(-9.0f, last_energy[ch][i]);
347 float diff = block->energy[i] - prev[ch] - last*alpha;
348 int q_en = lrintf(diff);
349 if (left >= 15) {
350 ff_opus_rc_enc_laplace(rc, &q_en, pmod[i << 1] << 7, pmod[(i << 1) + 1] << 6);
351 } else if (left >= 2) {
352 q_en = av_clip(q_en, -1, 1);
353 ff_opus_rc_enc_cdf(rc, 2*q_en + 3*(q_en < 0), ff_celt_model_energy_small);
354 } else if (left >= 1) {
355 q_en = av_clip(q_en, -1, 0);
356 ff_opus_rc_enc_log(rc, (q_en & 1), 1);
357 } else q_en = -1;
358
359 block->error_energy[i] = q_en - diff;
360 prev[ch] += beta * q_en;
361 }
362 }
363 }
364
celt_quant_coarse(CeltFrame * f,OpusRangeCoder * rc,float last_energy[][CELT_MAX_BANDS])365 static void celt_quant_coarse(CeltFrame *f, OpusRangeCoder *rc,
366 float last_energy[][CELT_MAX_BANDS])
367 {
368 uint32_t inter, intra;
369 OPUS_RC_CHECKPOINT_SPAWN(rc);
370
371 exp_quant_coarse(rc, f, last_energy, 1);
372 intra = OPUS_RC_CHECKPOINT_BITS(rc);
373
374 OPUS_RC_CHECKPOINT_ROLLBACK(rc);
375
376 exp_quant_coarse(rc, f, last_energy, 0);
377 inter = OPUS_RC_CHECKPOINT_BITS(rc);
378
379 if (inter > intra) { /* Unlikely */
380 OPUS_RC_CHECKPOINT_ROLLBACK(rc);
381 exp_quant_coarse(rc, f, last_energy, 1);
382 }
383 }
384
celt_quant_fine(CeltFrame * f,OpusRangeCoder * rc)385 static void celt_quant_fine(CeltFrame *f, OpusRangeCoder *rc)
386 {
387 for (int i = f->start_band; i < f->end_band; i++) {
388 if (!f->fine_bits[i])
389 continue;
390 for (int ch = 0; ch < f->channels; ch++) {
391 CeltBlock *block = &f->block[ch];
392 int quant, lim = (1 << f->fine_bits[i]);
393 float offset, diff = 0.5f - block->error_energy[i];
394 quant = av_clip(floor(diff*lim), 0, lim - 1);
395 ff_opus_rc_put_raw(rc, quant, f->fine_bits[i]);
396 offset = 0.5f - ((quant + 0.5f) * (1 << (14 - f->fine_bits[i])) / 16384.0f);
397 block->error_energy[i] -= offset;
398 }
399 }
400 }
401
celt_quant_final(OpusEncContext * s,OpusRangeCoder * rc,CeltFrame * f)402 static void celt_quant_final(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f)
403 {
404 for (int priority = 0; priority < 2; priority++) {
405 for (int i = f->start_band; i < f->end_band && (f->framebits - opus_rc_tell(rc)) >= f->channels; i++) {
406 if (f->fine_priority[i] != priority || f->fine_bits[i] >= CELT_MAX_FINE_BITS)
407 continue;
408 for (int ch = 0; ch < f->channels; ch++) {
409 CeltBlock *block = &f->block[ch];
410 const float err = block->error_energy[i];
411 const float offset = 0.5f * (1 << (14 - f->fine_bits[i] - 1)) / 16384.0f;
412 const int sign = FFABS(err + offset) < FFABS(err - offset);
413 ff_opus_rc_put_raw(rc, sign, 1);
414 block->error_energy[i] -= offset*(1 - 2*sign);
415 }
416 }
417 }
418 }
419
celt_encode_frame(OpusEncContext * s,OpusRangeCoder * rc,CeltFrame * f,int index)420 static void celt_encode_frame(OpusEncContext *s, OpusRangeCoder *rc,
421 CeltFrame *f, int index)
422 {
423 ff_opus_rc_enc_init(rc);
424
425 ff_opus_psy_celt_frame_init(&s->psyctx, f, index);
426
427 celt_frame_setup_input(s, f);
428
429 if (f->silence) {
430 if (f->framebits >= 16)
431 ff_opus_rc_enc_log(rc, 1, 15); /* Silence (if using explicit singalling) */
432 for (int ch = 0; ch < s->channels; ch++)
433 memset(s->last_quantized_energy[ch], 0.0f, sizeof(float)*CELT_MAX_BANDS);
434 return;
435 }
436
437 /* Filters */
438 celt_apply_preemph_filter(s, f);
439 if (f->pfilter) {
440 ff_opus_rc_enc_log(rc, 0, 15);
441 celt_enc_quant_pfilter(rc, f);
442 }
443
444 /* Transform */
445 celt_frame_mdct(s, f);
446
447 /* Need to handle transient/non-transient switches at any point during analysis */
448 while (ff_opus_psy_celt_frame_process(&s->psyctx, f, index))
449 celt_frame_mdct(s, f);
450
451 ff_opus_rc_enc_init(rc);
452
453 /* Silence */
454 ff_opus_rc_enc_log(rc, 0, 15);
455
456 /* Pitch filter */
457 if (!f->start_band && opus_rc_tell(rc) + 16 <= f->framebits)
458 celt_enc_quant_pfilter(rc, f);
459
460 /* Transient flag */
461 if (f->size && opus_rc_tell(rc) + 3 <= f->framebits)
462 ff_opus_rc_enc_log(rc, f->transient, 3);
463
464 /* Main encoding */
465 celt_quant_coarse (f, rc, s->last_quantized_energy);
466 celt_enc_tf (f, rc);
467 ff_celt_bitalloc (f, rc, 1);
468 celt_quant_fine (f, rc);
469 ff_celt_quant_bands(f, rc);
470
471 /* Anticollapse bit */
472 if (f->anticollapse_needed)
473 ff_opus_rc_put_raw(rc, f->anticollapse, 1);
474
475 /* Final per-band energy adjustments from leftover bits */
476 celt_quant_final(s, rc, f);
477
478 for (int ch = 0; ch < f->channels; ch++) {
479 CeltBlock *block = &f->block[ch];
480 for (int i = 0; i < CELT_MAX_BANDS; i++)
481 s->last_quantized_energy[ch][i] = block->energy[i] + block->error_energy[i];
482 }
483 }
484
write_opuslacing(uint8_t * dst,int v)485 static inline int write_opuslacing(uint8_t *dst, int v)
486 {
487 dst[0] = FFMIN(v - FFALIGN(v - 255, 4), v);
488 dst[1] = v - dst[0] >> 2;
489 return 1 + (v >= 252);
490 }
491
opus_packet_assembler(OpusEncContext * s,AVPacket * avpkt)492 static void opus_packet_assembler(OpusEncContext *s, AVPacket *avpkt)
493 {
494 int offset, fsize_needed;
495
496 /* Write toc */
497 opus_gen_toc(s, avpkt->data, &offset, &fsize_needed);
498
499 /* Frame sizes if needed */
500 if (fsize_needed) {
501 for (int i = 0; i < s->packet.frames - 1; i++) {
502 offset += write_opuslacing(avpkt->data + offset,
503 s->frame[i].framebits >> 3);
504 }
505 }
506
507 /* Packets */
508 for (int i = 0; i < s->packet.frames; i++) {
509 ff_opus_rc_enc_end(&s->rc[i], avpkt->data + offset,
510 s->frame[i].framebits >> 3);
511 offset += s->frame[i].framebits >> 3;
512 }
513
514 avpkt->size = offset;
515 }
516
517 /* Used as overlap for the first frame and padding for the last encoded packet */
spawn_empty_frame(OpusEncContext * s)518 static AVFrame *spawn_empty_frame(OpusEncContext *s)
519 {
520 AVFrame *f = av_frame_alloc();
521 int ret;
522 if (!f)
523 return NULL;
524 f->format = s->avctx->sample_fmt;
525 f->nb_samples = s->avctx->frame_size;
526 ret = av_channel_layout_copy(&f->ch_layout, &s->avctx->ch_layout);
527 if (ret < 0) {
528 av_frame_free(&f);
529 return NULL;
530 }
531 if (av_frame_get_buffer(f, 4)) {
532 av_frame_free(&f);
533 return NULL;
534 }
535 for (int i = 0; i < s->channels; i++) {
536 size_t bps = av_get_bytes_per_sample(f->format);
537 memset(f->extended_data[i], 0, bps*f->nb_samples);
538 }
539 return f;
540 }
541
opus_encode_frame(AVCodecContext * avctx,AVPacket * avpkt,const AVFrame * frame,int * got_packet_ptr)542 static int opus_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
543 const AVFrame *frame, int *got_packet_ptr)
544 {
545 OpusEncContext *s = avctx->priv_data;
546 int ret, frame_size, alloc_size = 0;
547
548 if (frame) { /* Add new frame to queue */
549 if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
550 return ret;
551 ff_bufqueue_add(avctx, &s->bufqueue, av_frame_clone(frame));
552 } else {
553 ff_opus_psy_signal_eof(&s->psyctx);
554 if (!s->afq.remaining_samples || !avctx->frame_number)
555 return 0; /* We've been flushed and there's nothing left to encode */
556 }
557
558 /* Run the psychoacoustic system */
559 if (ff_opus_psy_process(&s->psyctx, &s->packet))
560 return 0;
561
562 frame_size = OPUS_BLOCK_SIZE(s->packet.framesize);
563
564 if (!frame) {
565 /* This can go negative, that's not a problem, we only pad if positive */
566 int pad_empty = s->packet.frames*(frame_size/s->avctx->frame_size) - s->bufqueue.available + 1;
567 /* Pad with empty 2.5 ms frames to whatever framesize was decided,
568 * this should only happen at the very last flush frame. The frames
569 * allocated here will be freed (because they have no other references)
570 * after they get used by celt_frame_setup_input() */
571 for (int i = 0; i < pad_empty; i++) {
572 AVFrame *empty = spawn_empty_frame(s);
573 if (!empty)
574 return AVERROR(ENOMEM);
575 ff_bufqueue_add(avctx, &s->bufqueue, empty);
576 }
577 }
578
579 for (int i = 0; i < s->packet.frames; i++) {
580 celt_encode_frame(s, &s->rc[i], &s->frame[i], i);
581 alloc_size += s->frame[i].framebits >> 3;
582 }
583
584 /* Worst case toc + the frame lengths if needed */
585 alloc_size += 2 + s->packet.frames*2;
586
587 if ((ret = ff_alloc_packet(avctx, avpkt, alloc_size)) < 0)
588 return ret;
589
590 /* Assemble packet */
591 opus_packet_assembler(s, avpkt);
592
593 /* Update the psychoacoustic system */
594 ff_opus_psy_postencode_update(&s->psyctx, s->frame, s->rc);
595
596 /* Remove samples from queue and skip if needed */
597 ff_af_queue_remove(&s->afq, s->packet.frames*frame_size, &avpkt->pts, &avpkt->duration);
598 if (s->packet.frames*frame_size > avpkt->duration) {
599 uint8_t *side = av_packet_new_side_data(avpkt, AV_PKT_DATA_SKIP_SAMPLES, 10);
600 if (!side)
601 return AVERROR(ENOMEM);
602 AV_WL32(&side[4], s->packet.frames*frame_size - avpkt->duration + 120);
603 }
604
605 *got_packet_ptr = 1;
606
607 return 0;
608 }
609
opus_encode_end(AVCodecContext * avctx)610 static av_cold int opus_encode_end(AVCodecContext *avctx)
611 {
612 OpusEncContext *s = avctx->priv_data;
613
614 for (int i = 0; i < CELT_BLOCK_NB; i++)
615 ff_mdct15_uninit(&s->mdct[i]);
616
617 ff_celt_pvq_uninit(&s->pvq);
618 av_freep(&s->dsp);
619 av_freep(&s->frame);
620 av_freep(&s->rc);
621 ff_af_queue_close(&s->afq);
622 ff_opus_psy_end(&s->psyctx);
623 ff_bufqueue_discard_all(&s->bufqueue);
624
625 return 0;
626 }
627
opus_encode_init(AVCodecContext * avctx)628 static av_cold int opus_encode_init(AVCodecContext *avctx)
629 {
630 int ret, max_frames;
631 OpusEncContext *s = avctx->priv_data;
632
633 s->avctx = avctx;
634 s->channels = avctx->ch_layout.nb_channels;
635
636 /* Opus allows us to change the framesize on each packet (and each packet may
637 * have multiple frames in it) but we can't change the codec's frame size on
638 * runtime, so fix it to the lowest possible number of samples and use a queue
639 * to accumulate AVFrames until we have enough to encode whatever the encoder
640 * decides is the best */
641 avctx->frame_size = 120;
642 /* Initial padding will change if SILK is ever supported */
643 avctx->initial_padding = 120;
644
645 if (!avctx->bit_rate) {
646 int coupled = ff_opus_default_coupled_streams[s->channels - 1];
647 avctx->bit_rate = coupled*(96000) + (s->channels - coupled*2)*(48000);
648 } else if (avctx->bit_rate < 6000 || avctx->bit_rate > 255000 * s->channels) {
649 int64_t clipped_rate = av_clip(avctx->bit_rate, 6000, 255000 * s->channels);
650 av_log(avctx, AV_LOG_ERROR, "Unsupported bitrate %"PRId64" kbps, clipping to %"PRId64" kbps\n",
651 avctx->bit_rate/1000, clipped_rate/1000);
652 avctx->bit_rate = clipped_rate;
653 }
654
655 /* Extradata */
656 avctx->extradata_size = 19;
657 avctx->extradata = av_malloc(avctx->extradata_size + AV_INPUT_BUFFER_PADDING_SIZE);
658 if (!avctx->extradata)
659 return AVERROR(ENOMEM);
660 opus_write_extradata(avctx);
661
662 ff_af_queue_init(avctx, &s->afq);
663
664 if ((ret = ff_celt_pvq_init(&s->pvq, 1)) < 0)
665 return ret;
666
667 if (!(s->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT)))
668 return AVERROR(ENOMEM);
669
670 /* I have no idea why a base scaling factor of 68 works, could be the twiddles */
671 for (int i = 0; i < CELT_BLOCK_NB; i++)
672 if ((ret = ff_mdct15_init(&s->mdct[i], 0, i + 3, 68 << (CELT_BLOCK_NB - 1 - i))))
673 return AVERROR(ENOMEM);
674
675 /* Zero out previous energy (matters for inter first frame) */
676 for (int ch = 0; ch < s->channels; ch++)
677 memset(s->last_quantized_energy[ch], 0.0f, sizeof(float)*CELT_MAX_BANDS);
678
679 /* Allocate an empty frame to use as overlap for the first frame of audio */
680 ff_bufqueue_add(avctx, &s->bufqueue, spawn_empty_frame(s));
681 if (!ff_bufqueue_peek(&s->bufqueue, 0))
682 return AVERROR(ENOMEM);
683
684 if ((ret = ff_opus_psy_init(&s->psyctx, s->avctx, &s->bufqueue, &s->options)))
685 return ret;
686
687 /* Frame structs and range coder buffers */
688 max_frames = ceilf(FFMIN(s->options.max_delay_ms, 120.0f)/2.5f);
689 s->frame = av_malloc(max_frames*sizeof(CeltFrame));
690 if (!s->frame)
691 return AVERROR(ENOMEM);
692 s->rc = av_malloc(max_frames*sizeof(OpusRangeCoder));
693 if (!s->rc)
694 return AVERROR(ENOMEM);
695
696 for (int i = 0; i < max_frames; i++) {
697 s->frame[i].dsp = s->dsp;
698 s->frame[i].avctx = s->avctx;
699 s->frame[i].seed = 0;
700 s->frame[i].pvq = s->pvq;
701 s->frame[i].apply_phase_inv = s->options.apply_phase_inv;
702 s->frame[i].block[0].emph_coeff = s->frame[i].block[1].emph_coeff = 0.0f;
703 }
704
705 return 0;
706 }
707
708 #define OPUSENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
709 static const AVOption opusenc_options[] = {
710 { "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" },
711 { "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" },
712 { NULL },
713 };
714
715 static const AVClass opusenc_class = {
716 .class_name = "Opus encoder",
717 .item_name = av_default_item_name,
718 .option = opusenc_options,
719 .version = LIBAVUTIL_VERSION_INT,
720 };
721
722 static const FFCodecDefault opusenc_defaults[] = {
723 { "b", "0" },
724 { "compression_level", "10" },
725 { NULL },
726 };
727
728 const FFCodec ff_opus_encoder = {
729 .p.name = "opus",
730 .p.long_name = NULL_IF_CONFIG_SMALL("Opus"),
731 .p.type = AVMEDIA_TYPE_AUDIO,
732 .p.id = AV_CODEC_ID_OPUS,
733 .defaults = opusenc_defaults,
734 .p.priv_class = &opusenc_class,
735 .priv_data_size = sizeof(OpusEncContext),
736 .init = opus_encode_init,
737 FF_CODEC_ENCODE_CB(opus_encode_frame),
738 .close = opus_encode_end,
739 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE | FF_CODEC_CAP_INIT_CLEANUP,
740 .p.capabilities = AV_CODEC_CAP_EXPERIMENTAL | AV_CODEC_CAP_SMALL_LAST_FRAME | AV_CODEC_CAP_DELAY,
741 .p.supported_samplerates = (const int []){ 48000, 0 },
742 #if FF_API_OLD_CHANNEL_LAYOUT
743 .p.channel_layouts = (const uint64_t []){ AV_CH_LAYOUT_MONO,
744 AV_CH_LAYOUT_STEREO, 0 },
745 #endif
746 .p.ch_layouts = (const AVChannelLayout []){ AV_CHANNEL_LAYOUT_MONO,
747 AV_CHANNEL_LAYOUT_STEREO, { 0 } },
748 .p.sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
749 AV_SAMPLE_FMT_NONE },
750 };
751