3 * Copyright (c) 2017 Rostislav Pehlivanov <atomnuker@gmail.com>
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 #include "opusenc_psy.h"
27 #include "libavutil/float_dsp.h"
28 #include "libavutil/opt.h"
30 #include "bytestream.h"
31 #include "audio_frame_queue.h"
33 typedef struct OpusEncContext {
35 OpusEncOptions options;
36 OpusPsyContext psyctx;
37 AVCodecContext *avctx;
39 AVFloatDSPContext *dsp;
40 MDCT15Context *mdct[CELT_BLOCK_NB];
42 struct FFBufQueue bufqueue;
47 OpusPacketInfo packet;
54 /* Actual energy the decoder will have */
55 float last_quantized_energy[OPUS_MAX_CHANNELS][CELT_MAX_BANDS];
57 DECLARE_ALIGNED(32, float, scratch)[2048];
60 static void opus_write_extradata(AVCodecContext *avctx)
62 uint8_t *bs = avctx->extradata;
64 bytestream_put_buffer(&bs, "OpusHead", 8);
65 bytestream_put_byte (&bs, 0x1);
66 bytestream_put_byte (&bs, avctx->channels);
67 bytestream_put_le16 (&bs, avctx->initial_padding);
68 bytestream_put_le32 (&bs, avctx->sample_rate);
69 bytestream_put_le16 (&bs, 0x0);
70 bytestream_put_byte (&bs, 0x0); /* Default layout */
73 static int opus_gen_toc(OpusEncContext *s, uint8_t *toc, int *size, int *fsize_needed)
75 int tmp = 0x0, extended_toc = 0;
76 static const int toc_cfg[][OPUS_MODE_NB][OPUS_BANDWITH_NB] = {
77 /* Silk Hybrid Celt Layer */
78 /* NB MB WB SWB FB NB MB WB SWB FB NB MB WB SWB FB Bandwidth */
79 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 17, 0, 21, 25, 29 } }, /* 2.5 ms */
80 { { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 }, { 18, 0, 22, 26, 30 } }, /* 5 ms */
81 { { 1, 5, 9, 0, 0 }, { 0, 0, 0, 13, 15 }, { 19, 0, 23, 27, 31 } }, /* 10 ms */
82 { { 2, 6, 10, 0, 0 }, { 0, 0, 0, 14, 16 }, { 20, 0, 24, 28, 32 } }, /* 20 ms */
83 { { 3, 7, 11, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 40 ms */
84 { { 4, 8, 12, 0, 0 }, { 0, 0, 0, 0, 0 }, { 0, 0, 0, 0, 0 } }, /* 60 ms */
86 int cfg = toc_cfg[s->packet.framesize][s->packet.mode][s->packet.bandwidth];
90 if (s->packet.frames == 2) { /* 2 packets */
91 if (s->frame[0].framebits == s->frame[1].framebits) { /* same size */
93 } else { /* different size */
95 *fsize_needed = 1; /* put frame sizes in the packet */
97 } else if (s->packet.frames > 2) {
101 tmp |= (s->channels > 1) << 2; /* Stereo or mono */
102 tmp |= (cfg - 1) << 3; /* codec configuration */
105 for (int i = 0; i < (s->packet.frames - 1); i++)
106 *fsize_needed |= (s->frame[i].framebits != s->frame[i + 1].framebits);
107 tmp = (*fsize_needed) << 7; /* vbr flag */
108 tmp |= (0) << 6; /* padding flag */
109 tmp |= s->packet.frames;
112 *size = 1 + extended_toc;
116 static void celt_frame_setup_input(OpusEncContext *s, CeltFrame *f)
119 const int subframesize = s->avctx->frame_size;
120 int subframes = OPUS_BLOCK_SIZE(s->packet.framesize) / subframesize;
122 cur = ff_bufqueue_get(&s->bufqueue);
124 for (int ch = 0; ch < f->channels; ch++) {
125 CeltBlock *b = &f->block[ch];
126 const void *input = cur->extended_data[ch];
127 size_t bps = av_get_bytes_per_sample(cur->format);
128 memcpy(b->overlap, input, bps*cur->nb_samples);
133 for (int sf = 0; sf < subframes; sf++) {
134 if (sf != (subframes - 1))
135 cur = ff_bufqueue_get(&s->bufqueue);
137 cur = ff_bufqueue_peek(&s->bufqueue, 0);
139 for (int ch = 0; ch < f->channels; ch++) {
140 CeltBlock *b = &f->block[ch];
141 const void *input = cur->extended_data[ch];
142 const size_t bps = av_get_bytes_per_sample(cur->format);
143 const size_t left = (subframesize - cur->nb_samples)*bps;
144 const size_t len = FFMIN(subframesize, cur->nb_samples)*bps;
145 memcpy(&b->samples[sf*subframesize], input, len);
146 memset(&b->samples[cur->nb_samples], 0, left);
149 /* Last frame isn't popped off and freed yet - we need it for overlap */
150 if (sf != (subframes - 1))
155 /* Apply the pre emphasis filter */
156 static void celt_apply_preemph_filter(OpusEncContext *s, CeltFrame *f)
158 const int subframesize = s->avctx->frame_size;
159 const int subframes = OPUS_BLOCK_SIZE(s->packet.framesize) / subframesize;
162 for (int ch = 0; ch < f->channels; ch++) {
163 CeltBlock *b = &f->block[ch];
164 float m = b->emph_coeff;
165 for (int i = 0; i < CELT_OVERLAP; i++) {
166 float sample = b->overlap[i];
167 b->overlap[i] = sample - m;
168 m = sample * CELT_EMPH_COEFF;
173 /* Filter the samples but do not update the last subframe's coeff - overlap ^^^ */
174 for (int sf = 0; sf < subframes; sf++) {
175 for (int ch = 0; ch < f->channels; ch++) {
176 CeltBlock *b = &f->block[ch];
177 float m = b->emph_coeff;
178 for (int i = 0; i < subframesize; i++) {
179 float sample = b->samples[sf*subframesize + i];
180 b->samples[sf*subframesize + i] = sample - m;
181 m = sample * CELT_EMPH_COEFF;
183 if (sf != (subframes - 1))
189 /* Create the window and do the mdct */
190 static void celt_frame_mdct(OpusEncContext *s, CeltFrame *f)
192 float *win = s->scratch, *temp = s->scratch + 1920;
195 for (int ch = 0; ch < f->channels; ch++) {
196 CeltBlock *b = &f->block[ch];
197 float *src1 = b->overlap;
198 for (int t = 0; t < f->blocks; t++) {
199 float *src2 = &b->samples[CELT_OVERLAP*t];
200 s->dsp->vector_fmul(win, src1, ff_celt_window, 128);
201 s->dsp->vector_fmul_reverse(&win[CELT_OVERLAP], src2,
202 ff_celt_window - 8, 128);
204 s->mdct[0]->mdct(s->mdct[0], b->coeffs + t, win, f->blocks);
208 int blk_len = OPUS_BLOCK_SIZE(f->size), wlen = OPUS_BLOCK_SIZE(f->size + 1);
209 int rwin = blk_len - CELT_OVERLAP, lap_dst = (wlen - blk_len - CELT_OVERLAP) >> 1;
210 memset(win, 0, wlen*sizeof(float));
211 for (int ch = 0; ch < f->channels; ch++) {
212 CeltBlock *b = &f->block[ch];
215 s->dsp->vector_fmul(temp, b->overlap, ff_celt_window, 128);
216 memcpy(win + lap_dst, temp, CELT_OVERLAP*sizeof(float));
218 /* Samples, flat top window */
219 memcpy(&win[lap_dst + CELT_OVERLAP], b->samples, rwin*sizeof(float));
221 /* Samples, windowed */
222 s->dsp->vector_fmul_reverse(temp, b->samples + rwin,
223 ff_celt_window - 8, 128);
224 memcpy(win + lap_dst + blk_len, temp, CELT_OVERLAP*sizeof(float));
226 s->mdct[f->size]->mdct(s->mdct[f->size], b->coeffs, win, 1);
230 for (int ch = 0; ch < f->channels; ch++) {
231 CeltBlock *block = &f->block[ch];
232 for (int i = 0; i < CELT_MAX_BANDS; i++) {
234 int band_offset = ff_celt_freq_bands[i] << f->size;
235 int band_size = ff_celt_freq_range[i] << f->size;
236 float *coeffs = &block->coeffs[band_offset];
238 for (int j = 0; j < band_size; j++)
239 ener += coeffs[j]*coeffs[j];
241 block->lin_energy[i] = sqrtf(ener) + FLT_EPSILON;
242 ener = 1.0f/block->lin_energy[i];
244 for (int j = 0; j < band_size; j++)
247 block->energy[i] = log2f(block->lin_energy[i]) - ff_celt_mean_energy[i];
249 /* CELT_ENERGY_SILENCE is what the decoder uses and its not -infinity */
250 block->energy[i] = FFMAX(block->energy[i], CELT_ENERGY_SILENCE);
255 static void celt_enc_tf(CeltFrame *f, OpusRangeCoder *rc)
257 int tf_select = 0, diff = 0, tf_changed = 0, tf_select_needed;
258 int bits = f->transient ? 2 : 4;
260 tf_select_needed = ((f->size && (opus_rc_tell(rc) + bits + 1) <= f->framebits));
262 for (int i = f->start_band; i < f->end_band; i++) {
263 if ((opus_rc_tell(rc) + bits + tf_select_needed) <= f->framebits) {
264 const int tbit = (diff ^ 1) == f->tf_change[i];
265 ff_opus_rc_enc_log(rc, tbit, bits);
269 bits = f->transient ? 4 : 5;
272 if (tf_select_needed && ff_celt_tf_select[f->size][f->transient][0][tf_changed] !=
273 ff_celt_tf_select[f->size][f->transient][1][tf_changed]) {
274 ff_opus_rc_enc_log(rc, f->tf_select, 1);
275 tf_select = f->tf_select;
278 for (int i = f->start_band; i < f->end_band; i++)
279 f->tf_change[i] = ff_celt_tf_select[f->size][f->transient][tf_select][f->tf_change[i]];
282 static void celt_enc_quant_pfilter(OpusRangeCoder *rc, CeltFrame *f)
284 float gain = f->pf_gain;
285 int txval, octave = f->pf_octave, period = f->pf_period, tapset = f->pf_tapset;
287 ff_opus_rc_enc_log(rc, f->pfilter, 1);
292 txval = FFMIN(octave, 6);
293 ff_opus_rc_enc_uint(rc, txval, 6);
296 txval = av_clip(period - (16 << octave) + 1, 0, (1 << (4 + octave)) - 1);
297 ff_opus_rc_put_raw(rc, period, 4 + octave);
298 period = txval + (16 << octave) - 1;
300 txval = FFMIN(((int)(gain / 0.09375f)) - 1, 7);
301 ff_opus_rc_put_raw(rc, txval, 3);
302 gain = 0.09375f * (txval + 1);
304 if ((opus_rc_tell(rc) + 2) <= f->framebits)
305 ff_opus_rc_enc_cdf(rc, tapset, ff_celt_model_tapset);
308 /* Finally create the coeffs */
309 for (int i = 0; i < 2; i++) {
310 CeltBlock *block = &f->block[i];
312 block->pf_period_new = FFMAX(period, CELT_POSTFILTER_MINPERIOD);
313 block->pf_gains_new[0] = gain * ff_celt_postfilter_taps[tapset][0];
314 block->pf_gains_new[1] = gain * ff_celt_postfilter_taps[tapset][1];
315 block->pf_gains_new[2] = gain * ff_celt_postfilter_taps[tapset][2];
319 static void exp_quant_coarse(OpusRangeCoder *rc, CeltFrame *f,
320 float last_energy[][CELT_MAX_BANDS], int intra)
322 float alpha, beta, prev[2] = { 0, 0 };
323 const uint8_t *pmod = ff_celt_coarse_energy_dist[f->size][intra];
325 /* Inter is really just differential coding */
326 if (opus_rc_tell(rc) + 3 <= f->framebits)
327 ff_opus_rc_enc_log(rc, intra, 3);
333 beta = 1.0f - (4915.0f/32768.0f);
335 alpha = ff_celt_alpha_coef[f->size];
336 beta = ff_celt_beta_coef[f->size];
339 for (int i = f->start_band; i < f->end_band; i++) {
340 for (int ch = 0; ch < f->channels; ch++) {
341 CeltBlock *block = &f->block[ch];
342 const int left = f->framebits - opus_rc_tell(rc);
343 const float last = FFMAX(-9.0f, last_energy[ch][i]);
344 float diff = block->energy[i] - prev[ch] - last*alpha;
345 int q_en = lrintf(diff);
347 ff_opus_rc_enc_laplace(rc, &q_en, pmod[i << 1] << 7, pmod[(i << 1) + 1] << 6);
348 } else if (left >= 2) {
349 q_en = av_clip(q_en, -1, 1);
350 ff_opus_rc_enc_cdf(rc, 2*q_en + 3*(q_en < 0), ff_celt_model_energy_small);
351 } else if (left >= 1) {
352 q_en = av_clip(q_en, -1, 0);
353 ff_opus_rc_enc_log(rc, (q_en & 1), 1);
356 block->error_energy[i] = q_en - diff;
357 prev[ch] += beta * q_en;
362 static void celt_quant_coarse(CeltFrame *f, OpusRangeCoder *rc,
363 float last_energy[][CELT_MAX_BANDS])
365 uint32_t inter, intra;
366 OPUS_RC_CHECKPOINT_SPAWN(rc);
368 exp_quant_coarse(rc, f, last_energy, 1);
369 intra = OPUS_RC_CHECKPOINT_BITS(rc);
371 OPUS_RC_CHECKPOINT_ROLLBACK(rc);
373 exp_quant_coarse(rc, f, last_energy, 0);
374 inter = OPUS_RC_CHECKPOINT_BITS(rc);
376 if (inter > intra) { /* Unlikely */
377 OPUS_RC_CHECKPOINT_ROLLBACK(rc);
378 exp_quant_coarse(rc, f, last_energy, 1);
382 static void celt_quant_fine(CeltFrame *f, OpusRangeCoder *rc)
384 for (int i = f->start_band; i < f->end_band; i++) {
385 if (!f->fine_bits[i])
387 for (int ch = 0; ch < f->channels; ch++) {
388 CeltBlock *block = &f->block[ch];
389 int quant, lim = (1 << f->fine_bits[i]);
390 float offset, diff = 0.5f - block->error_energy[i];
391 quant = av_clip(floor(diff*lim), 0, lim - 1);
392 ff_opus_rc_put_raw(rc, quant, f->fine_bits[i]);
393 offset = 0.5f - ((quant + 0.5f) * (1 << (14 - f->fine_bits[i])) / 16384.0f);
394 block->error_energy[i] -= offset;
399 static void celt_quant_final(OpusEncContext *s, OpusRangeCoder *rc, CeltFrame *f)
401 for (int priority = 0; priority < 2; priority++) {
402 for (int i = f->start_band; i < f->end_band && (f->framebits - opus_rc_tell(rc)) >= f->channels; i++) {
403 if (f->fine_priority[i] != priority || f->fine_bits[i] >= CELT_MAX_FINE_BITS)
405 for (int ch = 0; ch < f->channels; ch++) {
406 CeltBlock *block = &f->block[ch];
407 const float err = block->error_energy[i];
408 const float offset = 0.5f * (1 << (14 - f->fine_bits[i] - 1)) / 16384.0f;
409 const int sign = FFABS(err + offset) < FFABS(err - offset);
410 ff_opus_rc_put_raw(rc, sign, 1);
411 block->error_energy[i] -= offset*(1 - 2*sign);
417 static void celt_encode_frame(OpusEncContext *s, OpusRangeCoder *rc,
418 CeltFrame *f, int index)
420 ff_opus_rc_enc_init(rc);
422 ff_opus_psy_celt_frame_init(&s->psyctx, f, index);
424 celt_frame_setup_input(s, f);
427 if (f->framebits >= 16)
428 ff_opus_rc_enc_log(rc, 1, 15); /* Silence (if using explicit singalling) */
429 for (int ch = 0; ch < s->channels; ch++)
430 memset(s->last_quantized_energy[ch], 0.0f, sizeof(float)*CELT_MAX_BANDS);
435 celt_apply_preemph_filter(s, f);
437 ff_opus_rc_enc_log(rc, 0, 15);
438 celt_enc_quant_pfilter(rc, f);
442 celt_frame_mdct(s, f);
444 /* Need to handle transient/non-transient switches at any point during analysis */
445 while (ff_opus_psy_celt_frame_process(&s->psyctx, f, index))
446 celt_frame_mdct(s, f);
448 ff_opus_rc_enc_init(rc);
451 ff_opus_rc_enc_log(rc, 0, 15);
454 if (!f->start_band && opus_rc_tell(rc) + 16 <= f->framebits)
455 celt_enc_quant_pfilter(rc, f);
458 if (f->size && opus_rc_tell(rc) + 3 <= f->framebits)
459 ff_opus_rc_enc_log(rc, f->transient, 3);
462 celt_quant_coarse (f, rc, s->last_quantized_energy);
464 ff_celt_bitalloc (f, rc, 1);
465 celt_quant_fine (f, rc);
466 ff_celt_quant_bands(f, rc);
468 /* Anticollapse bit */
469 if (f->anticollapse_needed)
470 ff_opus_rc_put_raw(rc, f->anticollapse, 1);
472 /* Final per-band energy adjustments from leftover bits */
473 celt_quant_final(s, rc, f);
475 for (int ch = 0; ch < f->channels; ch++) {
476 CeltBlock *block = &f->block[ch];
477 for (int i = 0; i < CELT_MAX_BANDS; i++)
478 s->last_quantized_energy[ch][i] = block->energy[i] + block->error_energy[i];
482 static inline int write_opuslacing(uint8_t *dst, int v)
484 dst[0] = FFMIN(v - FFALIGN(v - 255, 4), v);
485 dst[1] = v - dst[0] >> 2;
486 return 1 + (v >= 252);
489 static void opus_packet_assembler(OpusEncContext *s, AVPacket *avpkt)
491 int offset, fsize_needed;
494 opus_gen_toc(s, avpkt->data, &offset, &fsize_needed);
496 /* Frame sizes if needed */
498 for (int i = 0; i < s->packet.frames - 1; i++) {
499 offset += write_opuslacing(avpkt->data + offset,
500 s->frame[i].framebits >> 3);
505 for (int i = 0; i < s->packet.frames; i++) {
506 ff_opus_rc_enc_end(&s->rc[i], avpkt->data + offset,
507 s->frame[i].framebits >> 3);
508 offset += s->frame[i].framebits >> 3;
511 avpkt->size = offset;
514 /* Used as overlap for the first frame and padding for the last encoded packet */
515 static AVFrame *spawn_empty_frame(OpusEncContext *s)
517 AVFrame *f = av_frame_alloc();
520 f->format = s->avctx->sample_fmt;
521 f->nb_samples = s->avctx->frame_size;
522 f->channel_layout = s->avctx->channel_layout;
523 if (av_frame_get_buffer(f, 4)) {
527 for (int i = 0; i < s->channels; i++) {
528 size_t bps = av_get_bytes_per_sample(f->format);
529 memset(f->extended_data[i], 0, bps*f->nb_samples);
534 static int opus_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
535 const AVFrame *frame, int *got_packet_ptr)
537 OpusEncContext *s = avctx->priv_data;
538 int ret, frame_size, alloc_size = 0;
540 if (frame) { /* Add new frame to queue */
541 if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
543 ff_bufqueue_add(avctx, &s->bufqueue, av_frame_clone(frame));
545 ff_opus_psy_signal_eof(&s->psyctx);
546 if (!s->afq.remaining_samples)
547 return 0; /* We've been flushed and there's nothing left to encode */
550 /* Run the psychoacoustic system */
551 if (ff_opus_psy_process(&s->psyctx, &s->packet))
554 frame_size = OPUS_BLOCK_SIZE(s->packet.framesize);
557 /* This can go negative, that's not a problem, we only pad if positive */
558 int pad_empty = s->packet.frames*(frame_size/s->avctx->frame_size) - s->bufqueue.available + 1;
559 /* Pad with empty 2.5 ms frames to whatever framesize was decided,
560 * this should only happen at the very last flush frame. The frames
561 * allocated here will be freed (because they have no other references)
562 * after they get used by celt_frame_setup_input() */
563 for (int i = 0; i < pad_empty; i++) {
564 AVFrame *empty = spawn_empty_frame(s);
566 return AVERROR(ENOMEM);
567 ff_bufqueue_add(avctx, &s->bufqueue, empty);
571 for (int i = 0; i < s->packet.frames; i++) {
572 celt_encode_frame(s, &s->rc[i], &s->frame[i], i);
573 alloc_size += s->frame[i].framebits >> 3;
576 /* Worst case toc + the frame lengths if needed */
577 alloc_size += 2 + s->packet.frames*2;
579 if ((ret = ff_alloc_packet2(avctx, avpkt, alloc_size, 0)) < 0)
582 /* Assemble packet */
583 opus_packet_assembler(s, avpkt);
585 /* Update the psychoacoustic system */
586 ff_opus_psy_postencode_update(&s->psyctx, s->frame, s->rc);
588 /* Remove samples from queue and skip if needed */
589 ff_af_queue_remove(&s->afq, s->packet.frames*frame_size, &avpkt->pts, &avpkt->duration);
590 if (s->packet.frames*frame_size > avpkt->duration) {
591 uint8_t *side = av_packet_new_side_data(avpkt, AV_PKT_DATA_SKIP_SAMPLES, 10);
593 return AVERROR(ENOMEM);
594 AV_WL32(&side[4], s->packet.frames*frame_size - avpkt->duration + 120);
602 static av_cold int opus_encode_end(AVCodecContext *avctx)
604 OpusEncContext *s = avctx->priv_data;
606 for (int i = 0; i < CELT_BLOCK_NB; i++)
607 ff_mdct15_uninit(&s->mdct[i]);
609 ff_celt_pvq_uninit(&s->pvq);
613 ff_af_queue_close(&s->afq);
614 ff_opus_psy_end(&s->psyctx);
615 ff_bufqueue_discard_all(&s->bufqueue);
616 av_freep(&avctx->extradata);
621 static av_cold int opus_encode_init(AVCodecContext *avctx)
624 OpusEncContext *s = avctx->priv_data;
627 s->channels = avctx->channels;
629 /* Opus allows us to change the framesize on each packet (and each packet may
630 * have multiple frames in it) but we can't change the codec's frame size on
631 * runtime, so fix it to the lowest possible number of samples and use a queue
632 * to accumulate AVFrames until we have enough to encode whatever the encoder
633 * decides is the best */
634 avctx->frame_size = 120;
635 /* Initial padding will change if SILK is ever supported */
636 avctx->initial_padding = 120;
638 if (!avctx->bit_rate) {
639 int coupled = ff_opus_default_coupled_streams[s->channels - 1];
640 avctx->bit_rate = coupled*(96000) + (s->channels - coupled*2)*(48000);
641 } else if (avctx->bit_rate < 6000 || avctx->bit_rate > 255000 * s->channels) {
642 int64_t clipped_rate = av_clip(avctx->bit_rate, 6000, 255000 * s->channels);
643 av_log(avctx, AV_LOG_ERROR, "Unsupported bitrate %"PRId64" kbps, clipping to %"PRId64" kbps\n",
644 avctx->bit_rate/1000, clipped_rate/1000);
645 avctx->bit_rate = clipped_rate;
649 avctx->extradata_size = 19;
650 avctx->extradata = av_malloc(avctx->extradata_size + AV_INPUT_BUFFER_PADDING_SIZE);
651 if (!avctx->extradata)
652 return AVERROR(ENOMEM);
653 opus_write_extradata(avctx);
655 ff_af_queue_init(avctx, &s->afq);
657 if ((ret = ff_celt_pvq_init(&s->pvq, 1)) < 0)
660 if (!(s->dsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT)))
661 return AVERROR(ENOMEM);
663 /* I have no idea why a base scaling factor of 68 works, could be the twiddles */
664 for (int i = 0; i < CELT_BLOCK_NB; i++)
665 if ((ret = ff_mdct15_init(&s->mdct[i], 0, i + 3, 68 << (CELT_BLOCK_NB - 1 - i))))
666 return AVERROR(ENOMEM);
668 /* Zero out previous energy (matters for inter first frame) */
669 for (int ch = 0; ch < s->channels; ch++)
670 memset(s->last_quantized_energy[ch], 0.0f, sizeof(float)*CELT_MAX_BANDS);
672 /* Allocate an empty frame to use as overlap for the first frame of audio */
673 ff_bufqueue_add(avctx, &s->bufqueue, spawn_empty_frame(s));
674 if (!ff_bufqueue_peek(&s->bufqueue, 0))
675 return AVERROR(ENOMEM);
677 if ((ret = ff_opus_psy_init(&s->psyctx, s->avctx, &s->bufqueue, &s->options)))
680 /* Frame structs and range coder buffers */
681 max_frames = ceilf(FFMIN(s->options.max_delay_ms, 120.0f)/2.5f);
682 s->frame = av_malloc(max_frames*sizeof(CeltFrame));
684 return AVERROR(ENOMEM);
685 s->rc = av_malloc(max_frames*sizeof(OpusRangeCoder));
687 return AVERROR(ENOMEM);
689 for (int i = 0; i < max_frames; i++) {
690 s->frame[i].dsp = s->dsp;
691 s->frame[i].avctx = s->avctx;
692 s->frame[i].seed = 0;
693 s->frame[i].pvq = s->pvq;
694 s->frame[i].apply_phase_inv = 1;
695 s->frame[i].block[0].emph_coeff = s->frame[i].block[1].emph_coeff = 0.0f;
701 #define OPUSENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
702 static const AVOption opusenc_options[] = {
703 { "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" },
707 static const AVClass opusenc_class = {
708 .class_name = "Opus encoder",
709 .item_name = av_default_item_name,
710 .option = opusenc_options,
711 .version = LIBAVUTIL_VERSION_INT,
714 static const AVCodecDefault opusenc_defaults[] = {
716 { "compression_level", "10" },
720 AVCodec ff_opus_encoder = {
722 .long_name = NULL_IF_CONFIG_SMALL("Opus"),
723 .type = AVMEDIA_TYPE_AUDIO,
724 .id = AV_CODEC_ID_OPUS,
725 .defaults = opusenc_defaults,
726 .priv_class = &opusenc_class,
727 .priv_data_size = sizeof(OpusEncContext),
728 .init = opus_encode_init,
729 .encode2 = opus_encode_frame,
730 .close = opus_encode_end,
731 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE | FF_CODEC_CAP_INIT_CLEANUP,
732 .capabilities = AV_CODEC_CAP_EXPERIMENTAL | AV_CODEC_CAP_SMALL_LAST_FRAME | AV_CODEC_CAP_DELAY,
733 .supported_samplerates = (const int []){ 48000, 0 },
734 .channel_layouts = (const uint64_t []){ AV_CH_LAYOUT_MONO,
735 AV_CH_LAYOUT_STEREO, 0 },
736 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
737 AV_SAMPLE_FMT_NONE },