3 * Copyright (C) 2008 Konstantin Shishkov
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
27 /***********************************
29 * add sane pulse detection
30 ***********************************/
32 #include "libavutil/float_dsp.h"
33 #include "libavutil/opt.h"
37 #include "mpeg4audio.h"
44 #include "aacenctab.h"
45 #include "aacenc_utils.h"
49 struct AACProfileOptions {
51 struct AACEncOptions opts;
55 * List of currently supported profiles, anything not listed isn't supported.
57 static const struct AACProfileOptions aacenc_profiles[] = {
59 { /* Main profile, all advanced encoding abilities enabled */
64 .intensity_stereo = 1,
68 { /* Default profile, these are the settings that get set by default */
72 .pred = OPT_NEEDS_MAIN,
73 .intensity_stereo = 1,
76 {FF_PROFILE_MPEG2_AAC_LOW,
77 { /* Strict MPEG 2 Part 7 compliance profile */
82 .intensity_stereo = 1,
88 * Make AAC audio config object.
89 * @see 1.6.2.1 "Syntax - AudioSpecificConfig"
91 static void put_audio_specific_config(AVCodecContext *avctx)
94 AACEncContext *s = avctx->priv_data;
95 int channels = s->channels - (s->channels == 8 ? 1 : 0);
97 init_put_bits(&pb, avctx->extradata, avctx->extradata_size);
98 put_bits(&pb, 5, s->profile+1); //profile
99 put_bits(&pb, 4, s->samplerate_index); //sample rate index
100 put_bits(&pb, 4, channels);
102 put_bits(&pb, 1, 0); //frame length - 1024 samples
103 put_bits(&pb, 1, 0); //does not depend on core coder
104 put_bits(&pb, 1, 0); //is not extension
106 //Explicitly Mark SBR absent
107 put_bits(&pb, 11, 0x2b7); //sync extension
108 put_bits(&pb, 5, AOT_SBR);
113 void ff_quantize_band_cost_cache_init(struct AACEncContext *s)
116 for (sf = 0; sf < 256; sf++) {
117 for (g = 0; g < 128; g++) {
118 s->quantize_band_cost_cache[sf][g].bits = -1;
123 #define WINDOW_FUNC(type) \
124 static void apply_ ##type ##_window(AVFloatDSPContext *fdsp, \
125 SingleChannelElement *sce, \
128 WINDOW_FUNC(only_long)
130 const float *lwindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
131 const float *pwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
132 float *out = sce->ret_buf;
134 fdsp->vector_fmul (out, audio, lwindow, 1024);
135 fdsp->vector_fmul_reverse(out + 1024, audio + 1024, pwindow, 1024);
138 WINDOW_FUNC(long_start)
140 const float *lwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
141 const float *swindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
142 float *out = sce->ret_buf;
144 fdsp->vector_fmul(out, audio, lwindow, 1024);
145 memcpy(out + 1024, audio + 1024, sizeof(out[0]) * 448);
146 fdsp->vector_fmul_reverse(out + 1024 + 448, audio + 1024 + 448, swindow, 128);
147 memset(out + 1024 + 576, 0, sizeof(out[0]) * 448);
150 WINDOW_FUNC(long_stop)
152 const float *lwindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
153 const float *swindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;
154 float *out = sce->ret_buf;
156 memset(out, 0, sizeof(out[0]) * 448);
157 fdsp->vector_fmul(out + 448, audio + 448, swindow, 128);
158 memcpy(out + 576, audio + 576, sizeof(out[0]) * 448);
159 fdsp->vector_fmul_reverse(out + 1024, audio + 1024, lwindow, 1024);
162 WINDOW_FUNC(eight_short)
164 const float *swindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
165 const float *pwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;
166 const float *in = audio + 448;
167 float *out = sce->ret_buf;
170 for (w = 0; w < 8; w++) {
171 fdsp->vector_fmul (out, in, w ? pwindow : swindow, 128);
174 fdsp->vector_fmul_reverse(out, in, swindow, 128);
179 static void (*const apply_window[4])(AVFloatDSPContext *fdsp,
180 SingleChannelElement *sce,
181 const float *audio) = {
182 [ONLY_LONG_SEQUENCE] = apply_only_long_window,
183 [LONG_START_SEQUENCE] = apply_long_start_window,
184 [EIGHT_SHORT_SEQUENCE] = apply_eight_short_window,
185 [LONG_STOP_SEQUENCE] = apply_long_stop_window
188 static void apply_window_and_mdct(AACEncContext *s, SingleChannelElement *sce,
192 float *output = sce->ret_buf;
194 apply_window[sce->ics.window_sequence[0]](s->fdsp, sce, audio);
196 if (sce->ics.window_sequence[0] != EIGHT_SHORT_SEQUENCE)
197 s->mdct1024.mdct_calc(&s->mdct1024, sce->coeffs, output);
199 for (i = 0; i < 1024; i += 128)
200 s->mdct128.mdct_calc(&s->mdct128, &sce->coeffs[i], output + i*2);
201 memcpy(audio, audio + 1024, sizeof(audio[0]) * 1024);
202 memcpy(sce->pcoeffs, sce->coeffs, sizeof(sce->pcoeffs));
206 * Encode ics_info element.
207 * @see Table 4.6 (syntax of ics_info)
209 static void put_ics_info(AACEncContext *s, IndividualChannelStream *info)
213 put_bits(&s->pb, 1, 0); // ics_reserved bit
214 put_bits(&s->pb, 2, info->window_sequence[0]);
215 put_bits(&s->pb, 1, info->use_kb_window[0]);
216 if (info->window_sequence[0] != EIGHT_SHORT_SEQUENCE) {
217 put_bits(&s->pb, 6, info->max_sfb);
218 put_bits(&s->pb, 1, !!info->predictor_present);
220 put_bits(&s->pb, 4, info->max_sfb);
221 for (w = 1; w < 8; w++)
222 put_bits(&s->pb, 1, !info->group_len[w]);
228 * @see 4.6.8.1 "Joint Coding - M/S Stereo"
230 static void encode_ms_info(PutBitContext *pb, ChannelElement *cpe)
234 put_bits(pb, 2, cpe->ms_mode);
235 if (cpe->ms_mode == 1)
236 for (w = 0; w < cpe->ch[0].ics.num_windows; w += cpe->ch[0].ics.group_len[w])
237 for (i = 0; i < cpe->ch[0].ics.max_sfb; i++)
238 put_bits(pb, 1, cpe->ms_mask[w*16 + i]);
242 * Produce integer coefficients from scalefactors provided by the model.
244 static void adjust_frame_information(ChannelElement *cpe, int chans)
249 for (ch = 0; ch < chans; ch++) {
250 IndividualChannelStream *ics = &cpe->ch[ch].ics;
252 cpe->ch[ch].pulse.num_pulse = 0;
253 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
254 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
255 for (cmaxsfb = ics->num_swb; cmaxsfb > 0 && cpe->ch[ch].zeroes[w*16+cmaxsfb-1]; cmaxsfb--)
257 maxsfb = FFMAX(maxsfb, cmaxsfb);
260 ics->max_sfb = maxsfb;
262 //adjust zero bands for window groups
263 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
264 for (g = 0; g < ics->max_sfb; g++) {
266 for (w2 = w; w2 < w + ics->group_len[w]; w2++) {
267 if (!cpe->ch[ch].zeroes[w2*16 + g]) {
272 cpe->ch[ch].zeroes[w*16 + g] = i;
277 if (chans > 1 && cpe->common_window) {
278 IndividualChannelStream *ics0 = &cpe->ch[0].ics;
279 IndividualChannelStream *ics1 = &cpe->ch[1].ics;
281 ics0->max_sfb = FFMAX(ics0->max_sfb, ics1->max_sfb);
282 ics1->max_sfb = ics0->max_sfb;
283 for (w = 0; w < ics0->num_windows*16; w += 16)
284 for (i = 0; i < ics0->max_sfb; i++)
285 if (cpe->ms_mask[w+i])
287 if (msc == 0 || ics0->max_sfb == 0)
290 cpe->ms_mode = msc < ics0->max_sfb * ics0->num_windows ? 1 : 2;
294 static void apply_intensity_stereo(ChannelElement *cpe)
297 IndividualChannelStream *ics = &cpe->ch[0].ics;
298 if (!cpe->common_window)
300 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
301 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
302 int start = (w+w2) * 128;
303 for (g = 0; g < ics->num_swb; g++) {
304 int p = -1 + 2 * (cpe->ch[1].band_type[w*16+g] - 14);
305 float scale = cpe->ch[0].is_ener[w*16+g];
306 if (!cpe->is_mask[w*16 + g]) {
307 start += ics->swb_sizes[g];
310 if (cpe->ms_mask[w*16 + g])
312 for (i = 0; i < ics->swb_sizes[g]; i++) {
313 float sum = (cpe->ch[0].coeffs[start+i] + p*cpe->ch[1].coeffs[start+i])*scale;
314 cpe->ch[0].coeffs[start+i] = sum;
315 cpe->ch[1].coeffs[start+i] = 0.0f;
317 start += ics->swb_sizes[g];
323 static void apply_mid_side_stereo(ChannelElement *cpe)
326 IndividualChannelStream *ics = &cpe->ch[0].ics;
327 if (!cpe->common_window)
329 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
330 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
331 int start = (w+w2) * 128;
332 for (g = 0; g < ics->num_swb; g++) {
333 if (!cpe->ms_mask[w*16 + g] && !cpe->is_mask[w*16 + g]) {
334 start += ics->swb_sizes[g];
337 for (i = 0; i < ics->swb_sizes[g]; i++) {
338 float L = (cpe->ch[0].coeffs[start+i] + cpe->ch[1].coeffs[start+i]) * 0.5f;
339 float R = L - cpe->ch[1].coeffs[start+i];
340 cpe->ch[0].coeffs[start+i] = L;
341 cpe->ch[1].coeffs[start+i] = R;
343 start += ics->swb_sizes[g];
350 * Encode scalefactor band coding type.
352 static void encode_band_info(AACEncContext *s, SingleChannelElement *sce)
356 if (s->coder->set_special_band_scalefactors)
357 s->coder->set_special_band_scalefactors(s, sce);
359 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
360 s->coder->encode_window_bands_info(s, sce, w, sce->ics.group_len[w], s->lambda);
364 * Encode scalefactors.
366 static void encode_scale_factors(AVCodecContext *avctx, AACEncContext *s,
367 SingleChannelElement *sce)
369 int diff, off_sf = sce->sf_idx[0], off_pns = sce->sf_idx[0] - NOISE_OFFSET;
370 int off_is = 0, noise_flag = 1;
373 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
374 for (i = 0; i < sce->ics.max_sfb; i++) {
375 if (!sce->zeroes[w*16 + i]) {
376 if (sce->band_type[w*16 + i] == NOISE_BT) {
377 diff = sce->sf_idx[w*16 + i] - off_pns;
378 off_pns = sce->sf_idx[w*16 + i];
379 if (noise_flag-- > 0) {
380 put_bits(&s->pb, NOISE_PRE_BITS, diff + NOISE_PRE);
383 } else if (sce->band_type[w*16 + i] == INTENSITY_BT ||
384 sce->band_type[w*16 + i] == INTENSITY_BT2) {
385 diff = sce->sf_idx[w*16 + i] - off_is;
386 off_is = sce->sf_idx[w*16 + i];
388 diff = sce->sf_idx[w*16 + i] - off_sf;
389 off_sf = sce->sf_idx[w*16 + i];
391 diff += SCALE_DIFF_ZERO;
392 av_assert0(diff >= 0 && diff <= 120);
393 put_bits(&s->pb, ff_aac_scalefactor_bits[diff], ff_aac_scalefactor_code[diff]);
402 static void encode_pulses(AACEncContext *s, Pulse *pulse)
406 put_bits(&s->pb, 1, !!pulse->num_pulse);
407 if (!pulse->num_pulse)
410 put_bits(&s->pb, 2, pulse->num_pulse - 1);
411 put_bits(&s->pb, 6, pulse->start);
412 for (i = 0; i < pulse->num_pulse; i++) {
413 put_bits(&s->pb, 5, pulse->pos[i]);
414 put_bits(&s->pb, 4, pulse->amp[i]);
419 * Encode spectral coefficients processed by psychoacoustic model.
421 static void encode_spectral_coeffs(AACEncContext *s, SingleChannelElement *sce)
425 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
427 for (i = 0; i < sce->ics.max_sfb; i++) {
428 if (sce->zeroes[w*16 + i]) {
429 start += sce->ics.swb_sizes[i];
432 for (w2 = w; w2 < w + sce->ics.group_len[w]; w2++) {
433 s->coder->quantize_and_encode_band(s, &s->pb,
434 &sce->coeffs[start + w2*128],
435 NULL, sce->ics.swb_sizes[i],
436 sce->sf_idx[w*16 + i],
437 sce->band_type[w*16 + i],
439 sce->ics.window_clipping[w]);
441 start += sce->ics.swb_sizes[i];
447 * Downscale spectral coefficients for near-clipping windows to avoid artifacts
449 static void avoid_clipping(AACEncContext *s, SingleChannelElement *sce)
453 if (sce->ics.clip_avoidance_factor < 1.0f) {
454 for (w = 0; w < sce->ics.num_windows; w++) {
456 for (i = 0; i < sce->ics.max_sfb; i++) {
457 float *swb_coeffs = &sce->coeffs[start + w*128];
458 for (j = 0; j < sce->ics.swb_sizes[i]; j++)
459 swb_coeffs[j] *= sce->ics.clip_avoidance_factor;
460 start += sce->ics.swb_sizes[i];
467 * Encode one channel of audio data.
469 static int encode_individual_channel(AVCodecContext *avctx, AACEncContext *s,
470 SingleChannelElement *sce,
473 put_bits(&s->pb, 8, sce->sf_idx[0]);
474 if (!common_window) {
475 put_ics_info(s, &sce->ics);
476 if (s->coder->encode_main_pred)
477 s->coder->encode_main_pred(s, sce);
479 encode_band_info(s, sce);
480 encode_scale_factors(avctx, s, sce);
481 encode_pulses(s, &sce->pulse);
482 put_bits(&s->pb, 1, !!sce->tns.present);
483 if (s->coder->encode_tns_info)
484 s->coder->encode_tns_info(s, sce);
485 put_bits(&s->pb, 1, 0); //ssr
486 encode_spectral_coeffs(s, sce);
491 * Write some auxiliary information about the created AAC file.
493 static void put_bitstream_info(AACEncContext *s, const char *name)
495 int i, namelen, padbits;
497 namelen = strlen(name) + 2;
498 put_bits(&s->pb, 3, TYPE_FIL);
499 put_bits(&s->pb, 4, FFMIN(namelen, 15));
501 put_bits(&s->pb, 8, namelen - 14);
502 put_bits(&s->pb, 4, 0); //extension type - filler
503 padbits = -put_bits_count(&s->pb) & 7;
504 avpriv_align_put_bits(&s->pb);
505 for (i = 0; i < namelen - 2; i++)
506 put_bits(&s->pb, 8, name[i]);
507 put_bits(&s->pb, 12 - padbits, 0);
511 * Copy input samples.
512 * Channels are reordered from libavcodec's default order to AAC order.
514 static void copy_input_samples(AACEncContext *s, const AVFrame *frame)
517 int end = 2048 + (frame ? frame->nb_samples : 0);
518 const uint8_t *channel_map = aac_chan_maps[s->channels - 1];
520 /* copy and remap input samples */
521 for (ch = 0; ch < s->channels; ch++) {
522 /* copy last 1024 samples of previous frame to the start of the current frame */
523 memcpy(&s->planar_samples[ch][1024], &s->planar_samples[ch][2048], 1024 * sizeof(s->planar_samples[0][0]));
525 /* copy new samples and zero any remaining samples */
527 memcpy(&s->planar_samples[ch][2048],
528 frame->extended_data[channel_map[ch]],
529 frame->nb_samples * sizeof(s->planar_samples[0][0]));
531 memset(&s->planar_samples[ch][end], 0,
532 (3072 - end) * sizeof(s->planar_samples[0][0]));
536 static int aac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
537 const AVFrame *frame, int *got_packet_ptr)
539 AACEncContext *s = avctx->priv_data;
540 float **samples = s->planar_samples, *samples2, *la, *overlap;
542 SingleChannelElement *sce;
543 int i, its, ch, w, chans, tag, start_ch, ret, frame_bits;
544 int target_bits, rate_bits, too_many_bits, too_few_bits;
545 int ms_mode = 0, is_mode = 0, tns_mode = 0, pred_mode = 0;
546 int chan_el_counter[4];
547 FFPsyWindowInfo windows[AAC_MAX_CHANNELS];
549 if (s->last_frame == 2)
552 /* add current frame to queue */
554 if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
558 copy_input_samples(s, frame);
560 ff_psy_preprocess(s->psypp, s->planar_samples, s->channels);
562 if (!avctx->frame_number)
566 for (i = 0; i < s->chan_map[0]; i++) {
567 FFPsyWindowInfo* wi = windows + start_ch;
568 tag = s->chan_map[i+1];
569 chans = tag == TYPE_CPE ? 2 : 1;
571 for (ch = 0; ch < chans; ch++) {
572 IndividualChannelStream *ics = &cpe->ch[ch].ics;
573 int cur_channel = start_ch + ch;
574 float clip_avoidance_factor;
575 overlap = &samples[cur_channel][0];
576 samples2 = overlap + 1024;
577 la = samples2 + (448+64);
580 if (tag == TYPE_LFE) {
581 wi[ch].window_type[0] = ONLY_LONG_SEQUENCE;
582 wi[ch].window_shape = 0;
583 wi[ch].num_windows = 1;
584 wi[ch].grouping[0] = 1;
586 /* Only the lowest 12 coefficients are used in a LFE channel.
587 * The expression below results in only the bottom 8 coefficients
588 * being used for 11.025kHz to 16kHz sample rates.
590 ics->num_swb = s->samplerate_index >= 8 ? 1 : 3;
592 wi[ch] = s->psy.model->window(&s->psy, samples2, la, cur_channel,
593 ics->window_sequence[0]);
595 ics->window_sequence[1] = ics->window_sequence[0];
596 ics->window_sequence[0] = wi[ch].window_type[0];
597 ics->use_kb_window[1] = ics->use_kb_window[0];
598 ics->use_kb_window[0] = wi[ch].window_shape;
599 ics->num_windows = wi[ch].num_windows;
600 ics->swb_sizes = s->psy.bands [ics->num_windows == 8];
601 ics->num_swb = tag == TYPE_LFE ? ics->num_swb : s->psy.num_bands[ics->num_windows == 8];
602 ics->swb_offset = wi[ch].window_type[0] == EIGHT_SHORT_SEQUENCE ?
603 ff_swb_offset_128 [s->samplerate_index]:
604 ff_swb_offset_1024[s->samplerate_index];
605 ics->tns_max_bands = wi[ch].window_type[0] == EIGHT_SHORT_SEQUENCE ?
606 ff_tns_max_bands_128 [s->samplerate_index]:
607 ff_tns_max_bands_1024[s->samplerate_index];
608 clip_avoidance_factor = 0.0f;
609 for (w = 0; w < ics->num_windows; w++)
610 ics->group_len[w] = wi[ch].grouping[w];
611 for (w = 0; w < ics->num_windows; w++) {
612 if (wi[ch].clipping[w] > CLIP_AVOIDANCE_FACTOR) {
613 ics->window_clipping[w] = 1;
614 clip_avoidance_factor = FFMAX(clip_avoidance_factor, wi[ch].clipping[w]);
616 ics->window_clipping[w] = 0;
619 if (clip_avoidance_factor > CLIP_AVOIDANCE_FACTOR) {
620 ics->clip_avoidance_factor = CLIP_AVOIDANCE_FACTOR / clip_avoidance_factor;
622 ics->clip_avoidance_factor = 1.0f;
625 apply_window_and_mdct(s, &cpe->ch[ch], overlap);
626 if (isnan(cpe->ch->coeffs[0])) {
627 av_log(avctx, AV_LOG_ERROR, "Input contains NaN\n");
628 return AVERROR(EINVAL);
630 avoid_clipping(s, &cpe->ch[ch]);
634 if ((ret = ff_alloc_packet2(avctx, avpkt, 8192 * s->channels, 0)) < 0)
636 frame_bits = its = 0;
638 init_put_bits(&s->pb, avpkt->data, avpkt->size);
640 if ((avctx->frame_number & 0xFF)==1 && !(avctx->flags & AV_CODEC_FLAG_BITEXACT))
641 put_bitstream_info(s, LIBAVCODEC_IDENT);
644 memset(chan_el_counter, 0, sizeof(chan_el_counter));
645 for (i = 0; i < s->chan_map[0]; i++) {
646 FFPsyWindowInfo* wi = windows + start_ch;
647 const float *coeffs[2];
648 tag = s->chan_map[i+1];
649 chans = tag == TYPE_CPE ? 2 : 1;
651 cpe->common_window = 0;
652 memset(cpe->is_mask, 0, sizeof(cpe->is_mask));
653 memset(cpe->ms_mask, 0, sizeof(cpe->ms_mask));
654 put_bits(&s->pb, 3, tag);
655 put_bits(&s->pb, 4, chan_el_counter[tag]++);
656 for (ch = 0; ch < chans; ch++) {
658 coeffs[ch] = sce->coeffs;
659 sce->ics.predictor_present = 0;
660 memset(&sce->ics.prediction_used, 0, sizeof(sce->ics.prediction_used));
661 memset(&sce->tns, 0, sizeof(TemporalNoiseShaping));
662 for (w = 0; w < 128; w++)
663 if (sce->band_type[w] > RESERVED_BT)
664 sce->band_type[w] = 0;
666 s->psy.bitres.alloc = -1;
667 s->psy.bitres.bits = avctx->frame_bits / s->channels;
668 s->psy.model->analyze(&s->psy, start_ch, coeffs, wi);
669 if (s->psy.bitres.alloc > 0) {
670 /* Lambda unused here on purpose, we need to take psy's unscaled allocation */
671 target_bits += s->psy.bitres.alloc
672 * (s->lambda / (avctx->global_quality ? avctx->global_quality : 120));
673 s->psy.bitres.alloc /= chans;
676 for (ch = 0; ch < chans; ch++) {
677 s->cur_channel = start_ch + ch;
678 if (s->options.pns && s->coder->mark_pns)
679 s->coder->mark_pns(s, avctx, &cpe->ch[ch]);
680 s->coder->search_for_quantizers(avctx, s, &cpe->ch[ch], s->lambda);
683 && wi[0].window_type[0] == wi[1].window_type[0]
684 && wi[0].window_shape == wi[1].window_shape) {
686 cpe->common_window = 1;
687 for (w = 0; w < wi[0].num_windows; w++) {
688 if (wi[0].grouping[w] != wi[1].grouping[w]) {
689 cpe->common_window = 0;
694 for (ch = 0; ch < chans; ch++) { /* TNS and PNS */
696 s->cur_channel = start_ch + ch;
697 if (s->options.pns && s->coder->search_for_pns)
698 s->coder->search_for_pns(s, avctx, sce);
699 if (s->options.tns && s->coder->search_for_tns)
700 s->coder->search_for_tns(s, sce);
701 if (s->options.tns && s->coder->apply_tns_filt)
702 s->coder->apply_tns_filt(s, sce);
703 if (sce->tns.present)
706 s->cur_channel = start_ch;
707 if (s->options.intensity_stereo) { /* Intensity Stereo */
708 if (s->coder->search_for_is)
709 s->coder->search_for_is(s, avctx, cpe);
710 if (cpe->is_mode) is_mode = 1;
711 apply_intensity_stereo(cpe);
713 if (s->options.pred) { /* Prediction */
714 for (ch = 0; ch < chans; ch++) {
716 s->cur_channel = start_ch + ch;
717 if (s->options.pred && s->coder->search_for_pred)
718 s->coder->search_for_pred(s, sce);
719 if (cpe->ch[ch].ics.predictor_present) pred_mode = 1;
721 if (s->coder->adjust_common_pred)
722 s->coder->adjust_common_pred(s, cpe);
723 for (ch = 0; ch < chans; ch++) {
725 s->cur_channel = start_ch + ch;
726 if (s->options.pred && s->coder->apply_main_pred)
727 s->coder->apply_main_pred(s, sce);
729 s->cur_channel = start_ch;
731 if (s->options.mid_side) { /* Mid/Side stereo */
732 if (s->options.mid_side == -1 && s->coder->search_for_ms)
733 s->coder->search_for_ms(s, cpe);
734 else if (cpe->common_window)
735 memset(cpe->ms_mask, 1, sizeof(cpe->ms_mask));
736 apply_mid_side_stereo(cpe);
738 adjust_frame_information(cpe, chans);
740 put_bits(&s->pb, 1, cpe->common_window);
741 if (cpe->common_window) {
742 put_ics_info(s, &cpe->ch[0].ics);
743 if (s->coder->encode_main_pred)
744 s->coder->encode_main_pred(s, &cpe->ch[0]);
745 encode_ms_info(&s->pb, cpe);
746 if (cpe->ms_mode) ms_mode = 1;
749 for (ch = 0; ch < chans; ch++) {
750 s->cur_channel = start_ch + ch;
751 encode_individual_channel(avctx, s, &cpe->ch[ch], cpe->common_window);
756 if (avctx->flags & CODEC_FLAG_QSCALE) {
757 /* When using a constant Q-scale, don't mess with lambda */
761 /* rate control stuff
762 * allow between the nominal bitrate, and what psy's bit reservoir says to target
763 * but drift towards the nominal bitrate always
765 frame_bits = put_bits_count(&s->pb);
766 rate_bits = avctx->bit_rate * 1024 / avctx->sample_rate;
767 rate_bits = FFMIN(rate_bits, 6144 * s->channels - 3);
768 too_many_bits = FFMAX(target_bits, rate_bits);
769 too_many_bits = FFMIN(too_many_bits, 6144 * s->channels - 3);
770 too_few_bits = FFMIN(FFMAX(rate_bits - rate_bits/4, target_bits), too_many_bits);
772 /* When using ABR, be strict (but only for increasing) */
773 too_few_bits = too_few_bits - too_few_bits/8;
774 too_many_bits = too_many_bits + too_many_bits/2;
776 if ( its == 0 /* for steady-state Q-scale tracking */
777 || (its < 5 && (frame_bits < too_few_bits || frame_bits > too_many_bits))
778 || frame_bits >= 6144 * s->channels - 3 )
780 float ratio = ((float)rate_bits) / frame_bits;
782 if (frame_bits >= too_few_bits && frame_bits <= too_many_bits) {
784 * This path is for steady-state Q-scale tracking
785 * When frame bits fall within the stable range, we still need to adjust
786 * lambda to maintain it like so in a stable fashion (large jumps in lambda
787 * create artifacts and should be avoided), but slowly
789 ratio = sqrtf(sqrtf(ratio));
790 ratio = av_clipf(ratio, 0.9f, 1.1f);
792 /* Not so fast though */
793 ratio = sqrtf(ratio);
795 s->lambda = FFMIN(s->lambda * ratio, 65536.f);
797 /* Keep iterating if we must reduce and lambda is in the sky */
798 if ((s->lambda < 300.f || ratio > 0.9f) && (s->lambda > 10.f || ratio < 1.1f)) {
801 if (is_mode || ms_mode || tns_mode || pred_mode) {
802 for (i = 0; i < s->chan_map[0]; i++) {
803 // Must restore coeffs
804 chans = tag == TYPE_CPE ? 2 : 1;
806 for (ch = 0; ch < chans; ch++)
807 memcpy(cpe->ch[ch].coeffs, cpe->ch[ch].pcoeffs, sizeof(cpe->ch[ch].coeffs));
817 put_bits(&s->pb, 3, TYPE_END);
818 flush_put_bits(&s->pb);
819 avctx->frame_bits = put_bits_count(&s->pb);
820 s->lambda_sum += s->lambda;
826 ff_af_queue_remove(&s->afq, avctx->frame_size, &avpkt->pts,
829 avpkt->size = put_bits_count(&s->pb) >> 3;
834 static av_cold int aac_encode_end(AVCodecContext *avctx)
836 AACEncContext *s = avctx->priv_data;
838 av_log(avctx, AV_LOG_INFO, "Qavg: %.3f\n", s->lambda_sum / s->lambda_count);
840 ff_mdct_end(&s->mdct1024);
841 ff_mdct_end(&s->mdct128);
845 ff_psy_preprocess_end(s->psypp);
846 av_freep(&s->buffer.samples);
849 ff_af_queue_close(&s->afq);
853 static av_cold int dsp_init(AVCodecContext *avctx, AACEncContext *s)
857 s->fdsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);
859 return AVERROR(ENOMEM);
862 ff_kbd_window_init(ff_aac_kbd_long_1024, 4.0, 1024);
863 ff_kbd_window_init(ff_aac_kbd_short_128, 6.0, 128);
864 ff_init_ff_sine_windows(10);
865 ff_init_ff_sine_windows(7);
867 if ((ret = ff_mdct_init(&s->mdct1024, 11, 0, 32768.0)) < 0)
869 if ((ret = ff_mdct_init(&s->mdct128, 8, 0, 32768.0)) < 0)
875 static av_cold int alloc_buffers(AVCodecContext *avctx, AACEncContext *s)
878 FF_ALLOCZ_ARRAY_OR_GOTO(avctx, s->buffer.samples, s->channels, 3 * 1024 * sizeof(s->buffer.samples[0]), alloc_fail);
879 FF_ALLOCZ_ARRAY_OR_GOTO(avctx, s->cpe, s->chan_map[0], sizeof(ChannelElement), alloc_fail);
880 FF_ALLOCZ_OR_GOTO(avctx, avctx->extradata, 5 + AV_INPUT_BUFFER_PADDING_SIZE, alloc_fail);
882 for(ch = 0; ch < s->channels; ch++)
883 s->planar_samples[ch] = s->buffer.samples + 3 * 1024 * ch;
887 return AVERROR(ENOMEM);
890 static av_cold int aac_encode_init(AVCodecContext *avctx)
892 AACEncContext *s = avctx->priv_data;
893 const AACEncOptions *p_opt = NULL;
895 const uint8_t *sizes[2];
896 uint8_t grouping[AAC_MAX_CHANNELS];
899 s->channels = avctx->channels;
900 s->chan_map = aac_chan_configs[s->channels-1];
901 s->random_state = 0x1f2e3d4c;
902 s->lambda = avctx->global_quality > 0 ? avctx->global_quality : 120;
903 avctx->extradata_size = 5;
904 avctx->frame_size = 1024;
905 avctx->initial_padding = 1024;
906 avctx->bit_rate = (int)FFMIN(
907 6144 * s->channels / 1024.0 * avctx->sample_rate,
909 avctx->profile = avctx->profile == FF_PROFILE_UNKNOWN ? FF_PROFILE_AAC_LOW :
912 for (i = 0; i < 16; i++)
913 if (avctx->sample_rate == avpriv_mpeg4audio_sample_rates[i])
915 s->samplerate_index = i;
917 ERROR_IF(s->samplerate_index == 16 ||
918 s->samplerate_index >= ff_aac_swb_size_1024_len ||
919 s->samplerate_index >= ff_aac_swb_size_128_len,
920 "Unsupported sample rate %d\n", avctx->sample_rate);
921 ERROR_IF(s->channels > AAC_MAX_CHANNELS || s->channels == 7,
922 "Unsupported number of channels: %d\n", s->channels);
923 WARN_IF(1024.0 * avctx->bit_rate / avctx->sample_rate > 6144 * s->channels,
924 "Too many bits per frame requested, clamping to max\n");
926 for (i = 0; i < FF_ARRAY_ELEMS(aacenc_profiles); i++) {
927 if (avctx->profile == aacenc_profiles[i].profile) {
928 p_opt = &aacenc_profiles[i].opts;
932 ERROR_IF(!p_opt, "Unsupported encoding profile: %d\n", avctx->profile);
933 AAC_OPT_SET(&s->options, p_opt, 1, coder);
934 AAC_OPT_SET(&s->options, p_opt, 0, pns);
935 AAC_OPT_SET(&s->options, p_opt, 0, tns);
936 AAC_OPT_SET(&s->options, p_opt, 0, pred);
937 AAC_OPT_SET(&s->options, p_opt, 1, mid_side);
938 AAC_OPT_SET(&s->options, p_opt, 0, intensity_stereo);
939 if (avctx->profile == FF_PROFILE_MPEG2_AAC_LOW)
940 s->profile = FF_PROFILE_AAC_LOW;
942 s->profile = avctx->profile;
943 s->coder = &ff_aac_coders[s->options.coder];
945 if (s->options.coder != AAC_CODER_TWOLOOP) {
946 s->options.intensity_stereo = 0;
950 if ((ret = dsp_init(avctx, s)) < 0)
953 if ((ret = alloc_buffers(avctx, s)) < 0)
956 put_audio_specific_config(avctx);
958 sizes[0] = ff_aac_swb_size_1024[s->samplerate_index];
959 sizes[1] = ff_aac_swb_size_128[s->samplerate_index];
960 lengths[0] = ff_aac_num_swb_1024[s->samplerate_index];
961 lengths[1] = ff_aac_num_swb_128[s->samplerate_index];
962 for (i = 0; i < s->chan_map[0]; i++)
963 grouping[i] = s->chan_map[i + 1] == TYPE_CPE;
964 if ((ret = ff_psy_init(&s->psy, avctx, 2, sizes, lengths,
965 s->chan_map[0], grouping)) < 0)
967 s->psypp = ff_psy_preprocess_init(avctx);
968 ff_lpc_init(&s->lpc, 2*avctx->frame_size, TNS_MAX_ORDER, FF_LPC_TYPE_LEVINSON);
971 ff_aac_coder_init_mips(s);
975 ff_af_queue_init(avctx, &s->afq);
979 aac_encode_end(avctx);
983 #define AACENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
984 static const AVOption aacenc_options[] = {
985 {"aac_coder", "Coding algorithm", offsetof(AACEncContext, options.coder), AV_OPT_TYPE_INT, {.i64 = AAC_CODER_TWOLOOP}, -1, AAC_CODER_NB-1, AACENC_FLAGS, "coder"},
986 {"faac", "FAAC-inspired method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_FAAC}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
987 {"anmr", "ANMR method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_ANMR}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
988 {"twoloop", "Two loop searching method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_TWOLOOP}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
989 {"fast", "Constant quantizer", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_FAST}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
990 {"aac_ms", "Force M/S stereo coding", offsetof(AACEncContext, options.mid_side), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
991 {"aac_is", "Intensity stereo coding", offsetof(AACEncContext, options.intensity_stereo), AV_OPT_TYPE_BOOL, {.i64 = OPT_AUTO}, -1, 1, AACENC_FLAGS},
992 {"aac_pns", "Perceptual noise substitution", offsetof(AACEncContext, options.pns), AV_OPT_TYPE_BOOL, {.i64 = OPT_AUTO}, -1, 1, AACENC_FLAGS},
993 {"aac_tns", "Temporal noise shaping", offsetof(AACEncContext, options.tns), AV_OPT_TYPE_BOOL, {.i64 = OPT_AUTO}, -1, 1, AACENC_FLAGS},
994 {"aac_pred", "AAC-Main prediction", offsetof(AACEncContext, options.pred), AV_OPT_TYPE_BOOL, {.i64 = OPT_AUTO}, -1, 1, AACENC_FLAGS},
998 static const AVClass aacenc_class = {
1000 av_default_item_name,
1002 LIBAVUTIL_VERSION_INT,
1005 AVCodec ff_aac_encoder = {
1007 .long_name = NULL_IF_CONFIG_SMALL("AAC (Advanced Audio Coding)"),
1008 .type = AVMEDIA_TYPE_AUDIO,
1009 .id = AV_CODEC_ID_AAC,
1010 .priv_data_size = sizeof(AACEncContext),
1011 .init = aac_encode_init,
1012 .encode2 = aac_encode_frame,
1013 .close = aac_encode_end,
1014 .supported_samplerates = mpeg4audio_sample_rates,
1015 .capabilities = AV_CODEC_CAP_SMALL_LAST_FRAME | AV_CODEC_CAP_DELAY |
1016 AV_CODEC_CAP_EXPERIMENTAL,
1017 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
1018 AV_SAMPLE_FMT_NONE },
1019 .priv_class = &aacenc_class,