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"
50 * Make AAC audio config object.
51 * @see 1.6.2.1 "Syntax - AudioSpecificConfig"
53 static void put_audio_specific_config(AVCodecContext *avctx)
56 AACEncContext *s = avctx->priv_data;
57 int channels = s->channels - (s->channels == 8 ? 1 : 0);
59 init_put_bits(&pb, avctx->extradata, avctx->extradata_size);
60 put_bits(&pb, 5, s->profile+1); //profile
61 put_bits(&pb, 4, s->samplerate_index); //sample rate index
62 put_bits(&pb, 4, channels);
64 put_bits(&pb, 1, 0); //frame length - 1024 samples
65 put_bits(&pb, 1, 0); //does not depend on core coder
66 put_bits(&pb, 1, 0); //is not extension
68 //Explicitly Mark SBR absent
69 put_bits(&pb, 11, 0x2b7); //sync extension
70 put_bits(&pb, 5, AOT_SBR);
75 void ff_quantize_band_cost_cache_init(struct AACEncContext *s)
78 for (sf = 0; sf < 256; sf++) {
79 for (g = 0; g < 128; g++) {
80 s->quantize_band_cost_cache[sf][g].bits = -1;
85 #define WINDOW_FUNC(type) \
86 static void apply_ ##type ##_window(AVFloatDSPContext *fdsp, \
87 SingleChannelElement *sce, \
90 WINDOW_FUNC(only_long)
92 const float *lwindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
93 const float *pwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
94 float *out = sce->ret_buf;
96 fdsp->vector_fmul (out, audio, lwindow, 1024);
97 fdsp->vector_fmul_reverse(out + 1024, audio + 1024, pwindow, 1024);
100 WINDOW_FUNC(long_start)
102 const float *lwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
103 const float *swindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
104 float *out = sce->ret_buf;
106 fdsp->vector_fmul(out, audio, lwindow, 1024);
107 memcpy(out + 1024, audio + 1024, sizeof(out[0]) * 448);
108 fdsp->vector_fmul_reverse(out + 1024 + 448, audio + 1024 + 448, swindow, 128);
109 memset(out + 1024 + 576, 0, sizeof(out[0]) * 448);
112 WINDOW_FUNC(long_stop)
114 const float *lwindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
115 const float *swindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;
116 float *out = sce->ret_buf;
118 memset(out, 0, sizeof(out[0]) * 448);
119 fdsp->vector_fmul(out + 448, audio + 448, swindow, 128);
120 memcpy(out + 576, audio + 576, sizeof(out[0]) * 448);
121 fdsp->vector_fmul_reverse(out + 1024, audio + 1024, lwindow, 1024);
124 WINDOW_FUNC(eight_short)
126 const float *swindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
127 const float *pwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;
128 const float *in = audio + 448;
129 float *out = sce->ret_buf;
132 for (w = 0; w < 8; w++) {
133 fdsp->vector_fmul (out, in, w ? pwindow : swindow, 128);
136 fdsp->vector_fmul_reverse(out, in, swindow, 128);
141 static void (*const apply_window[4])(AVFloatDSPContext *fdsp,
142 SingleChannelElement *sce,
143 const float *audio) = {
144 [ONLY_LONG_SEQUENCE] = apply_only_long_window,
145 [LONG_START_SEQUENCE] = apply_long_start_window,
146 [EIGHT_SHORT_SEQUENCE] = apply_eight_short_window,
147 [LONG_STOP_SEQUENCE] = apply_long_stop_window
150 static void apply_window_and_mdct(AACEncContext *s, SingleChannelElement *sce,
154 float *output = sce->ret_buf;
156 apply_window[sce->ics.window_sequence[0]](s->fdsp, sce, audio);
158 if (sce->ics.window_sequence[0] != EIGHT_SHORT_SEQUENCE)
159 s->mdct1024.mdct_calc(&s->mdct1024, sce->coeffs, output);
161 for (i = 0; i < 1024; i += 128)
162 s->mdct128.mdct_calc(&s->mdct128, &sce->coeffs[i], output + i*2);
163 memcpy(audio, audio + 1024, sizeof(audio[0]) * 1024);
164 memcpy(sce->pcoeffs, sce->coeffs, sizeof(sce->pcoeffs));
168 * Encode ics_info element.
169 * @see Table 4.6 (syntax of ics_info)
171 static void put_ics_info(AACEncContext *s, IndividualChannelStream *info)
175 put_bits(&s->pb, 1, 0); // ics_reserved bit
176 put_bits(&s->pb, 2, info->window_sequence[0]);
177 put_bits(&s->pb, 1, info->use_kb_window[0]);
178 if (info->window_sequence[0] != EIGHT_SHORT_SEQUENCE) {
179 put_bits(&s->pb, 6, info->max_sfb);
180 put_bits(&s->pb, 1, !!info->predictor_present);
182 put_bits(&s->pb, 4, info->max_sfb);
183 for (w = 1; w < 8; w++)
184 put_bits(&s->pb, 1, !info->group_len[w]);
190 * @see 4.6.8.1 "Joint Coding - M/S Stereo"
192 static void encode_ms_info(PutBitContext *pb, ChannelElement *cpe)
196 put_bits(pb, 2, cpe->ms_mode);
197 if (cpe->ms_mode == 1)
198 for (w = 0; w < cpe->ch[0].ics.num_windows; w += cpe->ch[0].ics.group_len[w])
199 for (i = 0; i < cpe->ch[0].ics.max_sfb; i++)
200 put_bits(pb, 1, cpe->ms_mask[w*16 + i]);
204 * Produce integer coefficients from scalefactors provided by the model.
206 static void adjust_frame_information(ChannelElement *cpe, int chans)
211 for (ch = 0; ch < chans; ch++) {
212 IndividualChannelStream *ics = &cpe->ch[ch].ics;
214 cpe->ch[ch].pulse.num_pulse = 0;
215 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
216 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
217 for (cmaxsfb = ics->num_swb; cmaxsfb > 0 && cpe->ch[ch].zeroes[w*16+cmaxsfb-1]; cmaxsfb--)
219 maxsfb = FFMAX(maxsfb, cmaxsfb);
222 ics->max_sfb = maxsfb;
224 //adjust zero bands for window groups
225 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
226 for (g = 0; g < ics->max_sfb; g++) {
228 for (w2 = w; w2 < w + ics->group_len[w]; w2++) {
229 if (!cpe->ch[ch].zeroes[w2*16 + g]) {
234 cpe->ch[ch].zeroes[w*16 + g] = i;
239 if (chans > 1 && cpe->common_window) {
240 IndividualChannelStream *ics0 = &cpe->ch[0].ics;
241 IndividualChannelStream *ics1 = &cpe->ch[1].ics;
243 ics0->max_sfb = FFMAX(ics0->max_sfb, ics1->max_sfb);
244 ics1->max_sfb = ics0->max_sfb;
245 for (w = 0; w < ics0->num_windows*16; w += 16)
246 for (i = 0; i < ics0->max_sfb; i++)
247 if (cpe->ms_mask[w+i])
249 if (msc == 0 || ics0->max_sfb == 0)
252 cpe->ms_mode = msc < ics0->max_sfb * ics0->num_windows ? 1 : 2;
256 static void apply_intensity_stereo(ChannelElement *cpe)
259 IndividualChannelStream *ics = &cpe->ch[0].ics;
260 if (!cpe->common_window)
262 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
263 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
264 int start = (w+w2) * 128;
265 for (g = 0; g < ics->num_swb; g++) {
266 int p = -1 + 2 * (cpe->ch[1].band_type[w*16+g] - 14);
267 float scale = cpe->ch[0].is_ener[w*16+g];
268 if (!cpe->is_mask[w*16 + g]) {
269 start += ics->swb_sizes[g];
272 if (cpe->ms_mask[w*16 + g])
274 for (i = 0; i < ics->swb_sizes[g]; i++) {
275 float sum = (cpe->ch[0].coeffs[start+i] + p*cpe->ch[1].coeffs[start+i])*scale;
276 cpe->ch[0].coeffs[start+i] = sum;
277 cpe->ch[1].coeffs[start+i] = 0.0f;
279 start += ics->swb_sizes[g];
285 static void apply_mid_side_stereo(ChannelElement *cpe)
288 IndividualChannelStream *ics = &cpe->ch[0].ics;
289 if (!cpe->common_window)
291 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
292 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
293 int start = (w+w2) * 128;
294 for (g = 0; g < ics->num_swb; g++) {
295 if (!cpe->ms_mask[w*16 + g] && !cpe->is_mask[w*16 + g]) {
296 start += ics->swb_sizes[g];
299 for (i = 0; i < ics->swb_sizes[g]; i++) {
300 float L = (cpe->ch[0].coeffs[start+i] + cpe->ch[1].coeffs[start+i]) * 0.5f;
301 float R = L - cpe->ch[1].coeffs[start+i];
302 cpe->ch[0].coeffs[start+i] = L;
303 cpe->ch[1].coeffs[start+i] = R;
305 start += ics->swb_sizes[g];
312 * Encode scalefactor band coding type.
314 static void encode_band_info(AACEncContext *s, SingleChannelElement *sce)
318 if (s->coder->set_special_band_scalefactors)
319 s->coder->set_special_band_scalefactors(s, sce);
321 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
322 s->coder->encode_window_bands_info(s, sce, w, sce->ics.group_len[w], s->lambda);
326 * Encode scalefactors.
328 static void encode_scale_factors(AVCodecContext *avctx, AACEncContext *s,
329 SingleChannelElement *sce)
331 int diff, off_sf = sce->sf_idx[0], off_pns = sce->sf_idx[0] - NOISE_OFFSET;
332 int off_is = 0, noise_flag = 1;
335 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
336 for (i = 0; i < sce->ics.max_sfb; i++) {
337 if (!sce->zeroes[w*16 + i]) {
338 if (sce->band_type[w*16 + i] == NOISE_BT) {
339 diff = sce->sf_idx[w*16 + i] - off_pns;
340 off_pns = sce->sf_idx[w*16 + i];
341 if (noise_flag-- > 0) {
342 put_bits(&s->pb, NOISE_PRE_BITS, diff + NOISE_PRE);
345 } else if (sce->band_type[w*16 + i] == INTENSITY_BT ||
346 sce->band_type[w*16 + i] == INTENSITY_BT2) {
347 diff = sce->sf_idx[w*16 + i] - off_is;
348 off_is = sce->sf_idx[w*16 + i];
350 diff = sce->sf_idx[w*16 + i] - off_sf;
351 off_sf = sce->sf_idx[w*16 + i];
353 diff += SCALE_DIFF_ZERO;
354 av_assert0(diff >= 0 && diff <= 120);
355 put_bits(&s->pb, ff_aac_scalefactor_bits[diff], ff_aac_scalefactor_code[diff]);
364 static void encode_pulses(AACEncContext *s, Pulse *pulse)
368 put_bits(&s->pb, 1, !!pulse->num_pulse);
369 if (!pulse->num_pulse)
372 put_bits(&s->pb, 2, pulse->num_pulse - 1);
373 put_bits(&s->pb, 6, pulse->start);
374 for (i = 0; i < pulse->num_pulse; i++) {
375 put_bits(&s->pb, 5, pulse->pos[i]);
376 put_bits(&s->pb, 4, pulse->amp[i]);
381 * Encode spectral coefficients processed by psychoacoustic model.
383 static void encode_spectral_coeffs(AACEncContext *s, SingleChannelElement *sce)
387 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
389 for (i = 0; i < sce->ics.max_sfb; i++) {
390 if (sce->zeroes[w*16 + i]) {
391 start += sce->ics.swb_sizes[i];
394 for (w2 = w; w2 < w + sce->ics.group_len[w]; w2++) {
395 s->coder->quantize_and_encode_band(s, &s->pb,
396 &sce->coeffs[start + w2*128],
397 NULL, sce->ics.swb_sizes[i],
398 sce->sf_idx[w*16 + i],
399 sce->band_type[w*16 + i],
401 sce->ics.window_clipping[w]);
403 start += sce->ics.swb_sizes[i];
409 * Downscale spectral coefficients for near-clipping windows to avoid artifacts
411 static void avoid_clipping(AACEncContext *s, SingleChannelElement *sce)
415 if (sce->ics.clip_avoidance_factor < 1.0f) {
416 for (w = 0; w < sce->ics.num_windows; w++) {
418 for (i = 0; i < sce->ics.max_sfb; i++) {
419 float *swb_coeffs = &sce->coeffs[start + w*128];
420 for (j = 0; j < sce->ics.swb_sizes[i]; j++)
421 swb_coeffs[j] *= sce->ics.clip_avoidance_factor;
422 start += sce->ics.swb_sizes[i];
429 * Encode one channel of audio data.
431 static int encode_individual_channel(AVCodecContext *avctx, AACEncContext *s,
432 SingleChannelElement *sce,
435 put_bits(&s->pb, 8, sce->sf_idx[0]);
436 if (!common_window) {
437 put_ics_info(s, &sce->ics);
438 if (s->coder->encode_main_pred)
439 s->coder->encode_main_pred(s, sce);
440 if (s->coder->encode_ltp_info)
441 s->coder->encode_ltp_info(s, sce, 0);
443 encode_band_info(s, sce);
444 encode_scale_factors(avctx, s, sce);
445 encode_pulses(s, &sce->pulse);
446 put_bits(&s->pb, 1, !!sce->tns.present);
447 if (s->coder->encode_tns_info)
448 s->coder->encode_tns_info(s, sce);
449 put_bits(&s->pb, 1, 0); //ssr
450 encode_spectral_coeffs(s, sce);
455 * Write some auxiliary information about the created AAC file.
457 static void put_bitstream_info(AACEncContext *s, const char *name)
459 int i, namelen, padbits;
461 namelen = strlen(name) + 2;
462 put_bits(&s->pb, 3, TYPE_FIL);
463 put_bits(&s->pb, 4, FFMIN(namelen, 15));
465 put_bits(&s->pb, 8, namelen - 14);
466 put_bits(&s->pb, 4, 0); //extension type - filler
467 padbits = -put_bits_count(&s->pb) & 7;
468 avpriv_align_put_bits(&s->pb);
469 for (i = 0; i < namelen - 2; i++)
470 put_bits(&s->pb, 8, name[i]);
471 put_bits(&s->pb, 12 - padbits, 0);
475 * Copy input samples.
476 * Channels are reordered from libavcodec's default order to AAC order.
478 static void copy_input_samples(AACEncContext *s, const AVFrame *frame)
481 int end = 2048 + (frame ? frame->nb_samples : 0);
482 const uint8_t *channel_map = aac_chan_maps[s->channels - 1];
484 /* copy and remap input samples */
485 for (ch = 0; ch < s->channels; ch++) {
486 /* copy last 1024 samples of previous frame to the start of the current frame */
487 memcpy(&s->planar_samples[ch][1024], &s->planar_samples[ch][2048], 1024 * sizeof(s->planar_samples[0][0]));
489 /* copy new samples and zero any remaining samples */
491 memcpy(&s->planar_samples[ch][2048],
492 frame->extended_data[channel_map[ch]],
493 frame->nb_samples * sizeof(s->planar_samples[0][0]));
495 memset(&s->planar_samples[ch][end], 0,
496 (3072 - end) * sizeof(s->planar_samples[0][0]));
500 static int aac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
501 const AVFrame *frame, int *got_packet_ptr)
503 AACEncContext *s = avctx->priv_data;
504 float **samples = s->planar_samples, *samples2, *la, *overlap;
506 SingleChannelElement *sce;
507 IndividualChannelStream *ics;
508 int i, its, ch, w, chans, tag, start_ch, ret, frame_bits;
509 int target_bits, rate_bits, too_many_bits, too_few_bits;
510 int ms_mode = 0, is_mode = 0, tns_mode = 0, pred_mode = 0;
511 int chan_el_counter[4];
512 FFPsyWindowInfo windows[AAC_MAX_CHANNELS];
514 if (s->last_frame == 2)
517 /* add current frame to queue */
519 if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
523 copy_input_samples(s, frame);
525 ff_psy_preprocess(s->psypp, s->planar_samples, s->channels);
527 if (!avctx->frame_number)
531 for (i = 0; i < s->chan_map[0]; i++) {
532 FFPsyWindowInfo* wi = windows + start_ch;
533 tag = s->chan_map[i+1];
534 chans = tag == TYPE_CPE ? 2 : 1;
536 for (ch = 0; ch < chans; ch++) {
537 float clip_avoidance_factor;
540 s->cur_channel = start_ch + ch;
541 overlap = &samples[s->cur_channel][0];
542 samples2 = overlap + 1024;
543 la = samples2 + (448+64);
546 if (tag == TYPE_LFE) {
547 wi[ch].window_type[0] = ONLY_LONG_SEQUENCE;
548 wi[ch].window_shape = 0;
549 wi[ch].num_windows = 1;
550 wi[ch].grouping[0] = 1;
552 /* Only the lowest 12 coefficients are used in a LFE channel.
553 * The expression below results in only the bottom 8 coefficients
554 * being used for 11.025kHz to 16kHz sample rates.
556 ics->num_swb = s->samplerate_index >= 8 ? 1 : 3;
558 wi[ch] = s->psy.model->window(&s->psy, samples2, la, s->cur_channel,
559 ics->window_sequence[0]);
561 ics->window_sequence[1] = ics->window_sequence[0];
562 ics->window_sequence[0] = wi[ch].window_type[0];
563 ics->use_kb_window[1] = ics->use_kb_window[0];
564 ics->use_kb_window[0] = wi[ch].window_shape;
565 ics->num_windows = wi[ch].num_windows;
566 ics->swb_sizes = s->psy.bands [ics->num_windows == 8];
567 ics->num_swb = tag == TYPE_LFE ? ics->num_swb : s->psy.num_bands[ics->num_windows == 8];
568 ics->swb_offset = wi[ch].window_type[0] == EIGHT_SHORT_SEQUENCE ?
569 ff_swb_offset_128 [s->samplerate_index]:
570 ff_swb_offset_1024[s->samplerate_index];
571 ics->tns_max_bands = wi[ch].window_type[0] == EIGHT_SHORT_SEQUENCE ?
572 ff_tns_max_bands_128 [s->samplerate_index]:
573 ff_tns_max_bands_1024[s->samplerate_index];
574 clip_avoidance_factor = 0.0f;
575 for (w = 0; w < ics->num_windows; w++)
576 ics->group_len[w] = wi[ch].grouping[w];
577 for (w = 0; w < ics->num_windows; w++) {
578 if (wi[ch].clipping[w] > CLIP_AVOIDANCE_FACTOR) {
579 ics->window_clipping[w] = 1;
580 clip_avoidance_factor = FFMAX(clip_avoidance_factor, wi[ch].clipping[w]);
582 ics->window_clipping[w] = 0;
585 if (clip_avoidance_factor > CLIP_AVOIDANCE_FACTOR) {
586 ics->clip_avoidance_factor = CLIP_AVOIDANCE_FACTOR / clip_avoidance_factor;
588 ics->clip_avoidance_factor = 1.0f;
591 apply_window_and_mdct(s, sce, overlap);
593 if (s->options.ltp && s->coder->update_ltp) {
594 s->coder->update_ltp(s, sce);
595 apply_window[sce->ics.window_sequence[0]](s->fdsp, sce, &sce->ltp_state[0]);
596 s->mdct1024.mdct_calc(&s->mdct1024, sce->lcoeffs, sce->ret_buf);
599 if (isnan(cpe->ch->coeffs[0])) {
600 av_log(avctx, AV_LOG_ERROR, "Input contains NaN\n");
601 return AVERROR(EINVAL);
603 avoid_clipping(s, sce);
607 if ((ret = ff_alloc_packet2(avctx, avpkt, 8192 * s->channels, 0)) < 0)
609 frame_bits = its = 0;
611 init_put_bits(&s->pb, avpkt->data, avpkt->size);
613 if ((avctx->frame_number & 0xFF)==1 && !(avctx->flags & AV_CODEC_FLAG_BITEXACT))
614 put_bitstream_info(s, LIBAVCODEC_IDENT);
617 memset(chan_el_counter, 0, sizeof(chan_el_counter));
618 for (i = 0; i < s->chan_map[0]; i++) {
619 FFPsyWindowInfo* wi = windows + start_ch;
620 const float *coeffs[2];
621 tag = s->chan_map[i+1];
622 chans = tag == TYPE_CPE ? 2 : 1;
624 cpe->common_window = 0;
625 memset(cpe->is_mask, 0, sizeof(cpe->is_mask));
626 memset(cpe->ms_mask, 0, sizeof(cpe->ms_mask));
627 put_bits(&s->pb, 3, tag);
628 put_bits(&s->pb, 4, chan_el_counter[tag]++);
629 for (ch = 0; ch < chans; ch++) {
631 coeffs[ch] = sce->coeffs;
632 sce->ics.predictor_present = 0;
633 sce->ics.ltp.present = 0;
634 memset(sce->ics.ltp.used, 0, sizeof(sce->ics.ltp.used));
635 memset(sce->ics.prediction_used, 0, sizeof(sce->ics.prediction_used));
636 memset(&sce->tns, 0, sizeof(TemporalNoiseShaping));
637 for (w = 0; w < 128; w++)
638 if (sce->band_type[w] > RESERVED_BT)
639 sce->band_type[w] = 0;
641 s->psy.bitres.alloc = -1;
642 s->psy.bitres.bits = avctx->frame_bits / s->channels;
643 s->psy.model->analyze(&s->psy, start_ch, coeffs, wi);
644 if (s->psy.bitres.alloc > 0) {
645 /* Lambda unused here on purpose, we need to take psy's unscaled allocation */
646 target_bits += s->psy.bitres.alloc
647 * (s->lambda / (avctx->global_quality ? avctx->global_quality : 120));
648 s->psy.bitres.alloc /= chans;
651 for (ch = 0; ch < chans; ch++) {
652 s->cur_channel = start_ch + ch;
653 if (s->options.pns && s->coder->mark_pns)
654 s->coder->mark_pns(s, avctx, &cpe->ch[ch]);
655 s->coder->search_for_quantizers(avctx, s, &cpe->ch[ch], s->lambda);
658 && wi[0].window_type[0] == wi[1].window_type[0]
659 && wi[0].window_shape == wi[1].window_shape) {
661 cpe->common_window = 1;
662 for (w = 0; w < wi[0].num_windows; w++) {
663 if (wi[0].grouping[w] != wi[1].grouping[w]) {
664 cpe->common_window = 0;
669 for (ch = 0; ch < chans; ch++) { /* TNS and PNS */
671 s->cur_channel = start_ch + ch;
672 if (s->options.pns && s->coder->search_for_pns)
673 s->coder->search_for_pns(s, avctx, sce);
674 if (s->options.tns && s->coder->search_for_tns)
675 s->coder->search_for_tns(s, sce);
676 if (s->options.tns && s->coder->apply_tns_filt)
677 s->coder->apply_tns_filt(s, sce);
678 if (sce->tns.present)
681 s->cur_channel = start_ch;
682 if (s->options.intensity_stereo) { /* Intensity Stereo */
683 if (s->coder->search_for_is)
684 s->coder->search_for_is(s, avctx, cpe);
685 if (cpe->is_mode) is_mode = 1;
686 apply_intensity_stereo(cpe);
688 if (s->options.pred) { /* Prediction */
689 for (ch = 0; ch < chans; ch++) {
691 s->cur_channel = start_ch + ch;
692 if (s->options.pred && s->coder->search_for_pred)
693 s->coder->search_for_pred(s, sce);
694 if (cpe->ch[ch].ics.predictor_present) pred_mode = 1;
696 if (s->coder->adjust_common_pred)
697 s->coder->adjust_common_pred(s, cpe);
698 for (ch = 0; ch < chans; ch++) {
700 s->cur_channel = start_ch + ch;
701 if (s->options.pred && s->coder->apply_main_pred)
702 s->coder->apply_main_pred(s, sce);
704 s->cur_channel = start_ch;
706 if (s->options.mid_side) { /* Mid/Side stereo */
707 if (s->options.mid_side == -1 && s->coder->search_for_ms)
708 s->coder->search_for_ms(s, cpe);
709 else if (cpe->common_window)
710 memset(cpe->ms_mask, 1, sizeof(cpe->ms_mask));
711 apply_mid_side_stereo(cpe);
713 adjust_frame_information(cpe, chans);
714 if (s->options.ltp) { /* LTP */
715 for (ch = 0; ch < chans; ch++) {
717 s->cur_channel = start_ch + ch;
718 if (s->coder->search_for_ltp)
719 s->coder->search_for_ltp(s, sce, cpe->common_window);
720 if (sce->ics.ltp.present) pred_mode = 1;
722 s->cur_channel = start_ch;
723 if (s->coder->adjust_common_ltp)
724 s->coder->adjust_common_ltp(s, cpe);
727 put_bits(&s->pb, 1, cpe->common_window);
728 if (cpe->common_window) {
729 put_ics_info(s, &cpe->ch[0].ics);
730 if (s->coder->encode_main_pred)
731 s->coder->encode_main_pred(s, &cpe->ch[0]);
732 if (s->coder->encode_ltp_info)
733 s->coder->encode_ltp_info(s, &cpe->ch[0], 1);
734 encode_ms_info(&s->pb, cpe);
735 if (cpe->ms_mode) ms_mode = 1;
738 for (ch = 0; ch < chans; ch++) {
739 s->cur_channel = start_ch + ch;
740 encode_individual_channel(avctx, s, &cpe->ch[ch], cpe->common_window);
745 if (avctx->flags & CODEC_FLAG_QSCALE) {
746 /* When using a constant Q-scale, don't mess with lambda */
750 /* rate control stuff
751 * allow between the nominal bitrate, and what psy's bit reservoir says to target
752 * but drift towards the nominal bitrate always
754 frame_bits = put_bits_count(&s->pb);
755 rate_bits = avctx->bit_rate * 1024 / avctx->sample_rate;
756 rate_bits = FFMIN(rate_bits, 6144 * s->channels - 3);
757 too_many_bits = FFMAX(target_bits, rate_bits);
758 too_many_bits = FFMIN(too_many_bits, 6144 * s->channels - 3);
759 too_few_bits = FFMIN(FFMAX(rate_bits - rate_bits/4, target_bits), too_many_bits);
761 /* When using ABR, be strict (but only for increasing) */
762 too_few_bits = too_few_bits - too_few_bits/8;
763 too_many_bits = too_many_bits + too_many_bits/2;
765 if ( its == 0 /* for steady-state Q-scale tracking */
766 || (its < 5 && (frame_bits < too_few_bits || frame_bits > too_many_bits))
767 || frame_bits >= 6144 * s->channels - 3 )
769 float ratio = ((float)rate_bits) / frame_bits;
771 if (frame_bits >= too_few_bits && frame_bits <= too_many_bits) {
773 * This path is for steady-state Q-scale tracking
774 * When frame bits fall within the stable range, we still need to adjust
775 * lambda to maintain it like so in a stable fashion (large jumps in lambda
776 * create artifacts and should be avoided), but slowly
778 ratio = sqrtf(sqrtf(ratio));
779 ratio = av_clipf(ratio, 0.9f, 1.1f);
781 /* Not so fast though */
782 ratio = sqrtf(ratio);
784 s->lambda = FFMIN(s->lambda * ratio, 65536.f);
786 /* Keep iterating if we must reduce and lambda is in the sky */
787 if ((s->lambda < 300.f || ratio > 0.9f) && (s->lambda > 10.f || ratio < 1.1f)) {
790 if (is_mode || ms_mode || tns_mode || pred_mode) {
791 for (i = 0; i < s->chan_map[0]; i++) {
792 // Must restore coeffs
793 chans = tag == TYPE_CPE ? 2 : 1;
795 for (ch = 0; ch < chans; ch++)
796 memcpy(cpe->ch[ch].coeffs, cpe->ch[ch].pcoeffs, sizeof(cpe->ch[ch].coeffs));
806 if (s->options.ltp && s->coder->ltp_insert_new_frame)
807 s->coder->ltp_insert_new_frame(s);
809 put_bits(&s->pb, 3, TYPE_END);
810 flush_put_bits(&s->pb);
811 avctx->frame_bits = put_bits_count(&s->pb);
812 s->lambda_sum += s->lambda;
818 ff_af_queue_remove(&s->afq, avctx->frame_size, &avpkt->pts,
821 avpkt->size = put_bits_count(&s->pb) >> 3;
826 static av_cold int aac_encode_end(AVCodecContext *avctx)
828 AACEncContext *s = avctx->priv_data;
830 av_log(avctx, AV_LOG_INFO, "Qavg: %.3f\n", s->lambda_sum / s->lambda_count);
832 ff_mdct_end(&s->mdct1024);
833 ff_mdct_end(&s->mdct128);
837 ff_psy_preprocess_end(s->psypp);
838 av_freep(&s->buffer.samples);
841 ff_af_queue_close(&s->afq);
845 static av_cold int dsp_init(AVCodecContext *avctx, AACEncContext *s)
849 s->fdsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);
851 return AVERROR(ENOMEM);
854 ff_kbd_window_init(ff_aac_kbd_long_1024, 4.0, 1024);
855 ff_kbd_window_init(ff_aac_kbd_short_128, 6.0, 128);
856 ff_init_ff_sine_windows(10);
857 ff_init_ff_sine_windows(7);
859 if ((ret = ff_mdct_init(&s->mdct1024, 11, 0, 32768.0)) < 0)
861 if ((ret = ff_mdct_init(&s->mdct128, 8, 0, 32768.0)) < 0)
867 static av_cold int alloc_buffers(AVCodecContext *avctx, AACEncContext *s)
870 FF_ALLOCZ_ARRAY_OR_GOTO(avctx, s->buffer.samples, s->channels, 3 * 1024 * sizeof(s->buffer.samples[0]), alloc_fail);
871 FF_ALLOCZ_ARRAY_OR_GOTO(avctx, s->cpe, s->chan_map[0], sizeof(ChannelElement), alloc_fail);
872 FF_ALLOCZ_OR_GOTO(avctx, avctx->extradata, 5 + AV_INPUT_BUFFER_PADDING_SIZE, alloc_fail);
874 for(ch = 0; ch < s->channels; ch++)
875 s->planar_samples[ch] = s->buffer.samples + 3 * 1024 * ch;
879 return AVERROR(ENOMEM);
882 static av_cold int aac_encode_init(AVCodecContext *avctx)
884 AACEncContext *s = avctx->priv_data;
886 const uint8_t *sizes[2];
887 uint8_t grouping[AAC_MAX_CHANNELS];
890 s->channels = avctx->channels;
891 s->chan_map = aac_chan_configs[s->channels-1];
892 s->random_state = 0x1f2e3d4c;
893 s->lambda = avctx->global_quality > 0 ? avctx->global_quality : 120;
894 avctx->extradata_size = 5;
895 avctx->frame_size = 1024;
896 avctx->initial_padding = 1024;
897 avctx->bit_rate = (int)FFMIN(
898 6144 * s->channels / 1024.0 * avctx->sample_rate,
900 avctx->profile = avctx->profile == FF_PROFILE_UNKNOWN ? FF_PROFILE_AAC_LOW :
903 for (i = 0; i < 16; i++)
904 if (avctx->sample_rate == avpriv_mpeg4audio_sample_rates[i])
906 s->samplerate_index = i;
908 ERROR_IF(s->samplerate_index == 16 ||
909 s->samplerate_index >= ff_aac_swb_size_1024_len ||
910 s->samplerate_index >= ff_aac_swb_size_128_len,
911 "Unsupported sample rate %d\n", avctx->sample_rate);
912 ERROR_IF(s->channels > AAC_MAX_CHANNELS || s->channels == 7,
913 "Unsupported number of channels: %d\n", s->channels);
914 WARN_IF(1024.0 * avctx->bit_rate / avctx->sample_rate > 6144 * s->channels,
915 "Too many bits per frame requested, clamping to max\n");
917 for (i = 0; i < FF_ARRAY_ELEMS(aacenc_profiles); i++)
918 if (avctx->profile == aacenc_profiles[i])
920 ERROR_IF(i == FF_ARRAY_ELEMS(aacenc_profiles),
921 "Unsupported encoding profile: %d\n", avctx->profile);
922 if (avctx->profile == FF_PROFILE_MPEG2_AAC_LOW) {
923 avctx->profile = FF_PROFILE_AAC_LOW;
924 ERROR_IF(s->options.pred,
925 "Main prediction unavailable in the \"mpeg2_aac_low\" profile\n");
926 ERROR_IF(s->options.ltp,
927 "LTP prediction unavailable in the \"mpeg2_aac_low\" profile\n");
928 WARN_IF(s->options.pns,
929 "PNS unavailable in the \"mpeg2_aac_low\" profile, turning off\n");
931 } else if (avctx->profile == FF_PROFILE_AAC_LTP) {
933 ERROR_IF(s->options.pred,
934 "Main prediction unavailable in the \"aac_ltp\" profile\n");
935 } else if (avctx->profile == FF_PROFILE_AAC_MAIN) {
937 ERROR_IF(s->options.ltp,
938 "LTP prediction unavailable in the \"aac_main\" profile\n");
939 } else if (s->options.ltp) {
940 avctx->profile = FF_PROFILE_AAC_LTP;
942 "Chainging profile to \"aac_ltp\"\n");
943 ERROR_IF(s->options.pred,
944 "Main prediction unavailable in the \"aac_ltp\" profile\n");
945 } else if (s->options.pred) {
946 avctx->profile = FF_PROFILE_AAC_MAIN;
948 "Chainging profile to \"aac_main\"\n");
949 ERROR_IF(s->options.pred,
950 "LTP prediction unavailable in the \"aac_main\" profile\n");
952 s->profile = avctx->profile;
953 s->coder = &ff_aac_coders[s->options.coder];
955 if (s->options.coder != AAC_CODER_TWOLOOP) {
956 s->options.intensity_stereo = 0;
960 if ((ret = dsp_init(avctx, s)) < 0)
963 if ((ret = alloc_buffers(avctx, s)) < 0)
966 put_audio_specific_config(avctx);
968 sizes[0] = ff_aac_swb_size_1024[s->samplerate_index];
969 sizes[1] = ff_aac_swb_size_128[s->samplerate_index];
970 lengths[0] = ff_aac_num_swb_1024[s->samplerate_index];
971 lengths[1] = ff_aac_num_swb_128[s->samplerate_index];
972 for (i = 0; i < s->chan_map[0]; i++)
973 grouping[i] = s->chan_map[i + 1] == TYPE_CPE;
974 if ((ret = ff_psy_init(&s->psy, avctx, 2, sizes, lengths,
975 s->chan_map[0], grouping)) < 0)
977 s->psypp = ff_psy_preprocess_init(avctx);
978 ff_lpc_init(&s->lpc, 2*avctx->frame_size, TNS_MAX_ORDER, FF_LPC_TYPE_LEVINSON);
981 ff_aac_coder_init_mips(s);
985 ff_af_queue_init(avctx, &s->afq);
989 aac_encode_end(avctx);
993 #define AACENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
994 static const AVOption aacenc_options[] = {
995 {"aac_coder", "Coding algorithm", offsetof(AACEncContext, options.coder), AV_OPT_TYPE_INT, {.i64 = AAC_CODER_TWOLOOP}, -1, AAC_CODER_NB-1, AACENC_FLAGS, "coder"},
996 {"faac", "FAAC-inspired method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_FAAC}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
997 {"anmr", "ANMR method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_ANMR}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
998 {"twoloop", "Two loop searching method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_TWOLOOP}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
999 {"fast", "Constant quantizer", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_FAST}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
1000 {"aac_ms", "Force M/S stereo coding", offsetof(AACEncContext, options.mid_side), AV_OPT_TYPE_BOOL, {.i64 = -1}, -1, 1, AACENC_FLAGS},
1001 {"aac_is", "Intensity stereo coding", offsetof(AACEncContext, options.intensity_stereo), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1002 {"aac_pns", "Perceptual noise substitution", offsetof(AACEncContext, options.pns), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1003 {"aac_tns", "Temporal noise shaping", offsetof(AACEncContext, options.tns), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1004 {"aac_ltp", "Long term prediction", offsetof(AACEncContext, options.ltp), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
1005 {"aac_pred", "AAC-Main prediction", offsetof(AACEncContext, options.pred), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
1009 static const AVClass aacenc_class = {
1011 av_default_item_name,
1013 LIBAVUTIL_VERSION_INT,
1016 AVCodec ff_aac_encoder = {
1018 .long_name = NULL_IF_CONFIG_SMALL("AAC (Advanced Audio Coding)"),
1019 .type = AVMEDIA_TYPE_AUDIO,
1020 .id = AV_CODEC_ID_AAC,
1021 .priv_data_size = sizeof(AACEncContext),
1022 .init = aac_encode_init,
1023 .encode2 = aac_encode_frame,
1024 .close = aac_encode_end,
1025 .supported_samplerates = mpeg4audio_sample_rates,
1026 .capabilities = AV_CODEC_CAP_SMALL_LAST_FRAME | AV_CODEC_CAP_DELAY |
1027 AV_CODEC_CAP_EXPERIMENTAL,
1028 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
1029 AV_SAMPLE_FMT_NONE },
1030 .priv_class = &aacenc_class,