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/libm.h"
33 #include "libavutil/thread.h"
34 #include "libavutil/float_dsp.h"
35 #include "libavutil/opt.h"
39 #include "mpeg4audio.h"
46 #include "aacenctab.h"
47 #include "aacenc_utils.h"
51 static AVOnce aac_table_init = AV_ONCE_INIT;
53 static void put_pce(PutBitContext *pb, AVCodecContext *avctx)
56 AACEncContext *s = avctx->priv_data;
57 AACPCEInfo *pce = &s->pce;
58 const int bitexact = avctx->flags & AV_CODEC_FLAG_BITEXACT;
59 const char *aux_data = bitexact ? "Lavc" : LIBAVCODEC_IDENT;
63 put_bits(pb, 2, avctx->profile);
64 put_bits(pb, 4, s->samplerate_index);
66 put_bits(pb, 4, pce->num_ele[0]); /* Front */
67 put_bits(pb, 4, pce->num_ele[1]); /* Side */
68 put_bits(pb, 4, pce->num_ele[2]); /* Back */
69 put_bits(pb, 2, pce->num_ele[3]); /* LFE */
70 put_bits(pb, 3, 0); /* Assoc data */
71 put_bits(pb, 4, 0); /* CCs */
73 put_bits(pb, 1, 0); /* Stereo mixdown */
74 put_bits(pb, 1, 0); /* Mono mixdown */
75 put_bits(pb, 1, 0); /* Something else */
77 for (i = 0; i < 4; i++) {
78 for (j = 0; j < pce->num_ele[i]; j++) {
80 put_bits(pb, 1, pce->pairing[i][j]);
81 put_bits(pb, 4, pce->index[i][j]);
85 avpriv_align_put_bits(pb);
86 put_bits(pb, 8, strlen(aux_data));
87 avpriv_put_string(pb, aux_data, 0);
91 * Make AAC audio config object.
92 * @see 1.6.2.1 "Syntax - AudioSpecificConfig"
94 static int put_audio_specific_config(AVCodecContext *avctx)
97 AACEncContext *s = avctx->priv_data;
98 int channels = (!s->needs_pce)*(s->channels - (s->channels == 8 ? 1 : 0));
99 const int max_size = 32;
101 avctx->extradata = av_mallocz(max_size);
102 if (!avctx->extradata)
103 return AVERROR(ENOMEM);
105 init_put_bits(&pb, avctx->extradata, max_size);
106 put_bits(&pb, 5, s->profile+1); //profile
107 put_bits(&pb, 4, s->samplerate_index); //sample rate index
108 put_bits(&pb, 4, channels);
110 put_bits(&pb, 1, 0); //frame length - 1024 samples
111 put_bits(&pb, 1, 0); //does not depend on core coder
112 put_bits(&pb, 1, 0); //is not extension
116 //Explicitly Mark SBR absent
117 put_bits(&pb, 11, 0x2b7); //sync extension
118 put_bits(&pb, 5, AOT_SBR);
121 avctx->extradata_size = put_bits_count(&pb) >> 3;
126 void ff_quantize_band_cost_cache_init(struct AACEncContext *s)
128 ++s->quantize_band_cost_cache_generation;
129 if (s->quantize_band_cost_cache_generation == 0) {
130 memset(s->quantize_band_cost_cache, 0, sizeof(s->quantize_band_cost_cache));
131 s->quantize_band_cost_cache_generation = 1;
135 #define WINDOW_FUNC(type) \
136 static void apply_ ##type ##_window(AVFloatDSPContext *fdsp, \
137 SingleChannelElement *sce, \
140 WINDOW_FUNC(only_long)
142 const float *lwindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
143 const float *pwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
144 float *out = sce->ret_buf;
146 fdsp->vector_fmul (out, audio, lwindow, 1024);
147 fdsp->vector_fmul_reverse(out + 1024, audio + 1024, pwindow, 1024);
150 WINDOW_FUNC(long_start)
152 const float *lwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
153 const float *swindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
154 float *out = sce->ret_buf;
156 fdsp->vector_fmul(out, audio, lwindow, 1024);
157 memcpy(out + 1024, audio + 1024, sizeof(out[0]) * 448);
158 fdsp->vector_fmul_reverse(out + 1024 + 448, audio + 1024 + 448, swindow, 128);
159 memset(out + 1024 + 576, 0, sizeof(out[0]) * 448);
162 WINDOW_FUNC(long_stop)
164 const float *lwindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
165 const float *swindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;
166 float *out = sce->ret_buf;
168 memset(out, 0, sizeof(out[0]) * 448);
169 fdsp->vector_fmul(out + 448, audio + 448, swindow, 128);
170 memcpy(out + 576, audio + 576, sizeof(out[0]) * 448);
171 fdsp->vector_fmul_reverse(out + 1024, audio + 1024, lwindow, 1024);
174 WINDOW_FUNC(eight_short)
176 const float *swindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
177 const float *pwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;
178 const float *in = audio + 448;
179 float *out = sce->ret_buf;
182 for (w = 0; w < 8; w++) {
183 fdsp->vector_fmul (out, in, w ? pwindow : swindow, 128);
186 fdsp->vector_fmul_reverse(out, in, swindow, 128);
191 static void (*const apply_window[4])(AVFloatDSPContext *fdsp,
192 SingleChannelElement *sce,
193 const float *audio) = {
194 [ONLY_LONG_SEQUENCE] = apply_only_long_window,
195 [LONG_START_SEQUENCE] = apply_long_start_window,
196 [EIGHT_SHORT_SEQUENCE] = apply_eight_short_window,
197 [LONG_STOP_SEQUENCE] = apply_long_stop_window
200 static void apply_window_and_mdct(AACEncContext *s, SingleChannelElement *sce,
204 const float *output = sce->ret_buf;
206 apply_window[sce->ics.window_sequence[0]](s->fdsp, sce, audio);
208 if (sce->ics.window_sequence[0] != EIGHT_SHORT_SEQUENCE)
209 s->mdct1024.mdct_calc(&s->mdct1024, sce->coeffs, output);
211 for (i = 0; i < 1024; i += 128)
212 s->mdct128.mdct_calc(&s->mdct128, &sce->coeffs[i], output + i*2);
213 memcpy(audio, audio + 1024, sizeof(audio[0]) * 1024);
214 memcpy(sce->pcoeffs, sce->coeffs, sizeof(sce->pcoeffs));
218 * Encode ics_info element.
219 * @see Table 4.6 (syntax of ics_info)
221 static void put_ics_info(AACEncContext *s, IndividualChannelStream *info)
225 put_bits(&s->pb, 1, 0); // ics_reserved bit
226 put_bits(&s->pb, 2, info->window_sequence[0]);
227 put_bits(&s->pb, 1, info->use_kb_window[0]);
228 if (info->window_sequence[0] != EIGHT_SHORT_SEQUENCE) {
229 put_bits(&s->pb, 6, info->max_sfb);
230 put_bits(&s->pb, 1, !!info->predictor_present);
232 put_bits(&s->pb, 4, info->max_sfb);
233 for (w = 1; w < 8; w++)
234 put_bits(&s->pb, 1, !info->group_len[w]);
240 * @see 4.6.8.1 "Joint Coding - M/S Stereo"
242 static void encode_ms_info(PutBitContext *pb, ChannelElement *cpe)
246 put_bits(pb, 2, cpe->ms_mode);
247 if (cpe->ms_mode == 1)
248 for (w = 0; w < cpe->ch[0].ics.num_windows; w += cpe->ch[0].ics.group_len[w])
249 for (i = 0; i < cpe->ch[0].ics.max_sfb; i++)
250 put_bits(pb, 1, cpe->ms_mask[w*16 + i]);
254 * Produce integer coefficients from scalefactors provided by the model.
256 static void adjust_frame_information(ChannelElement *cpe, int chans)
261 for (ch = 0; ch < chans; ch++) {
262 IndividualChannelStream *ics = &cpe->ch[ch].ics;
264 cpe->ch[ch].pulse.num_pulse = 0;
265 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
266 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
267 for (cmaxsfb = ics->num_swb; cmaxsfb > 0 && cpe->ch[ch].zeroes[w*16+cmaxsfb-1]; cmaxsfb--)
269 maxsfb = FFMAX(maxsfb, cmaxsfb);
272 ics->max_sfb = maxsfb;
274 //adjust zero bands for window groups
275 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
276 for (g = 0; g < ics->max_sfb; g++) {
278 for (w2 = w; w2 < w + ics->group_len[w]; w2++) {
279 if (!cpe->ch[ch].zeroes[w2*16 + g]) {
284 cpe->ch[ch].zeroes[w*16 + g] = i;
289 if (chans > 1 && cpe->common_window) {
290 IndividualChannelStream *ics0 = &cpe->ch[0].ics;
291 IndividualChannelStream *ics1 = &cpe->ch[1].ics;
293 ics0->max_sfb = FFMAX(ics0->max_sfb, ics1->max_sfb);
294 ics1->max_sfb = ics0->max_sfb;
295 for (w = 0; w < ics0->num_windows*16; w += 16)
296 for (i = 0; i < ics0->max_sfb; i++)
297 if (cpe->ms_mask[w+i])
299 if (msc == 0 || ics0->max_sfb == 0)
302 cpe->ms_mode = msc < ics0->max_sfb * ics0->num_windows ? 1 : 2;
306 static void apply_intensity_stereo(ChannelElement *cpe)
309 IndividualChannelStream *ics = &cpe->ch[0].ics;
310 if (!cpe->common_window)
312 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
313 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
314 int start = (w+w2) * 128;
315 for (g = 0; g < ics->num_swb; g++) {
316 int p = -1 + 2 * (cpe->ch[1].band_type[w*16+g] - 14);
317 float scale = cpe->ch[0].is_ener[w*16+g];
318 if (!cpe->is_mask[w*16 + g]) {
319 start += ics->swb_sizes[g];
322 if (cpe->ms_mask[w*16 + g])
324 for (i = 0; i < ics->swb_sizes[g]; i++) {
325 float sum = (cpe->ch[0].coeffs[start+i] + p*cpe->ch[1].coeffs[start+i])*scale;
326 cpe->ch[0].coeffs[start+i] = sum;
327 cpe->ch[1].coeffs[start+i] = 0.0f;
329 start += ics->swb_sizes[g];
335 static void apply_mid_side_stereo(ChannelElement *cpe)
338 IndividualChannelStream *ics = &cpe->ch[0].ics;
339 if (!cpe->common_window)
341 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
342 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
343 int start = (w+w2) * 128;
344 for (g = 0; g < ics->num_swb; g++) {
345 /* ms_mask can be used for other purposes in PNS and I/S,
346 * so must not apply M/S if any band uses either, even if
349 if (!cpe->ms_mask[w*16 + g] || cpe->is_mask[w*16 + g]
350 || cpe->ch[0].band_type[w*16 + g] >= NOISE_BT
351 || cpe->ch[1].band_type[w*16 + g] >= NOISE_BT) {
352 start += ics->swb_sizes[g];
355 for (i = 0; i < ics->swb_sizes[g]; i++) {
356 float L = (cpe->ch[0].coeffs[start+i] + cpe->ch[1].coeffs[start+i]) * 0.5f;
357 float R = L - cpe->ch[1].coeffs[start+i];
358 cpe->ch[0].coeffs[start+i] = L;
359 cpe->ch[1].coeffs[start+i] = R;
361 start += ics->swb_sizes[g];
368 * Encode scalefactor band coding type.
370 static void encode_band_info(AACEncContext *s, SingleChannelElement *sce)
374 if (s->coder->set_special_band_scalefactors)
375 s->coder->set_special_band_scalefactors(s, sce);
377 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
378 s->coder->encode_window_bands_info(s, sce, w, sce->ics.group_len[w], s->lambda);
382 * Encode scalefactors.
384 static void encode_scale_factors(AVCodecContext *avctx, AACEncContext *s,
385 SingleChannelElement *sce)
387 int diff, off_sf = sce->sf_idx[0], off_pns = sce->sf_idx[0] - NOISE_OFFSET;
388 int off_is = 0, noise_flag = 1;
391 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
392 for (i = 0; i < sce->ics.max_sfb; i++) {
393 if (!sce->zeroes[w*16 + i]) {
394 if (sce->band_type[w*16 + i] == NOISE_BT) {
395 diff = sce->sf_idx[w*16 + i] - off_pns;
396 off_pns = sce->sf_idx[w*16 + i];
397 if (noise_flag-- > 0) {
398 put_bits(&s->pb, NOISE_PRE_BITS, diff + NOISE_PRE);
401 } else if (sce->band_type[w*16 + i] == INTENSITY_BT ||
402 sce->band_type[w*16 + i] == INTENSITY_BT2) {
403 diff = sce->sf_idx[w*16 + i] - off_is;
404 off_is = sce->sf_idx[w*16 + i];
406 diff = sce->sf_idx[w*16 + i] - off_sf;
407 off_sf = sce->sf_idx[w*16 + i];
409 diff += SCALE_DIFF_ZERO;
410 av_assert0(diff >= 0 && diff <= 120);
411 put_bits(&s->pb, ff_aac_scalefactor_bits[diff], ff_aac_scalefactor_code[diff]);
420 static void encode_pulses(AACEncContext *s, Pulse *pulse)
424 put_bits(&s->pb, 1, !!pulse->num_pulse);
425 if (!pulse->num_pulse)
428 put_bits(&s->pb, 2, pulse->num_pulse - 1);
429 put_bits(&s->pb, 6, pulse->start);
430 for (i = 0; i < pulse->num_pulse; i++) {
431 put_bits(&s->pb, 5, pulse->pos[i]);
432 put_bits(&s->pb, 4, pulse->amp[i]);
437 * Encode spectral coefficients processed by psychoacoustic model.
439 static void encode_spectral_coeffs(AACEncContext *s, SingleChannelElement *sce)
443 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
445 for (i = 0; i < sce->ics.max_sfb; i++) {
446 if (sce->zeroes[w*16 + i]) {
447 start += sce->ics.swb_sizes[i];
450 for (w2 = w; w2 < w + sce->ics.group_len[w]; w2++) {
451 s->coder->quantize_and_encode_band(s, &s->pb,
452 &sce->coeffs[start + w2*128],
453 NULL, sce->ics.swb_sizes[i],
454 sce->sf_idx[w*16 + i],
455 sce->band_type[w*16 + i],
457 sce->ics.window_clipping[w]);
459 start += sce->ics.swb_sizes[i];
465 * Downscale spectral coefficients for near-clipping windows to avoid artifacts
467 static void avoid_clipping(AACEncContext *s, SingleChannelElement *sce)
471 if (sce->ics.clip_avoidance_factor < 1.0f) {
472 for (w = 0; w < sce->ics.num_windows; w++) {
474 for (i = 0; i < sce->ics.max_sfb; i++) {
475 float *swb_coeffs = &sce->coeffs[start + w*128];
476 for (j = 0; j < sce->ics.swb_sizes[i]; j++)
477 swb_coeffs[j] *= sce->ics.clip_avoidance_factor;
478 start += sce->ics.swb_sizes[i];
485 * Encode one channel of audio data.
487 static int encode_individual_channel(AVCodecContext *avctx, AACEncContext *s,
488 SingleChannelElement *sce,
491 put_bits(&s->pb, 8, sce->sf_idx[0]);
492 if (!common_window) {
493 put_ics_info(s, &sce->ics);
494 if (s->coder->encode_main_pred)
495 s->coder->encode_main_pred(s, sce);
496 if (s->coder->encode_ltp_info)
497 s->coder->encode_ltp_info(s, sce, 0);
499 encode_band_info(s, sce);
500 encode_scale_factors(avctx, s, sce);
501 encode_pulses(s, &sce->pulse);
502 put_bits(&s->pb, 1, !!sce->tns.present);
503 if (s->coder->encode_tns_info)
504 s->coder->encode_tns_info(s, sce);
505 put_bits(&s->pb, 1, 0); //ssr
506 encode_spectral_coeffs(s, sce);
511 * Write some auxiliary information about the created AAC file.
513 static void put_bitstream_info(AACEncContext *s, const char *name)
515 int i, namelen, padbits;
517 namelen = strlen(name) + 2;
518 put_bits(&s->pb, 3, TYPE_FIL);
519 put_bits(&s->pb, 4, FFMIN(namelen, 15));
521 put_bits(&s->pb, 8, namelen - 14);
522 put_bits(&s->pb, 4, 0); //extension type - filler
523 padbits = -put_bits_count(&s->pb) & 7;
524 avpriv_align_put_bits(&s->pb);
525 for (i = 0; i < namelen - 2; i++)
526 put_bits(&s->pb, 8, name[i]);
527 put_bits(&s->pb, 12 - padbits, 0);
531 * Copy input samples.
532 * Channels are reordered from libavcodec's default order to AAC order.
534 static void copy_input_samples(AACEncContext *s, const AVFrame *frame)
537 int end = 2048 + (frame ? frame->nb_samples : 0);
538 const uint8_t *channel_map = s->reorder_map;
540 /* copy and remap input samples */
541 for (ch = 0; ch < s->channels; ch++) {
542 /* copy last 1024 samples of previous frame to the start of the current frame */
543 memcpy(&s->planar_samples[ch][1024], &s->planar_samples[ch][2048], 1024 * sizeof(s->planar_samples[0][0]));
545 /* copy new samples and zero any remaining samples */
547 memcpy(&s->planar_samples[ch][2048],
548 frame->extended_data[channel_map[ch]],
549 frame->nb_samples * sizeof(s->planar_samples[0][0]));
551 memset(&s->planar_samples[ch][end], 0,
552 (3072 - end) * sizeof(s->planar_samples[0][0]));
556 static int aac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
557 const AVFrame *frame, int *got_packet_ptr)
559 AACEncContext *s = avctx->priv_data;
560 float **samples = s->planar_samples, *samples2, *la, *overlap;
562 SingleChannelElement *sce;
563 IndividualChannelStream *ics;
564 int i, its, ch, w, chans, tag, start_ch, ret, frame_bits;
565 int target_bits, rate_bits, too_many_bits, too_few_bits;
566 int ms_mode = 0, is_mode = 0, tns_mode = 0, pred_mode = 0;
567 int chan_el_counter[4];
568 FFPsyWindowInfo windows[AAC_MAX_CHANNELS];
570 /* add current frame to queue */
572 if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
575 if (!s->afq.remaining_samples || (!s->afq.frame_alloc && !s->afq.frame_count))
579 copy_input_samples(s, frame);
581 ff_psy_preprocess(s->psypp, s->planar_samples, s->channels);
583 if (!avctx->frame_number)
587 for (i = 0; i < s->chan_map[0]; i++) {
588 FFPsyWindowInfo* wi = windows + start_ch;
589 tag = s->chan_map[i+1];
590 chans = tag == TYPE_CPE ? 2 : 1;
592 for (ch = 0; ch < chans; ch++) {
594 float clip_avoidance_factor;
597 s->cur_channel = start_ch + ch;
598 overlap = &samples[s->cur_channel][0];
599 samples2 = overlap + 1024;
600 la = samples2 + (448+64);
603 if (tag == TYPE_LFE) {
604 wi[ch].window_type[0] = wi[ch].window_type[1] = ONLY_LONG_SEQUENCE;
605 wi[ch].window_shape = 0;
606 wi[ch].num_windows = 1;
607 wi[ch].grouping[0] = 1;
608 wi[ch].clipping[0] = 0;
610 /* Only the lowest 12 coefficients are used in a LFE channel.
611 * The expression below results in only the bottom 8 coefficients
612 * being used for 11.025kHz to 16kHz sample rates.
614 ics->num_swb = s->samplerate_index >= 8 ? 1 : 3;
616 wi[ch] = s->psy.model->window(&s->psy, samples2, la, s->cur_channel,
617 ics->window_sequence[0]);
619 ics->window_sequence[1] = ics->window_sequence[0];
620 ics->window_sequence[0] = wi[ch].window_type[0];
621 ics->use_kb_window[1] = ics->use_kb_window[0];
622 ics->use_kb_window[0] = wi[ch].window_shape;
623 ics->num_windows = wi[ch].num_windows;
624 ics->swb_sizes = s->psy.bands [ics->num_windows == 8];
625 ics->num_swb = tag == TYPE_LFE ? ics->num_swb : s->psy.num_bands[ics->num_windows == 8];
626 ics->max_sfb = FFMIN(ics->max_sfb, ics->num_swb);
627 ics->swb_offset = wi[ch].window_type[0] == EIGHT_SHORT_SEQUENCE ?
628 ff_swb_offset_128 [s->samplerate_index]:
629 ff_swb_offset_1024[s->samplerate_index];
630 ics->tns_max_bands = wi[ch].window_type[0] == EIGHT_SHORT_SEQUENCE ?
631 ff_tns_max_bands_128 [s->samplerate_index]:
632 ff_tns_max_bands_1024[s->samplerate_index];
634 for (w = 0; w < ics->num_windows; w++)
635 ics->group_len[w] = wi[ch].grouping[w];
637 /* Calculate input sample maximums and evaluate clipping risk */
638 clip_avoidance_factor = 0.0f;
639 for (w = 0; w < ics->num_windows; w++) {
640 const float *wbuf = overlap + w * 128;
641 const int wlen = 2048 / ics->num_windows;
644 /* mdct input is 2 * output */
645 for (j = 0; j < wlen; j++)
646 max = FFMAX(max, fabsf(wbuf[j]));
647 wi[ch].clipping[w] = max;
649 for (w = 0; w < ics->num_windows; w++) {
650 if (wi[ch].clipping[w] > CLIP_AVOIDANCE_FACTOR) {
651 ics->window_clipping[w] = 1;
652 clip_avoidance_factor = FFMAX(clip_avoidance_factor, wi[ch].clipping[w]);
654 ics->window_clipping[w] = 0;
657 if (clip_avoidance_factor > CLIP_AVOIDANCE_FACTOR) {
658 ics->clip_avoidance_factor = CLIP_AVOIDANCE_FACTOR / clip_avoidance_factor;
660 ics->clip_avoidance_factor = 1.0f;
663 apply_window_and_mdct(s, sce, overlap);
665 if (s->options.ltp && s->coder->update_ltp) {
666 s->coder->update_ltp(s, sce);
667 apply_window[sce->ics.window_sequence[0]](s->fdsp, sce, &sce->ltp_state[0]);
668 s->mdct1024.mdct_calc(&s->mdct1024, sce->lcoeffs, sce->ret_buf);
671 for (k = 0; k < 1024; k++) {
672 if (!(fabs(cpe->ch[ch].coeffs[k]) < 1E16)) { // Ensure headroom for energy calculation
673 av_log(avctx, AV_LOG_ERROR, "Input contains (near) NaN/+-Inf\n");
674 return AVERROR(EINVAL);
677 avoid_clipping(s, sce);
681 if ((ret = ff_alloc_packet2(avctx, avpkt, 8192 * s->channels, 0)) < 0)
683 frame_bits = its = 0;
685 init_put_bits(&s->pb, avpkt->data, avpkt->size);
687 if ((avctx->frame_number & 0xFF)==1 && !(avctx->flags & AV_CODEC_FLAG_BITEXACT))
688 put_bitstream_info(s, LIBAVCODEC_IDENT);
691 memset(chan_el_counter, 0, sizeof(chan_el_counter));
692 for (i = 0; i < s->chan_map[0]; i++) {
693 FFPsyWindowInfo* wi = windows + start_ch;
694 const float *coeffs[2];
695 tag = s->chan_map[i+1];
696 chans = tag == TYPE_CPE ? 2 : 1;
698 cpe->common_window = 0;
699 memset(cpe->is_mask, 0, sizeof(cpe->is_mask));
700 memset(cpe->ms_mask, 0, sizeof(cpe->ms_mask));
701 put_bits(&s->pb, 3, tag);
702 put_bits(&s->pb, 4, chan_el_counter[tag]++);
703 for (ch = 0; ch < chans; ch++) {
705 coeffs[ch] = sce->coeffs;
706 sce->ics.predictor_present = 0;
707 sce->ics.ltp.present = 0;
708 memset(sce->ics.ltp.used, 0, sizeof(sce->ics.ltp.used));
709 memset(sce->ics.prediction_used, 0, sizeof(sce->ics.prediction_used));
710 memset(&sce->tns, 0, sizeof(TemporalNoiseShaping));
711 for (w = 0; w < 128; w++)
712 if (sce->band_type[w] > RESERVED_BT)
713 sce->band_type[w] = 0;
715 s->psy.bitres.alloc = -1;
716 s->psy.bitres.bits = s->last_frame_pb_count / s->channels;
717 s->psy.model->analyze(&s->psy, start_ch, coeffs, wi);
718 if (s->psy.bitres.alloc > 0) {
719 /* Lambda unused here on purpose, we need to take psy's unscaled allocation */
720 target_bits += s->psy.bitres.alloc
721 * (s->lambda / (avctx->global_quality ? avctx->global_quality : 120));
722 s->psy.bitres.alloc /= chans;
725 for (ch = 0; ch < chans; ch++) {
726 s->cur_channel = start_ch + ch;
727 if (s->options.pns && s->coder->mark_pns)
728 s->coder->mark_pns(s, avctx, &cpe->ch[ch]);
729 s->coder->search_for_quantizers(avctx, s, &cpe->ch[ch], s->lambda);
732 && wi[0].window_type[0] == wi[1].window_type[0]
733 && wi[0].window_shape == wi[1].window_shape) {
735 cpe->common_window = 1;
736 for (w = 0; w < wi[0].num_windows; w++) {
737 if (wi[0].grouping[w] != wi[1].grouping[w]) {
738 cpe->common_window = 0;
743 for (ch = 0; ch < chans; ch++) { /* TNS and PNS */
745 s->cur_channel = start_ch + ch;
746 if (s->options.tns && s->coder->search_for_tns)
747 s->coder->search_for_tns(s, sce);
748 if (s->options.tns && s->coder->apply_tns_filt)
749 s->coder->apply_tns_filt(s, sce);
750 if (sce->tns.present)
752 if (s->options.pns && s->coder->search_for_pns)
753 s->coder->search_for_pns(s, avctx, sce);
755 s->cur_channel = start_ch;
756 if (s->options.intensity_stereo) { /* Intensity Stereo */
757 if (s->coder->search_for_is)
758 s->coder->search_for_is(s, avctx, cpe);
759 if (cpe->is_mode) is_mode = 1;
760 apply_intensity_stereo(cpe);
762 if (s->options.pred) { /* Prediction */
763 for (ch = 0; ch < chans; ch++) {
765 s->cur_channel = start_ch + ch;
766 if (s->options.pred && s->coder->search_for_pred)
767 s->coder->search_for_pred(s, sce);
768 if (cpe->ch[ch].ics.predictor_present) pred_mode = 1;
770 if (s->coder->adjust_common_pred)
771 s->coder->adjust_common_pred(s, cpe);
772 for (ch = 0; ch < chans; ch++) {
774 s->cur_channel = start_ch + ch;
775 if (s->options.pred && s->coder->apply_main_pred)
776 s->coder->apply_main_pred(s, sce);
778 s->cur_channel = start_ch;
780 if (s->options.mid_side) { /* Mid/Side stereo */
781 if (s->options.mid_side == -1 && s->coder->search_for_ms)
782 s->coder->search_for_ms(s, cpe);
783 else if (cpe->common_window)
784 memset(cpe->ms_mask, 1, sizeof(cpe->ms_mask));
785 apply_mid_side_stereo(cpe);
787 adjust_frame_information(cpe, chans);
788 if (s->options.ltp) { /* LTP */
789 for (ch = 0; ch < chans; ch++) {
791 s->cur_channel = start_ch + ch;
792 if (s->coder->search_for_ltp)
793 s->coder->search_for_ltp(s, sce, cpe->common_window);
794 if (sce->ics.ltp.present) pred_mode = 1;
796 s->cur_channel = start_ch;
797 if (s->coder->adjust_common_ltp)
798 s->coder->adjust_common_ltp(s, cpe);
801 put_bits(&s->pb, 1, cpe->common_window);
802 if (cpe->common_window) {
803 put_ics_info(s, &cpe->ch[0].ics);
804 if (s->coder->encode_main_pred)
805 s->coder->encode_main_pred(s, &cpe->ch[0]);
806 if (s->coder->encode_ltp_info)
807 s->coder->encode_ltp_info(s, &cpe->ch[0], 1);
808 encode_ms_info(&s->pb, cpe);
809 if (cpe->ms_mode) ms_mode = 1;
812 for (ch = 0; ch < chans; ch++) {
813 s->cur_channel = start_ch + ch;
814 encode_individual_channel(avctx, s, &cpe->ch[ch], cpe->common_window);
819 if (avctx->flags & AV_CODEC_FLAG_QSCALE) {
820 /* When using a constant Q-scale, don't mess with lambda */
824 /* rate control stuff
825 * allow between the nominal bitrate, and what psy's bit reservoir says to target
826 * but drift towards the nominal bitrate always
828 frame_bits = put_bits_count(&s->pb);
829 rate_bits = avctx->bit_rate * 1024 / avctx->sample_rate;
830 rate_bits = FFMIN(rate_bits, 6144 * s->channels - 3);
831 too_many_bits = FFMAX(target_bits, rate_bits);
832 too_many_bits = FFMIN(too_many_bits, 6144 * s->channels - 3);
833 too_few_bits = FFMIN(FFMAX(rate_bits - rate_bits/4, target_bits), too_many_bits);
835 /* When using ABR, be strict (but only for increasing) */
836 too_few_bits = too_few_bits - too_few_bits/8;
837 too_many_bits = too_many_bits + too_many_bits/2;
839 if ( its == 0 /* for steady-state Q-scale tracking */
840 || (its < 5 && (frame_bits < too_few_bits || frame_bits > too_many_bits))
841 || frame_bits >= 6144 * s->channels - 3 )
843 float ratio = ((float)rate_bits) / frame_bits;
845 if (frame_bits >= too_few_bits && frame_bits <= too_many_bits) {
847 * This path is for steady-state Q-scale tracking
848 * When frame bits fall within the stable range, we still need to adjust
849 * lambda to maintain it like so in a stable fashion (large jumps in lambda
850 * create artifacts and should be avoided), but slowly
852 ratio = sqrtf(sqrtf(ratio));
853 ratio = av_clipf(ratio, 0.9f, 1.1f);
855 /* Not so fast though */
856 ratio = sqrtf(ratio);
858 s->lambda = FFMIN(s->lambda * ratio, 65536.f);
860 /* Keep iterating if we must reduce and lambda is in the sky */
861 if (ratio > 0.9f && ratio < 1.1f) {
864 if (is_mode || ms_mode || tns_mode || pred_mode) {
865 for (i = 0; i < s->chan_map[0]; i++) {
866 // Must restore coeffs
867 chans = tag == TYPE_CPE ? 2 : 1;
869 for (ch = 0; ch < chans; ch++)
870 memcpy(cpe->ch[ch].coeffs, cpe->ch[ch].pcoeffs, sizeof(cpe->ch[ch].coeffs));
880 if (s->options.ltp && s->coder->ltp_insert_new_frame)
881 s->coder->ltp_insert_new_frame(s);
883 put_bits(&s->pb, 3, TYPE_END);
884 flush_put_bits(&s->pb);
886 s->last_frame_pb_count = put_bits_count(&s->pb);
888 s->lambda_sum += s->lambda;
891 ff_af_queue_remove(&s->afq, avctx->frame_size, &avpkt->pts,
894 avpkt->size = put_bits_count(&s->pb) >> 3;
899 static av_cold int aac_encode_end(AVCodecContext *avctx)
901 AACEncContext *s = avctx->priv_data;
903 av_log(avctx, AV_LOG_INFO, "Qavg: %.3f\n", s->lambda_sum / s->lambda_count);
905 ff_mdct_end(&s->mdct1024);
906 ff_mdct_end(&s->mdct128);
910 ff_psy_preprocess_end(s->psypp);
911 av_freep(&s->buffer.samples);
914 ff_af_queue_close(&s->afq);
918 static av_cold int dsp_init(AVCodecContext *avctx, AACEncContext *s)
922 s->fdsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);
924 return AVERROR(ENOMEM);
927 ff_kbd_window_init(ff_aac_kbd_long_1024, 4.0, 1024);
928 ff_kbd_window_init(ff_aac_kbd_short_128, 6.0, 128);
929 ff_init_ff_sine_windows(10);
930 ff_init_ff_sine_windows(7);
932 if ((ret = ff_mdct_init(&s->mdct1024, 11, 0, 32768.0)) < 0)
934 if ((ret = ff_mdct_init(&s->mdct128, 8, 0, 32768.0)) < 0)
940 static av_cold int alloc_buffers(AVCodecContext *avctx, AACEncContext *s)
943 FF_ALLOCZ_ARRAY_OR_GOTO(avctx, s->buffer.samples, s->channels, 3 * 1024 * sizeof(s->buffer.samples[0]), alloc_fail);
944 FF_ALLOCZ_ARRAY_OR_GOTO(avctx, s->cpe, s->chan_map[0], sizeof(ChannelElement), alloc_fail);
946 for(ch = 0; ch < s->channels; ch++)
947 s->planar_samples[ch] = s->buffer.samples + 3 * 1024 * ch;
951 return AVERROR(ENOMEM);
954 static av_cold void aac_encode_init_tables(void)
959 static av_cold int aac_encode_init(AVCodecContext *avctx)
961 AACEncContext *s = avctx->priv_data;
963 const uint8_t *sizes[2];
964 uint8_t grouping[AAC_MAX_CHANNELS];
968 s->last_frame_pb_count = 0;
969 avctx->frame_size = 1024;
970 avctx->initial_padding = 1024;
971 s->lambda = avctx->global_quality > 0 ? avctx->global_quality : 120;
973 /* Channel map and unspecified bitrate guessing */
974 s->channels = avctx->channels;
977 for (i = 0; i < FF_ARRAY_ELEMS(aac_normal_chan_layouts); i++) {
978 if (avctx->channel_layout == aac_normal_chan_layouts[i]) {
979 s->needs_pce = s->options.pce;
986 for (i = 0; i < FF_ARRAY_ELEMS(aac_pce_configs); i++)
987 if (avctx->channel_layout == aac_pce_configs[i].layout)
989 av_get_channel_layout_string(buf, sizeof(buf), -1, avctx->channel_layout);
990 ERROR_IF(i == FF_ARRAY_ELEMS(aac_pce_configs), "Unsupported channel layout \"%s\"\n", buf);
991 av_log(avctx, AV_LOG_INFO, "Using a PCE to encode channel layout \"%s\"\n", buf);
992 s->pce = aac_pce_configs[i];
993 s->reorder_map = s->pce.reorder_map;
994 s->chan_map = s->pce.config_map;
996 s->reorder_map = aac_chan_maps[s->channels - 1];
997 s->chan_map = aac_chan_configs[s->channels - 1];
1000 if (!avctx->bit_rate) {
1001 for (i = 1; i <= s->chan_map[0]; i++) {
1002 avctx->bit_rate += s->chan_map[i] == TYPE_CPE ? 128000 : /* Pair */
1003 s->chan_map[i] == TYPE_LFE ? 16000 : /* LFE */
1009 for (i = 0; i < 16; i++)
1010 if (avctx->sample_rate == avpriv_mpeg4audio_sample_rates[i])
1012 s->samplerate_index = i;
1013 ERROR_IF(s->samplerate_index == 16 ||
1014 s->samplerate_index >= ff_aac_swb_size_1024_len ||
1015 s->samplerate_index >= ff_aac_swb_size_128_len,
1016 "Unsupported sample rate %d\n", avctx->sample_rate);
1018 /* Bitrate limiting */
1019 WARN_IF(1024.0 * avctx->bit_rate / avctx->sample_rate > 6144 * s->channels,
1020 "Too many bits %f > %d per frame requested, clamping to max\n",
1021 1024.0 * avctx->bit_rate / avctx->sample_rate,
1022 6144 * s->channels);
1023 avctx->bit_rate = (int64_t)FFMIN(6144 * s->channels / 1024.0 * avctx->sample_rate,
1026 /* Profile and option setting */
1027 avctx->profile = avctx->profile == FF_PROFILE_UNKNOWN ? FF_PROFILE_AAC_LOW :
1029 for (i = 0; i < FF_ARRAY_ELEMS(aacenc_profiles); i++)
1030 if (avctx->profile == aacenc_profiles[i])
1032 if (avctx->profile == FF_PROFILE_MPEG2_AAC_LOW) {
1033 avctx->profile = FF_PROFILE_AAC_LOW;
1034 ERROR_IF(s->options.pred,
1035 "Main prediction unavailable in the \"mpeg2_aac_low\" profile\n");
1036 ERROR_IF(s->options.ltp,
1037 "LTP prediction unavailable in the \"mpeg2_aac_low\" profile\n");
1038 WARN_IF(s->options.pns,
1039 "PNS unavailable in the \"mpeg2_aac_low\" profile, turning off\n");
1041 } else if (avctx->profile == FF_PROFILE_AAC_LTP) {
1043 ERROR_IF(s->options.pred,
1044 "Main prediction unavailable in the \"aac_ltp\" profile\n");
1045 } else if (avctx->profile == FF_PROFILE_AAC_MAIN) {
1046 s->options.pred = 1;
1047 ERROR_IF(s->options.ltp,
1048 "LTP prediction unavailable in the \"aac_main\" profile\n");
1049 } else if (s->options.ltp) {
1050 avctx->profile = FF_PROFILE_AAC_LTP;
1052 "Chainging profile to \"aac_ltp\"\n");
1053 ERROR_IF(s->options.pred,
1054 "Main prediction unavailable in the \"aac_ltp\" profile\n");
1055 } else if (s->options.pred) {
1056 avctx->profile = FF_PROFILE_AAC_MAIN;
1058 "Chainging profile to \"aac_main\"\n");
1059 ERROR_IF(s->options.ltp,
1060 "LTP prediction unavailable in the \"aac_main\" profile\n");
1062 s->profile = avctx->profile;
1064 /* Coder limitations */
1065 s->coder = &ff_aac_coders[s->options.coder];
1066 if (s->options.coder == AAC_CODER_ANMR) {
1067 ERROR_IF(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL,
1068 "The ANMR coder is considered experimental, add -strict -2 to enable!\n");
1069 s->options.intensity_stereo = 0;
1072 ERROR_IF(s->options.ltp && avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL,
1073 "The LPT profile requires experimental compliance, add -strict -2 to enable!\n");
1075 /* M/S introduces horrible artifacts with multichannel files, this is temporary */
1076 if (s->channels > 3)
1077 s->options.mid_side = 0;
1079 if ((ret = dsp_init(avctx, s)) < 0)
1082 if ((ret = alloc_buffers(avctx, s)) < 0)
1085 if ((ret = put_audio_specific_config(avctx)))
1088 sizes[0] = ff_aac_swb_size_1024[s->samplerate_index];
1089 sizes[1] = ff_aac_swb_size_128[s->samplerate_index];
1090 lengths[0] = ff_aac_num_swb_1024[s->samplerate_index];
1091 lengths[1] = ff_aac_num_swb_128[s->samplerate_index];
1092 for (i = 0; i < s->chan_map[0]; i++)
1093 grouping[i] = s->chan_map[i + 1] == TYPE_CPE;
1094 if ((ret = ff_psy_init(&s->psy, avctx, 2, sizes, lengths,
1095 s->chan_map[0], grouping)) < 0)
1097 s->psypp = ff_psy_preprocess_init(avctx);
1098 ff_lpc_init(&s->lpc, 2*avctx->frame_size, TNS_MAX_ORDER, FF_LPC_TYPE_LEVINSON);
1099 s->random_state = 0x1f2e3d4c;
1101 s->abs_pow34 = abs_pow34_v;
1102 s->quant_bands = quantize_bands;
1105 ff_aac_dsp_init_x86(s);
1108 ff_aac_coder_init_mips(s);
1110 if ((ret = ff_thread_once(&aac_table_init, &aac_encode_init_tables)) != 0)
1111 return AVERROR_UNKNOWN;
1113 ff_af_queue_init(avctx, &s->afq);
1117 aac_encode_end(avctx);
1121 #define AACENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
1122 static const AVOption aacenc_options[] = {
1123 {"aac_coder", "Coding algorithm", offsetof(AACEncContext, options.coder), AV_OPT_TYPE_INT, {.i64 = AAC_CODER_FAST}, 0, AAC_CODER_NB-1, AACENC_FLAGS, "coder"},
1124 {"anmr", "ANMR method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_ANMR}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
1125 {"twoloop", "Two loop searching method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_TWOLOOP}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
1126 {"fast", "Default fast search", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_FAST}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
1127 {"aac_ms", "Force M/S stereo coding", offsetof(AACEncContext, options.mid_side), AV_OPT_TYPE_BOOL, {.i64 = -1}, -1, 1, AACENC_FLAGS},
1128 {"aac_is", "Intensity stereo coding", offsetof(AACEncContext, options.intensity_stereo), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1129 {"aac_pns", "Perceptual noise substitution", offsetof(AACEncContext, options.pns), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1130 {"aac_tns", "Temporal noise shaping", offsetof(AACEncContext, options.tns), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1131 {"aac_ltp", "Long term prediction", offsetof(AACEncContext, options.ltp), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
1132 {"aac_pred", "AAC-Main prediction", offsetof(AACEncContext, options.pred), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
1133 {"aac_pce", "Forces the use of PCEs", offsetof(AACEncContext, options.pce), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
1137 static const AVClass aacenc_class = {
1138 .class_name = "AAC encoder",
1139 .item_name = av_default_item_name,
1140 .option = aacenc_options,
1141 .version = LIBAVUTIL_VERSION_INT,
1144 static const AVCodecDefault aac_encode_defaults[] = {
1149 AVCodec ff_aac_encoder = {
1151 .long_name = NULL_IF_CONFIG_SMALL("AAC (Advanced Audio Coding)"),
1152 .type = AVMEDIA_TYPE_AUDIO,
1153 .id = AV_CODEC_ID_AAC,
1154 .priv_data_size = sizeof(AACEncContext),
1155 .init = aac_encode_init,
1156 .encode2 = aac_encode_frame,
1157 .close = aac_encode_end,
1158 .defaults = aac_encode_defaults,
1159 .supported_samplerates = mpeg4audio_sample_rates,
1160 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE,
1161 .capabilities = AV_CODEC_CAP_SMALL_LAST_FRAME | AV_CODEC_CAP_DELAY,
1162 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
1163 AV_SAMPLE_FMT_NONE },
1164 .priv_class = &aacenc_class,