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/float_dsp.h"
34 #include "libavutil/opt.h"
38 #include "mpeg4audio.h"
45 #include "aacenctab.h"
46 #include "aacenc_utils.h"
50 static void put_pce(PutBitContext *pb, AVCodecContext *avctx)
53 AACEncContext *s = avctx->priv_data;
54 AACPCEInfo *pce = &s->pce;
55 const int bitexact = avctx->flags & AV_CODEC_FLAG_BITEXACT;
56 const char *aux_data = bitexact ? "Lavc" : LIBAVCODEC_IDENT;
60 put_bits(pb, 2, avctx->profile);
61 put_bits(pb, 4, s->samplerate_index);
63 put_bits(pb, 4, pce->num_ele[0]); /* Front */
64 put_bits(pb, 4, pce->num_ele[1]); /* Side */
65 put_bits(pb, 4, pce->num_ele[2]); /* Back */
66 put_bits(pb, 2, pce->num_ele[3]); /* LFE */
67 put_bits(pb, 3, 0); /* Assoc data */
68 put_bits(pb, 4, 0); /* CCs */
70 put_bits(pb, 1, 0); /* Stereo mixdown */
71 put_bits(pb, 1, 0); /* Mono mixdown */
72 put_bits(pb, 1, 0); /* Something else */
74 for (i = 0; i < 4; i++) {
75 for (j = 0; j < pce->num_ele[i]; j++) {
77 put_bits(pb, 1, pce->pairing[i][j]);
78 put_bits(pb, 4, pce->index[i][j]);
83 put_bits(pb, 8, strlen(aux_data));
84 ff_put_string(pb, aux_data, 0);
88 * Make AAC audio config object.
89 * @see 1.6.2.1 "Syntax - AudioSpecificConfig"
91 static int put_audio_specific_config(AVCodecContext *avctx)
94 AACEncContext *s = avctx->priv_data;
95 int channels = (!s->needs_pce)*(s->channels - (s->channels == 8 ? 1 : 0));
96 const int max_size = 32;
98 avctx->extradata = av_mallocz(max_size);
99 if (!avctx->extradata)
100 return AVERROR(ENOMEM);
102 init_put_bits(&pb, avctx->extradata, max_size);
103 put_bits(&pb, 5, s->profile+1); //profile
104 put_bits(&pb, 4, s->samplerate_index); //sample rate index
105 put_bits(&pb, 4, channels);
107 put_bits(&pb, 1, 0); //frame length - 1024 samples
108 put_bits(&pb, 1, 0); //does not depend on core coder
109 put_bits(&pb, 1, 0); //is not extension
113 //Explicitly Mark SBR absent
114 put_bits(&pb, 11, 0x2b7); //sync extension
115 put_bits(&pb, 5, AOT_SBR);
118 avctx->extradata_size = put_bytes_output(&pb);
123 void ff_quantize_band_cost_cache_init(struct AACEncContext *s)
125 ++s->quantize_band_cost_cache_generation;
126 if (s->quantize_band_cost_cache_generation == 0) {
127 memset(s->quantize_band_cost_cache, 0, sizeof(s->quantize_band_cost_cache));
128 s->quantize_band_cost_cache_generation = 1;
132 #define WINDOW_FUNC(type) \
133 static void apply_ ##type ##_window(AVFloatDSPContext *fdsp, \
134 SingleChannelElement *sce, \
137 WINDOW_FUNC(only_long)
139 const float *lwindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
140 const float *pwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
141 float *out = sce->ret_buf;
143 fdsp->vector_fmul (out, audio, lwindow, 1024);
144 fdsp->vector_fmul_reverse(out + 1024, audio + 1024, pwindow, 1024);
147 WINDOW_FUNC(long_start)
149 const float *lwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
150 const float *swindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
151 float *out = sce->ret_buf;
153 fdsp->vector_fmul(out, audio, lwindow, 1024);
154 memcpy(out + 1024, audio + 1024, sizeof(out[0]) * 448);
155 fdsp->vector_fmul_reverse(out + 1024 + 448, audio + 1024 + 448, swindow, 128);
156 memset(out + 1024 + 576, 0, sizeof(out[0]) * 448);
159 WINDOW_FUNC(long_stop)
161 const float *lwindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
162 const float *swindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;
163 float *out = sce->ret_buf;
165 memset(out, 0, sizeof(out[0]) * 448);
166 fdsp->vector_fmul(out + 448, audio + 448, swindow, 128);
167 memcpy(out + 576, audio + 576, sizeof(out[0]) * 448);
168 fdsp->vector_fmul_reverse(out + 1024, audio + 1024, lwindow, 1024);
171 WINDOW_FUNC(eight_short)
173 const float *swindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
174 const float *pwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;
175 const float *in = audio + 448;
176 float *out = sce->ret_buf;
179 for (w = 0; w < 8; w++) {
180 fdsp->vector_fmul (out, in, w ? pwindow : swindow, 128);
183 fdsp->vector_fmul_reverse(out, in, swindow, 128);
188 static void (*const apply_window[4])(AVFloatDSPContext *fdsp,
189 SingleChannelElement *sce,
190 const float *audio) = {
191 [ONLY_LONG_SEQUENCE] = apply_only_long_window,
192 [LONG_START_SEQUENCE] = apply_long_start_window,
193 [EIGHT_SHORT_SEQUENCE] = apply_eight_short_window,
194 [LONG_STOP_SEQUENCE] = apply_long_stop_window
197 static void apply_window_and_mdct(AACEncContext *s, SingleChannelElement *sce,
201 const float *output = sce->ret_buf;
203 apply_window[sce->ics.window_sequence[0]](s->fdsp, sce, audio);
205 if (sce->ics.window_sequence[0] != EIGHT_SHORT_SEQUENCE)
206 s->mdct1024.mdct_calc(&s->mdct1024, sce->coeffs, output);
208 for (i = 0; i < 1024; i += 128)
209 s->mdct128.mdct_calc(&s->mdct128, &sce->coeffs[i], output + i*2);
210 memcpy(audio, audio + 1024, sizeof(audio[0]) * 1024);
211 memcpy(sce->pcoeffs, sce->coeffs, sizeof(sce->pcoeffs));
215 * Encode ics_info element.
216 * @see Table 4.6 (syntax of ics_info)
218 static void put_ics_info(AACEncContext *s, IndividualChannelStream *info)
222 put_bits(&s->pb, 1, 0); // ics_reserved bit
223 put_bits(&s->pb, 2, info->window_sequence[0]);
224 put_bits(&s->pb, 1, info->use_kb_window[0]);
225 if (info->window_sequence[0] != EIGHT_SHORT_SEQUENCE) {
226 put_bits(&s->pb, 6, info->max_sfb);
227 put_bits(&s->pb, 1, !!info->predictor_present);
229 put_bits(&s->pb, 4, info->max_sfb);
230 for (w = 1; w < 8; w++)
231 put_bits(&s->pb, 1, !info->group_len[w]);
237 * @see 4.6.8.1 "Joint Coding - M/S Stereo"
239 static void encode_ms_info(PutBitContext *pb, ChannelElement *cpe)
243 put_bits(pb, 2, cpe->ms_mode);
244 if (cpe->ms_mode == 1)
245 for (w = 0; w < cpe->ch[0].ics.num_windows; w += cpe->ch[0].ics.group_len[w])
246 for (i = 0; i < cpe->ch[0].ics.max_sfb; i++)
247 put_bits(pb, 1, cpe->ms_mask[w*16 + i]);
251 * Produce integer coefficients from scalefactors provided by the model.
253 static void adjust_frame_information(ChannelElement *cpe, int chans)
258 for (ch = 0; ch < chans; ch++) {
259 IndividualChannelStream *ics = &cpe->ch[ch].ics;
261 cpe->ch[ch].pulse.num_pulse = 0;
262 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
263 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
264 for (cmaxsfb = ics->num_swb; cmaxsfb > 0 && cpe->ch[ch].zeroes[w*16+cmaxsfb-1]; cmaxsfb--)
266 maxsfb = FFMAX(maxsfb, cmaxsfb);
269 ics->max_sfb = maxsfb;
271 //adjust zero bands for window groups
272 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
273 for (g = 0; g < ics->max_sfb; g++) {
275 for (w2 = w; w2 < w + ics->group_len[w]; w2++) {
276 if (!cpe->ch[ch].zeroes[w2*16 + g]) {
281 cpe->ch[ch].zeroes[w*16 + g] = i;
286 if (chans > 1 && cpe->common_window) {
287 IndividualChannelStream *ics0 = &cpe->ch[0].ics;
288 IndividualChannelStream *ics1 = &cpe->ch[1].ics;
290 ics0->max_sfb = FFMAX(ics0->max_sfb, ics1->max_sfb);
291 ics1->max_sfb = ics0->max_sfb;
292 for (w = 0; w < ics0->num_windows*16; w += 16)
293 for (i = 0; i < ics0->max_sfb; i++)
294 if (cpe->ms_mask[w+i])
296 if (msc == 0 || ics0->max_sfb == 0)
299 cpe->ms_mode = msc < ics0->max_sfb * ics0->num_windows ? 1 : 2;
303 static void apply_intensity_stereo(ChannelElement *cpe)
306 IndividualChannelStream *ics = &cpe->ch[0].ics;
307 if (!cpe->common_window)
309 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
310 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
311 int start = (w+w2) * 128;
312 for (g = 0; g < ics->num_swb; g++) {
313 int p = -1 + 2 * (cpe->ch[1].band_type[w*16+g] - 14);
314 float scale = cpe->ch[0].is_ener[w*16+g];
315 if (!cpe->is_mask[w*16 + g]) {
316 start += ics->swb_sizes[g];
319 if (cpe->ms_mask[w*16 + g])
321 for (i = 0; i < ics->swb_sizes[g]; i++) {
322 float sum = (cpe->ch[0].coeffs[start+i] + p*cpe->ch[1].coeffs[start+i])*scale;
323 cpe->ch[0].coeffs[start+i] = sum;
324 cpe->ch[1].coeffs[start+i] = 0.0f;
326 start += ics->swb_sizes[g];
332 static void apply_mid_side_stereo(ChannelElement *cpe)
335 IndividualChannelStream *ics = &cpe->ch[0].ics;
336 if (!cpe->common_window)
338 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
339 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
340 int start = (w+w2) * 128;
341 for (g = 0; g < ics->num_swb; g++) {
342 /* ms_mask can be used for other purposes in PNS and I/S,
343 * so must not apply M/S if any band uses either, even if
346 if (!cpe->ms_mask[w*16 + g] || cpe->is_mask[w*16 + g]
347 || cpe->ch[0].band_type[w*16 + g] >= NOISE_BT
348 || cpe->ch[1].band_type[w*16 + g] >= NOISE_BT) {
349 start += ics->swb_sizes[g];
352 for (i = 0; i < ics->swb_sizes[g]; i++) {
353 float L = (cpe->ch[0].coeffs[start+i] + cpe->ch[1].coeffs[start+i]) * 0.5f;
354 float R = L - cpe->ch[1].coeffs[start+i];
355 cpe->ch[0].coeffs[start+i] = L;
356 cpe->ch[1].coeffs[start+i] = R;
358 start += ics->swb_sizes[g];
365 * Encode scalefactor band coding type.
367 static void encode_band_info(AACEncContext *s, SingleChannelElement *sce)
371 if (s->coder->set_special_band_scalefactors)
372 s->coder->set_special_band_scalefactors(s, sce);
374 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
375 s->coder->encode_window_bands_info(s, sce, w, sce->ics.group_len[w], s->lambda);
379 * Encode scalefactors.
381 static void encode_scale_factors(AVCodecContext *avctx, AACEncContext *s,
382 SingleChannelElement *sce)
384 int diff, off_sf = sce->sf_idx[0], off_pns = sce->sf_idx[0] - NOISE_OFFSET;
385 int off_is = 0, noise_flag = 1;
388 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 if (sce->band_type[w*16 + i] == NOISE_BT) {
392 diff = sce->sf_idx[w*16 + i] - off_pns;
393 off_pns = sce->sf_idx[w*16 + i];
394 if (noise_flag-- > 0) {
395 put_bits(&s->pb, NOISE_PRE_BITS, diff + NOISE_PRE);
398 } else if (sce->band_type[w*16 + i] == INTENSITY_BT ||
399 sce->band_type[w*16 + i] == INTENSITY_BT2) {
400 diff = sce->sf_idx[w*16 + i] - off_is;
401 off_is = sce->sf_idx[w*16 + i];
403 diff = sce->sf_idx[w*16 + i] - off_sf;
404 off_sf = sce->sf_idx[w*16 + i];
406 diff += SCALE_DIFF_ZERO;
407 av_assert0(diff >= 0 && diff <= 120);
408 put_bits(&s->pb, ff_aac_scalefactor_bits[diff], ff_aac_scalefactor_code[diff]);
417 static void encode_pulses(AACEncContext *s, Pulse *pulse)
421 put_bits(&s->pb, 1, !!pulse->num_pulse);
422 if (!pulse->num_pulse)
425 put_bits(&s->pb, 2, pulse->num_pulse - 1);
426 put_bits(&s->pb, 6, pulse->start);
427 for (i = 0; i < pulse->num_pulse; i++) {
428 put_bits(&s->pb, 5, pulse->pos[i]);
429 put_bits(&s->pb, 4, pulse->amp[i]);
434 * Encode spectral coefficients processed by psychoacoustic model.
436 static void encode_spectral_coeffs(AACEncContext *s, SingleChannelElement *sce)
440 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
442 for (i = 0; i < sce->ics.max_sfb; i++) {
443 if (sce->zeroes[w*16 + i]) {
444 start += sce->ics.swb_sizes[i];
447 for (w2 = w; w2 < w + sce->ics.group_len[w]; w2++) {
448 s->coder->quantize_and_encode_band(s, &s->pb,
449 &sce->coeffs[start + w2*128],
450 NULL, sce->ics.swb_sizes[i],
451 sce->sf_idx[w*16 + i],
452 sce->band_type[w*16 + i],
454 sce->ics.window_clipping[w]);
456 start += sce->ics.swb_sizes[i];
462 * Downscale spectral coefficients for near-clipping windows to avoid artifacts
464 static void avoid_clipping(AACEncContext *s, SingleChannelElement *sce)
468 if (sce->ics.clip_avoidance_factor < 1.0f) {
469 for (w = 0; w < sce->ics.num_windows; w++) {
471 for (i = 0; i < sce->ics.max_sfb; i++) {
472 float *swb_coeffs = &sce->coeffs[start + w*128];
473 for (j = 0; j < sce->ics.swb_sizes[i]; j++)
474 swb_coeffs[j] *= sce->ics.clip_avoidance_factor;
475 start += sce->ics.swb_sizes[i];
482 * Encode one channel of audio data.
484 static int encode_individual_channel(AVCodecContext *avctx, AACEncContext *s,
485 SingleChannelElement *sce,
488 put_bits(&s->pb, 8, sce->sf_idx[0]);
489 if (!common_window) {
490 put_ics_info(s, &sce->ics);
491 if (s->coder->encode_main_pred)
492 s->coder->encode_main_pred(s, sce);
493 if (s->coder->encode_ltp_info)
494 s->coder->encode_ltp_info(s, sce, 0);
496 encode_band_info(s, sce);
497 encode_scale_factors(avctx, s, sce);
498 encode_pulses(s, &sce->pulse);
499 put_bits(&s->pb, 1, !!sce->tns.present);
500 if (s->coder->encode_tns_info)
501 s->coder->encode_tns_info(s, sce);
502 put_bits(&s->pb, 1, 0); //ssr
503 encode_spectral_coeffs(s, sce);
508 * Write some auxiliary information about the created AAC file.
510 static void put_bitstream_info(AACEncContext *s, const char *name)
512 int i, namelen, padbits;
514 namelen = strlen(name) + 2;
515 put_bits(&s->pb, 3, TYPE_FIL);
516 put_bits(&s->pb, 4, FFMIN(namelen, 15));
518 put_bits(&s->pb, 8, namelen - 14);
519 put_bits(&s->pb, 4, 0); //extension type - filler
520 padbits = -put_bits_count(&s->pb) & 7;
521 align_put_bits(&s->pb);
522 for (i = 0; i < namelen - 2; i++)
523 put_bits(&s->pb, 8, name[i]);
524 put_bits(&s->pb, 12 - padbits, 0);
528 * Copy input samples.
529 * Channels are reordered from libavcodec's default order to AAC order.
531 static void copy_input_samples(AACEncContext *s, const AVFrame *frame)
534 int end = 2048 + (frame ? frame->nb_samples : 0);
535 const uint8_t *channel_map = s->reorder_map;
537 /* copy and remap input samples */
538 for (ch = 0; ch < s->channels; ch++) {
539 /* copy last 1024 samples of previous frame to the start of the current frame */
540 memcpy(&s->planar_samples[ch][1024], &s->planar_samples[ch][2048], 1024 * sizeof(s->planar_samples[0][0]));
542 /* copy new samples and zero any remaining samples */
544 memcpy(&s->planar_samples[ch][2048],
545 frame->extended_data[channel_map[ch]],
546 frame->nb_samples * sizeof(s->planar_samples[0][0]));
548 memset(&s->planar_samples[ch][end], 0,
549 (3072 - end) * sizeof(s->planar_samples[0][0]));
553 static int aac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
554 const AVFrame *frame, int *got_packet_ptr)
556 AACEncContext *s = avctx->priv_data;
557 float **samples = s->planar_samples, *samples2, *la, *overlap;
559 SingleChannelElement *sce;
560 IndividualChannelStream *ics;
561 int i, its, ch, w, chans, tag, start_ch, ret, frame_bits;
562 int target_bits, rate_bits, too_many_bits, too_few_bits;
563 int ms_mode = 0, is_mode = 0, tns_mode = 0, pred_mode = 0;
564 int chan_el_counter[4];
565 FFPsyWindowInfo windows[AAC_MAX_CHANNELS];
567 /* add current frame to queue */
569 if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
572 if (!s->afq.remaining_samples || (!s->afq.frame_alloc && !s->afq.frame_count))
576 copy_input_samples(s, frame);
578 ff_psy_preprocess(s->psypp, s->planar_samples, s->channels);
580 if (!avctx->frame_number)
584 for (i = 0; i < s->chan_map[0]; i++) {
585 FFPsyWindowInfo* wi = windows + start_ch;
586 tag = s->chan_map[i+1];
587 chans = tag == TYPE_CPE ? 2 : 1;
589 for (ch = 0; ch < chans; ch++) {
591 float clip_avoidance_factor;
594 s->cur_channel = start_ch + ch;
595 overlap = &samples[s->cur_channel][0];
596 samples2 = overlap + 1024;
597 la = samples2 + (448+64);
600 if (tag == TYPE_LFE) {
601 wi[ch].window_type[0] = wi[ch].window_type[1] = ONLY_LONG_SEQUENCE;
602 wi[ch].window_shape = 0;
603 wi[ch].num_windows = 1;
604 wi[ch].grouping[0] = 1;
605 wi[ch].clipping[0] = 0;
607 /* Only the lowest 12 coefficients are used in a LFE channel.
608 * The expression below results in only the bottom 8 coefficients
609 * being used for 11.025kHz to 16kHz sample rates.
611 ics->num_swb = s->samplerate_index >= 8 ? 1 : 3;
613 wi[ch] = s->psy.model->window(&s->psy, samples2, la, s->cur_channel,
614 ics->window_sequence[0]);
616 ics->window_sequence[1] = ics->window_sequence[0];
617 ics->window_sequence[0] = wi[ch].window_type[0];
618 ics->use_kb_window[1] = ics->use_kb_window[0];
619 ics->use_kb_window[0] = wi[ch].window_shape;
620 ics->num_windows = wi[ch].num_windows;
621 ics->swb_sizes = s->psy.bands [ics->num_windows == 8];
622 ics->num_swb = tag == TYPE_LFE ? ics->num_swb : s->psy.num_bands[ics->num_windows == 8];
623 ics->max_sfb = FFMIN(ics->max_sfb, ics->num_swb);
624 ics->swb_offset = wi[ch].window_type[0] == EIGHT_SHORT_SEQUENCE ?
625 ff_swb_offset_128 [s->samplerate_index]:
626 ff_swb_offset_1024[s->samplerate_index];
627 ics->tns_max_bands = wi[ch].window_type[0] == EIGHT_SHORT_SEQUENCE ?
628 ff_tns_max_bands_128 [s->samplerate_index]:
629 ff_tns_max_bands_1024[s->samplerate_index];
631 for (w = 0; w < ics->num_windows; w++)
632 ics->group_len[w] = wi[ch].grouping[w];
634 /* Calculate input sample maximums and evaluate clipping risk */
635 clip_avoidance_factor = 0.0f;
636 for (w = 0; w < ics->num_windows; w++) {
637 const float *wbuf = overlap + w * 128;
638 const int wlen = 2048 / ics->num_windows;
641 /* mdct input is 2 * output */
642 for (j = 0; j < wlen; j++)
643 max = FFMAX(max, fabsf(wbuf[j]));
644 wi[ch].clipping[w] = max;
646 for (w = 0; w < ics->num_windows; w++) {
647 if (wi[ch].clipping[w] > CLIP_AVOIDANCE_FACTOR) {
648 ics->window_clipping[w] = 1;
649 clip_avoidance_factor = FFMAX(clip_avoidance_factor, wi[ch].clipping[w]);
651 ics->window_clipping[w] = 0;
654 if (clip_avoidance_factor > CLIP_AVOIDANCE_FACTOR) {
655 ics->clip_avoidance_factor = CLIP_AVOIDANCE_FACTOR / clip_avoidance_factor;
657 ics->clip_avoidance_factor = 1.0f;
660 apply_window_and_mdct(s, sce, overlap);
662 if (s->options.ltp && s->coder->update_ltp) {
663 s->coder->update_ltp(s, sce);
664 apply_window[sce->ics.window_sequence[0]](s->fdsp, sce, &sce->ltp_state[0]);
665 s->mdct1024.mdct_calc(&s->mdct1024, sce->lcoeffs, sce->ret_buf);
668 for (k = 0; k < 1024; k++) {
669 if (!(fabs(cpe->ch[ch].coeffs[k]) < 1E16)) { // Ensure headroom for energy calculation
670 av_log(avctx, AV_LOG_ERROR, "Input contains (near) NaN/+-Inf\n");
671 return AVERROR(EINVAL);
674 avoid_clipping(s, sce);
678 if ((ret = ff_alloc_packet2(avctx, avpkt, 8192 * s->channels, 0)) < 0)
680 frame_bits = its = 0;
682 init_put_bits(&s->pb, avpkt->data, avpkt->size);
684 if ((avctx->frame_number & 0xFF)==1 && !(avctx->flags & AV_CODEC_FLAG_BITEXACT))
685 put_bitstream_info(s, LIBAVCODEC_IDENT);
688 memset(chan_el_counter, 0, sizeof(chan_el_counter));
689 for (i = 0; i < s->chan_map[0]; i++) {
690 FFPsyWindowInfo* wi = windows + start_ch;
691 const float *coeffs[2];
692 tag = s->chan_map[i+1];
693 chans = tag == TYPE_CPE ? 2 : 1;
695 cpe->common_window = 0;
696 memset(cpe->is_mask, 0, sizeof(cpe->is_mask));
697 memset(cpe->ms_mask, 0, sizeof(cpe->ms_mask));
698 put_bits(&s->pb, 3, tag);
699 put_bits(&s->pb, 4, chan_el_counter[tag]++);
700 for (ch = 0; ch < chans; ch++) {
702 coeffs[ch] = sce->coeffs;
703 sce->ics.predictor_present = 0;
704 sce->ics.ltp.present = 0;
705 memset(sce->ics.ltp.used, 0, sizeof(sce->ics.ltp.used));
706 memset(sce->ics.prediction_used, 0, sizeof(sce->ics.prediction_used));
707 memset(&sce->tns, 0, sizeof(TemporalNoiseShaping));
708 for (w = 0; w < 128; w++)
709 if (sce->band_type[w] > RESERVED_BT)
710 sce->band_type[w] = 0;
712 s->psy.bitres.alloc = -1;
713 s->psy.bitres.bits = s->last_frame_pb_count / s->channels;
714 s->psy.model->analyze(&s->psy, start_ch, coeffs, wi);
715 if (s->psy.bitres.alloc > 0) {
716 /* Lambda unused here on purpose, we need to take psy's unscaled allocation */
717 target_bits += s->psy.bitres.alloc
718 * (s->lambda / (avctx->global_quality ? avctx->global_quality : 120));
719 s->psy.bitres.alloc /= chans;
722 for (ch = 0; ch < chans; ch++) {
723 s->cur_channel = start_ch + ch;
724 if (s->options.pns && s->coder->mark_pns)
725 s->coder->mark_pns(s, avctx, &cpe->ch[ch]);
726 s->coder->search_for_quantizers(avctx, s, &cpe->ch[ch], s->lambda);
729 && wi[0].window_type[0] == wi[1].window_type[0]
730 && wi[0].window_shape == wi[1].window_shape) {
732 cpe->common_window = 1;
733 for (w = 0; w < wi[0].num_windows; w++) {
734 if (wi[0].grouping[w] != wi[1].grouping[w]) {
735 cpe->common_window = 0;
740 for (ch = 0; ch < chans; ch++) { /* TNS and PNS */
742 s->cur_channel = start_ch + ch;
743 if (s->options.tns && s->coder->search_for_tns)
744 s->coder->search_for_tns(s, sce);
745 if (s->options.tns && s->coder->apply_tns_filt)
746 s->coder->apply_tns_filt(s, sce);
747 if (sce->tns.present)
749 if (s->options.pns && s->coder->search_for_pns)
750 s->coder->search_for_pns(s, avctx, sce);
752 s->cur_channel = start_ch;
753 if (s->options.intensity_stereo) { /* Intensity Stereo */
754 if (s->coder->search_for_is)
755 s->coder->search_for_is(s, avctx, cpe);
756 if (cpe->is_mode) is_mode = 1;
757 apply_intensity_stereo(cpe);
759 if (s->options.pred) { /* Prediction */
760 for (ch = 0; ch < chans; ch++) {
762 s->cur_channel = start_ch + ch;
763 if (s->options.pred && s->coder->search_for_pred)
764 s->coder->search_for_pred(s, sce);
765 if (cpe->ch[ch].ics.predictor_present) pred_mode = 1;
767 if (s->coder->adjust_common_pred)
768 s->coder->adjust_common_pred(s, cpe);
769 for (ch = 0; ch < chans; ch++) {
771 s->cur_channel = start_ch + ch;
772 if (s->options.pred && s->coder->apply_main_pred)
773 s->coder->apply_main_pred(s, sce);
775 s->cur_channel = start_ch;
777 if (s->options.mid_side) { /* Mid/Side stereo */
778 if (s->options.mid_side == -1 && s->coder->search_for_ms)
779 s->coder->search_for_ms(s, cpe);
780 else if (cpe->common_window)
781 memset(cpe->ms_mask, 1, sizeof(cpe->ms_mask));
782 apply_mid_side_stereo(cpe);
784 adjust_frame_information(cpe, chans);
785 if (s->options.ltp) { /* LTP */
786 for (ch = 0; ch < chans; ch++) {
788 s->cur_channel = start_ch + ch;
789 if (s->coder->search_for_ltp)
790 s->coder->search_for_ltp(s, sce, cpe->common_window);
791 if (sce->ics.ltp.present) pred_mode = 1;
793 s->cur_channel = start_ch;
794 if (s->coder->adjust_common_ltp)
795 s->coder->adjust_common_ltp(s, cpe);
798 put_bits(&s->pb, 1, cpe->common_window);
799 if (cpe->common_window) {
800 put_ics_info(s, &cpe->ch[0].ics);
801 if (s->coder->encode_main_pred)
802 s->coder->encode_main_pred(s, &cpe->ch[0]);
803 if (s->coder->encode_ltp_info)
804 s->coder->encode_ltp_info(s, &cpe->ch[0], 1);
805 encode_ms_info(&s->pb, cpe);
806 if (cpe->ms_mode) ms_mode = 1;
809 for (ch = 0; ch < chans; ch++) {
810 s->cur_channel = start_ch + ch;
811 encode_individual_channel(avctx, s, &cpe->ch[ch], cpe->common_window);
816 if (avctx->flags & AV_CODEC_FLAG_QSCALE) {
817 /* When using a constant Q-scale, don't mess with lambda */
821 /* rate control stuff
822 * allow between the nominal bitrate, and what psy's bit reservoir says to target
823 * but drift towards the nominal bitrate always
825 frame_bits = put_bits_count(&s->pb);
826 rate_bits = avctx->bit_rate * 1024 / avctx->sample_rate;
827 rate_bits = FFMIN(rate_bits, 6144 * s->channels - 3);
828 too_many_bits = FFMAX(target_bits, rate_bits);
829 too_many_bits = FFMIN(too_many_bits, 6144 * s->channels - 3);
830 too_few_bits = FFMIN(FFMAX(rate_bits - rate_bits/4, target_bits), too_many_bits);
832 /* When using ABR, be strict (but only for increasing) */
833 too_few_bits = too_few_bits - too_few_bits/8;
834 too_many_bits = too_many_bits + too_many_bits/2;
836 if ( its == 0 /* for steady-state Q-scale tracking */
837 || (its < 5 && (frame_bits < too_few_bits || frame_bits > too_many_bits))
838 || frame_bits >= 6144 * s->channels - 3 )
840 float ratio = ((float)rate_bits) / frame_bits;
842 if (frame_bits >= too_few_bits && frame_bits <= too_many_bits) {
844 * This path is for steady-state Q-scale tracking
845 * When frame bits fall within the stable range, we still need to adjust
846 * lambda to maintain it like so in a stable fashion (large jumps in lambda
847 * create artifacts and should be avoided), but slowly
849 ratio = sqrtf(sqrtf(ratio));
850 ratio = av_clipf(ratio, 0.9f, 1.1f);
852 /* Not so fast though */
853 ratio = sqrtf(ratio);
855 s->lambda = FFMIN(s->lambda * ratio, 65536.f);
857 /* Keep iterating if we must reduce and lambda is in the sky */
858 if (ratio > 0.9f && ratio < 1.1f) {
861 if (is_mode || ms_mode || tns_mode || pred_mode) {
862 for (i = 0; i < s->chan_map[0]; i++) {
863 // Must restore coeffs
864 chans = tag == TYPE_CPE ? 2 : 1;
866 for (ch = 0; ch < chans; ch++)
867 memcpy(cpe->ch[ch].coeffs, cpe->ch[ch].pcoeffs, sizeof(cpe->ch[ch].coeffs));
877 if (s->options.ltp && s->coder->ltp_insert_new_frame)
878 s->coder->ltp_insert_new_frame(s);
880 put_bits(&s->pb, 3, TYPE_END);
881 flush_put_bits(&s->pb);
883 s->last_frame_pb_count = put_bits_count(&s->pb);
884 avpkt->size = put_bytes_output(&s->pb);
886 s->lambda_sum += s->lambda;
889 ff_af_queue_remove(&s->afq, avctx->frame_size, &avpkt->pts,
896 static av_cold int aac_encode_end(AVCodecContext *avctx)
898 AACEncContext *s = avctx->priv_data;
900 av_log(avctx, AV_LOG_INFO, "Qavg: %.3f\n", s->lambda_sum / s->lambda_count);
902 ff_mdct_end(&s->mdct1024);
903 ff_mdct_end(&s->mdct128);
907 ff_psy_preprocess_end(s->psypp);
908 av_freep(&s->buffer.samples);
911 ff_af_queue_close(&s->afq);
915 static av_cold int dsp_init(AVCodecContext *avctx, AACEncContext *s)
919 s->fdsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);
921 return AVERROR(ENOMEM);
924 ff_aac_float_common_init();
926 if ((ret = ff_mdct_init(&s->mdct1024, 11, 0, 32768.0)) < 0)
928 if ((ret = ff_mdct_init(&s->mdct128, 8, 0, 32768.0)) < 0)
934 static av_cold int alloc_buffers(AVCodecContext *avctx, AACEncContext *s)
937 if (!FF_ALLOCZ_TYPED_ARRAY(s->buffer.samples, s->channels * 3 * 1024) ||
938 !FF_ALLOCZ_TYPED_ARRAY(s->cpe, s->chan_map[0]))
939 return AVERROR(ENOMEM);
941 for(ch = 0; ch < s->channels; ch++)
942 s->planar_samples[ch] = s->buffer.samples + 3 * 1024 * ch;
947 static av_cold int aac_encode_init(AVCodecContext *avctx)
949 AACEncContext *s = avctx->priv_data;
951 const uint8_t *sizes[2];
952 uint8_t grouping[AAC_MAX_CHANNELS];
956 s->last_frame_pb_count = 0;
957 avctx->frame_size = 1024;
958 avctx->initial_padding = 1024;
959 s->lambda = avctx->global_quality > 0 ? avctx->global_quality : 120;
961 /* Channel map and unspecified bitrate guessing */
962 s->channels = avctx->channels;
965 for (i = 0; i < FF_ARRAY_ELEMS(aac_normal_chan_layouts); i++) {
966 if (avctx->channel_layout == aac_normal_chan_layouts[i]) {
967 s->needs_pce = s->options.pce;
974 for (i = 0; i < FF_ARRAY_ELEMS(aac_pce_configs); i++)
975 if (avctx->channel_layout == aac_pce_configs[i].layout)
977 av_get_channel_layout_string(buf, sizeof(buf), -1, avctx->channel_layout);
978 ERROR_IF(i == FF_ARRAY_ELEMS(aac_pce_configs), "Unsupported channel layout \"%s\"\n", buf);
979 av_log(avctx, AV_LOG_INFO, "Using a PCE to encode channel layout \"%s\"\n", buf);
980 s->pce = aac_pce_configs[i];
981 s->reorder_map = s->pce.reorder_map;
982 s->chan_map = s->pce.config_map;
984 s->reorder_map = aac_chan_maps[s->channels - 1];
985 s->chan_map = aac_chan_configs[s->channels - 1];
988 if (!avctx->bit_rate) {
989 for (i = 1; i <= s->chan_map[0]; i++) {
990 avctx->bit_rate += s->chan_map[i] == TYPE_CPE ? 128000 : /* Pair */
991 s->chan_map[i] == TYPE_LFE ? 16000 : /* LFE */
997 for (i = 0; i < 16; i++)
998 if (avctx->sample_rate == avpriv_mpeg4audio_sample_rates[i])
1000 s->samplerate_index = i;
1001 ERROR_IF(s->samplerate_index == 16 ||
1002 s->samplerate_index >= ff_aac_swb_size_1024_len ||
1003 s->samplerate_index >= ff_aac_swb_size_128_len,
1004 "Unsupported sample rate %d\n", avctx->sample_rate);
1006 /* Bitrate limiting */
1007 WARN_IF(1024.0 * avctx->bit_rate / avctx->sample_rate > 6144 * s->channels,
1008 "Too many bits %f > %d per frame requested, clamping to max\n",
1009 1024.0 * avctx->bit_rate / avctx->sample_rate,
1010 6144 * s->channels);
1011 avctx->bit_rate = (int64_t)FFMIN(6144 * s->channels / 1024.0 * avctx->sample_rate,
1014 /* Profile and option setting */
1015 avctx->profile = avctx->profile == FF_PROFILE_UNKNOWN ? FF_PROFILE_AAC_LOW :
1017 for (i = 0; i < FF_ARRAY_ELEMS(aacenc_profiles); i++)
1018 if (avctx->profile == aacenc_profiles[i])
1020 if (avctx->profile == FF_PROFILE_MPEG2_AAC_LOW) {
1021 avctx->profile = FF_PROFILE_AAC_LOW;
1022 ERROR_IF(s->options.pred,
1023 "Main prediction unavailable in the \"mpeg2_aac_low\" profile\n");
1024 ERROR_IF(s->options.ltp,
1025 "LTP prediction unavailable in the \"mpeg2_aac_low\" profile\n");
1026 WARN_IF(s->options.pns,
1027 "PNS unavailable in the \"mpeg2_aac_low\" profile, turning off\n");
1029 } else if (avctx->profile == FF_PROFILE_AAC_LTP) {
1031 ERROR_IF(s->options.pred,
1032 "Main prediction unavailable in the \"aac_ltp\" profile\n");
1033 } else if (avctx->profile == FF_PROFILE_AAC_MAIN) {
1034 s->options.pred = 1;
1035 ERROR_IF(s->options.ltp,
1036 "LTP prediction unavailable in the \"aac_main\" profile\n");
1037 } else if (s->options.ltp) {
1038 avctx->profile = FF_PROFILE_AAC_LTP;
1040 "Chainging profile to \"aac_ltp\"\n");
1041 ERROR_IF(s->options.pred,
1042 "Main prediction unavailable in the \"aac_ltp\" profile\n");
1043 } else if (s->options.pred) {
1044 avctx->profile = FF_PROFILE_AAC_MAIN;
1046 "Chainging profile to \"aac_main\"\n");
1047 ERROR_IF(s->options.ltp,
1048 "LTP prediction unavailable in the \"aac_main\" profile\n");
1050 s->profile = avctx->profile;
1052 /* Coder limitations */
1053 s->coder = &ff_aac_coders[s->options.coder];
1054 if (s->options.coder == AAC_CODER_ANMR) {
1055 ERROR_IF(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL,
1056 "The ANMR coder is considered experimental, add -strict -2 to enable!\n");
1057 s->options.intensity_stereo = 0;
1060 ERROR_IF(s->options.ltp && avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL,
1061 "The LPT profile requires experimental compliance, add -strict -2 to enable!\n");
1063 /* M/S introduces horrible artifacts with multichannel files, this is temporary */
1064 if (s->channels > 3)
1065 s->options.mid_side = 0;
1067 if ((ret = dsp_init(avctx, s)) < 0)
1070 if ((ret = alloc_buffers(avctx, s)) < 0)
1073 if ((ret = put_audio_specific_config(avctx)))
1076 sizes[0] = ff_aac_swb_size_1024[s->samplerate_index];
1077 sizes[1] = ff_aac_swb_size_128[s->samplerate_index];
1078 lengths[0] = ff_aac_num_swb_1024[s->samplerate_index];
1079 lengths[1] = ff_aac_num_swb_128[s->samplerate_index];
1080 for (i = 0; i < s->chan_map[0]; i++)
1081 grouping[i] = s->chan_map[i + 1] == TYPE_CPE;
1082 if ((ret = ff_psy_init(&s->psy, avctx, 2, sizes, lengths,
1083 s->chan_map[0], grouping)) < 0)
1085 s->psypp = ff_psy_preprocess_init(avctx);
1086 ff_lpc_init(&s->lpc, 2*avctx->frame_size, TNS_MAX_ORDER, FF_LPC_TYPE_LEVINSON);
1087 s->random_state = 0x1f2e3d4c;
1089 s->abs_pow34 = abs_pow34_v;
1090 s->quant_bands = quantize_bands;
1093 ff_aac_dsp_init_x86(s);
1096 ff_aac_coder_init_mips(s);
1098 ff_af_queue_init(avctx, &s->afq);
1104 #define AACENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
1105 static const AVOption aacenc_options[] = {
1106 {"aac_coder", "Coding algorithm", offsetof(AACEncContext, options.coder), AV_OPT_TYPE_INT, {.i64 = AAC_CODER_FAST}, 0, AAC_CODER_NB-1, AACENC_FLAGS, "coder"},
1107 {"anmr", "ANMR method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_ANMR}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
1108 {"twoloop", "Two loop searching method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_TWOLOOP}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
1109 {"fast", "Default fast search", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_FAST}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
1110 {"aac_ms", "Force M/S stereo coding", offsetof(AACEncContext, options.mid_side), AV_OPT_TYPE_BOOL, {.i64 = -1}, -1, 1, AACENC_FLAGS},
1111 {"aac_is", "Intensity stereo coding", offsetof(AACEncContext, options.intensity_stereo), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1112 {"aac_pns", "Perceptual noise substitution", offsetof(AACEncContext, options.pns), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1113 {"aac_tns", "Temporal noise shaping", offsetof(AACEncContext, options.tns), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1114 {"aac_ltp", "Long term prediction", offsetof(AACEncContext, options.ltp), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
1115 {"aac_pred", "AAC-Main prediction", offsetof(AACEncContext, options.pred), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
1116 {"aac_pce", "Forces the use of PCEs", offsetof(AACEncContext, options.pce), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
1121 static const AVClass aacenc_class = {
1122 .class_name = "AAC encoder",
1123 .item_name = av_default_item_name,
1124 .option = aacenc_options,
1125 .version = LIBAVUTIL_VERSION_INT,
1128 static const AVCodecDefault aac_encode_defaults[] = {
1133 AVCodec ff_aac_encoder = {
1135 .long_name = NULL_IF_CONFIG_SMALL("AAC (Advanced Audio Coding)"),
1136 .type = AVMEDIA_TYPE_AUDIO,
1137 .id = AV_CODEC_ID_AAC,
1138 .priv_data_size = sizeof(AACEncContext),
1139 .init = aac_encode_init,
1140 .encode2 = aac_encode_frame,
1141 .close = aac_encode_end,
1142 .defaults = aac_encode_defaults,
1143 .supported_samplerates = mpeg4audio_sample_rates,
1144 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE | FF_CODEC_CAP_INIT_CLEANUP,
1145 .capabilities = AV_CODEC_CAP_SMALL_LAST_FRAME | AV_CODEC_CAP_DELAY,
1146 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
1147 AV_SAMPLE_FMT_NONE },
1148 .priv_class = &aacenc_class,