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"
46 #include "aacenctab.h"
47 #include "aacenc_utils.h"
51 static void put_pce(PutBitContext *pb, AVCodecContext *avctx)
54 AACEncContext *s = avctx->priv_data;
55 AACPCEInfo *pce = &s->pce;
56 const int bitexact = avctx->flags & AV_CODEC_FLAG_BITEXACT;
57 const char *aux_data = bitexact ? "Lavc" : LIBAVCODEC_IDENT;
61 put_bits(pb, 2, avctx->profile);
62 put_bits(pb, 4, s->samplerate_index);
64 put_bits(pb, 4, pce->num_ele[0]); /* Front */
65 put_bits(pb, 4, pce->num_ele[1]); /* Side */
66 put_bits(pb, 4, pce->num_ele[2]); /* Back */
67 put_bits(pb, 2, pce->num_ele[3]); /* LFE */
68 put_bits(pb, 3, 0); /* Assoc data */
69 put_bits(pb, 4, 0); /* CCs */
71 put_bits(pb, 1, 0); /* Stereo mixdown */
72 put_bits(pb, 1, 0); /* Mono mixdown */
73 put_bits(pb, 1, 0); /* Something else */
75 for (i = 0; i < 4; i++) {
76 for (j = 0; j < pce->num_ele[i]; j++) {
78 put_bits(pb, 1, pce->pairing[i][j]);
79 put_bits(pb, 4, pce->index[i][j]);
84 put_bits(pb, 8, strlen(aux_data));
85 ff_put_string(pb, aux_data, 0);
89 * Make AAC audio config object.
90 * @see 1.6.2.1 "Syntax - AudioSpecificConfig"
92 static int put_audio_specific_config(AVCodecContext *avctx)
95 AACEncContext *s = avctx->priv_data;
96 int channels = (!s->needs_pce)*(s->channels - (s->channels == 8 ? 1 : 0));
97 const int max_size = 32;
99 avctx->extradata = av_mallocz(max_size);
100 if (!avctx->extradata)
101 return AVERROR(ENOMEM);
103 init_put_bits(&pb, avctx->extradata, max_size);
104 put_bits(&pb, 5, s->profile+1); //profile
105 put_bits(&pb, 4, s->samplerate_index); //sample rate index
106 put_bits(&pb, 4, channels);
108 put_bits(&pb, 1, 0); //frame length - 1024 samples
109 put_bits(&pb, 1, 0); //does not depend on core coder
110 put_bits(&pb, 1, 0); //is not extension
114 //Explicitly Mark SBR absent
115 put_bits(&pb, 11, 0x2b7); //sync extension
116 put_bits(&pb, 5, AOT_SBR);
119 avctx->extradata_size = put_bits_count(&pb) >> 3;
124 void ff_quantize_band_cost_cache_init(struct AACEncContext *s)
126 ++s->quantize_band_cost_cache_generation;
127 if (s->quantize_band_cost_cache_generation == 0) {
128 memset(s->quantize_band_cost_cache, 0, sizeof(s->quantize_band_cost_cache));
129 s->quantize_band_cost_cache_generation = 1;
133 #define WINDOW_FUNC(type) \
134 static void apply_ ##type ##_window(AVFloatDSPContext *fdsp, \
135 SingleChannelElement *sce, \
138 WINDOW_FUNC(only_long)
140 const float *lwindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
141 const float *pwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
142 float *out = sce->ret_buf;
144 fdsp->vector_fmul (out, audio, lwindow, 1024);
145 fdsp->vector_fmul_reverse(out + 1024, audio + 1024, pwindow, 1024);
148 WINDOW_FUNC(long_start)
150 const float *lwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
151 const float *swindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
152 float *out = sce->ret_buf;
154 fdsp->vector_fmul(out, audio, lwindow, 1024);
155 memcpy(out + 1024, audio + 1024, sizeof(out[0]) * 448);
156 fdsp->vector_fmul_reverse(out + 1024 + 448, audio + 1024 + 448, swindow, 128);
157 memset(out + 1024 + 576, 0, sizeof(out[0]) * 448);
160 WINDOW_FUNC(long_stop)
162 const float *lwindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
163 const float *swindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;
164 float *out = sce->ret_buf;
166 memset(out, 0, sizeof(out[0]) * 448);
167 fdsp->vector_fmul(out + 448, audio + 448, swindow, 128);
168 memcpy(out + 576, audio + 576, sizeof(out[0]) * 448);
169 fdsp->vector_fmul_reverse(out + 1024, audio + 1024, lwindow, 1024);
172 WINDOW_FUNC(eight_short)
174 const float *swindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
175 const float *pwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;
176 const float *in = audio + 448;
177 float *out = sce->ret_buf;
180 for (w = 0; w < 8; w++) {
181 fdsp->vector_fmul (out, in, w ? pwindow : swindow, 128);
184 fdsp->vector_fmul_reverse(out, in, swindow, 128);
189 static void (*const apply_window[4])(AVFloatDSPContext *fdsp,
190 SingleChannelElement *sce,
191 const float *audio) = {
192 [ONLY_LONG_SEQUENCE] = apply_only_long_window,
193 [LONG_START_SEQUENCE] = apply_long_start_window,
194 [EIGHT_SHORT_SEQUENCE] = apply_eight_short_window,
195 [LONG_STOP_SEQUENCE] = apply_long_stop_window
198 static void apply_window_and_mdct(AACEncContext *s, SingleChannelElement *sce,
202 const float *output = sce->ret_buf;
204 apply_window[sce->ics.window_sequence[0]](s->fdsp, sce, audio);
206 if (sce->ics.window_sequence[0] != EIGHT_SHORT_SEQUENCE)
207 s->mdct1024.mdct_calc(&s->mdct1024, sce->coeffs, output);
209 for (i = 0; i < 1024; i += 128)
210 s->mdct128.mdct_calc(&s->mdct128, &sce->coeffs[i], output + i*2);
211 memcpy(audio, audio + 1024, sizeof(audio[0]) * 1024);
212 memcpy(sce->pcoeffs, sce->coeffs, sizeof(sce->pcoeffs));
216 * Encode ics_info element.
217 * @see Table 4.6 (syntax of ics_info)
219 static void put_ics_info(AACEncContext *s, IndividualChannelStream *info)
223 put_bits(&s->pb, 1, 0); // ics_reserved bit
224 put_bits(&s->pb, 2, info->window_sequence[0]);
225 put_bits(&s->pb, 1, info->use_kb_window[0]);
226 if (info->window_sequence[0] != EIGHT_SHORT_SEQUENCE) {
227 put_bits(&s->pb, 6, info->max_sfb);
228 put_bits(&s->pb, 1, !!info->predictor_present);
230 put_bits(&s->pb, 4, info->max_sfb);
231 for (w = 1; w < 8; w++)
232 put_bits(&s->pb, 1, !info->group_len[w]);
238 * @see 4.6.8.1 "Joint Coding - M/S Stereo"
240 static void encode_ms_info(PutBitContext *pb, ChannelElement *cpe)
244 put_bits(pb, 2, cpe->ms_mode);
245 if (cpe->ms_mode == 1)
246 for (w = 0; w < cpe->ch[0].ics.num_windows; w += cpe->ch[0].ics.group_len[w])
247 for (i = 0; i < cpe->ch[0].ics.max_sfb; i++)
248 put_bits(pb, 1, cpe->ms_mask[w*16 + i]);
252 * Produce integer coefficients from scalefactors provided by the model.
254 static void adjust_frame_information(ChannelElement *cpe, int chans)
259 for (ch = 0; ch < chans; ch++) {
260 IndividualChannelStream *ics = &cpe->ch[ch].ics;
262 cpe->ch[ch].pulse.num_pulse = 0;
263 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
264 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
265 for (cmaxsfb = ics->num_swb; cmaxsfb > 0 && cpe->ch[ch].zeroes[w*16+cmaxsfb-1]; cmaxsfb--)
267 maxsfb = FFMAX(maxsfb, cmaxsfb);
270 ics->max_sfb = maxsfb;
272 //adjust zero bands for window groups
273 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
274 for (g = 0; g < ics->max_sfb; g++) {
276 for (w2 = w; w2 < w + ics->group_len[w]; w2++) {
277 if (!cpe->ch[ch].zeroes[w2*16 + g]) {
282 cpe->ch[ch].zeroes[w*16 + g] = i;
287 if (chans > 1 && cpe->common_window) {
288 IndividualChannelStream *ics0 = &cpe->ch[0].ics;
289 IndividualChannelStream *ics1 = &cpe->ch[1].ics;
291 ics0->max_sfb = FFMAX(ics0->max_sfb, ics1->max_sfb);
292 ics1->max_sfb = ics0->max_sfb;
293 for (w = 0; w < ics0->num_windows*16; w += 16)
294 for (i = 0; i < ics0->max_sfb; i++)
295 if (cpe->ms_mask[w+i])
297 if (msc == 0 || ics0->max_sfb == 0)
300 cpe->ms_mode = msc < ics0->max_sfb * ics0->num_windows ? 1 : 2;
304 static void apply_intensity_stereo(ChannelElement *cpe)
307 IndividualChannelStream *ics = &cpe->ch[0].ics;
308 if (!cpe->common_window)
310 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
311 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
312 int start = (w+w2) * 128;
313 for (g = 0; g < ics->num_swb; g++) {
314 int p = -1 + 2 * (cpe->ch[1].band_type[w*16+g] - 14);
315 float scale = cpe->ch[0].is_ener[w*16+g];
316 if (!cpe->is_mask[w*16 + g]) {
317 start += ics->swb_sizes[g];
320 if (cpe->ms_mask[w*16 + g])
322 for (i = 0; i < ics->swb_sizes[g]; i++) {
323 float sum = (cpe->ch[0].coeffs[start+i] + p*cpe->ch[1].coeffs[start+i])*scale;
324 cpe->ch[0].coeffs[start+i] = sum;
325 cpe->ch[1].coeffs[start+i] = 0.0f;
327 start += ics->swb_sizes[g];
333 static void apply_mid_side_stereo(ChannelElement *cpe)
336 IndividualChannelStream *ics = &cpe->ch[0].ics;
337 if (!cpe->common_window)
339 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
340 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
341 int start = (w+w2) * 128;
342 for (g = 0; g < ics->num_swb; g++) {
343 /* ms_mask can be used for other purposes in PNS and I/S,
344 * so must not apply M/S if any band uses either, even if
347 if (!cpe->ms_mask[w*16 + g] || cpe->is_mask[w*16 + g]
348 || cpe->ch[0].band_type[w*16 + g] >= NOISE_BT
349 || cpe->ch[1].band_type[w*16 + g] >= NOISE_BT) {
350 start += ics->swb_sizes[g];
353 for (i = 0; i < ics->swb_sizes[g]; i++) {
354 float L = (cpe->ch[0].coeffs[start+i] + cpe->ch[1].coeffs[start+i]) * 0.5f;
355 float R = L - cpe->ch[1].coeffs[start+i];
356 cpe->ch[0].coeffs[start+i] = L;
357 cpe->ch[1].coeffs[start+i] = R;
359 start += ics->swb_sizes[g];
366 * Encode scalefactor band coding type.
368 static void encode_band_info(AACEncContext *s, SingleChannelElement *sce)
372 if (s->coder->set_special_band_scalefactors)
373 s->coder->set_special_band_scalefactors(s, sce);
375 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
376 s->coder->encode_window_bands_info(s, sce, w, sce->ics.group_len[w], s->lambda);
380 * Encode scalefactors.
382 static void encode_scale_factors(AVCodecContext *avctx, AACEncContext *s,
383 SingleChannelElement *sce)
385 int diff, off_sf = sce->sf_idx[0], off_pns = sce->sf_idx[0] - NOISE_OFFSET;
386 int off_is = 0, noise_flag = 1;
389 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
390 for (i = 0; i < sce->ics.max_sfb; i++) {
391 if (!sce->zeroes[w*16 + i]) {
392 if (sce->band_type[w*16 + i] == NOISE_BT) {
393 diff = sce->sf_idx[w*16 + i] - off_pns;
394 off_pns = sce->sf_idx[w*16 + i];
395 if (noise_flag-- > 0) {
396 put_bits(&s->pb, NOISE_PRE_BITS, diff + NOISE_PRE);
399 } else if (sce->band_type[w*16 + i] == INTENSITY_BT ||
400 sce->band_type[w*16 + i] == INTENSITY_BT2) {
401 diff = sce->sf_idx[w*16 + i] - off_is;
402 off_is = sce->sf_idx[w*16 + i];
404 diff = sce->sf_idx[w*16 + i] - off_sf;
405 off_sf = sce->sf_idx[w*16 + i];
407 diff += SCALE_DIFF_ZERO;
408 av_assert0(diff >= 0 && diff <= 120);
409 put_bits(&s->pb, ff_aac_scalefactor_bits[diff], ff_aac_scalefactor_code[diff]);
418 static void encode_pulses(AACEncContext *s, Pulse *pulse)
422 put_bits(&s->pb, 1, !!pulse->num_pulse);
423 if (!pulse->num_pulse)
426 put_bits(&s->pb, 2, pulse->num_pulse - 1);
427 put_bits(&s->pb, 6, pulse->start);
428 for (i = 0; i < pulse->num_pulse; i++) {
429 put_bits(&s->pb, 5, pulse->pos[i]);
430 put_bits(&s->pb, 4, pulse->amp[i]);
435 * Encode spectral coefficients processed by psychoacoustic model.
437 static void encode_spectral_coeffs(AACEncContext *s, SingleChannelElement *sce)
441 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
443 for (i = 0; i < sce->ics.max_sfb; i++) {
444 if (sce->zeroes[w*16 + i]) {
445 start += sce->ics.swb_sizes[i];
448 for (w2 = w; w2 < w + sce->ics.group_len[w]; w2++) {
449 s->coder->quantize_and_encode_band(s, &s->pb,
450 &sce->coeffs[start + w2*128],
451 NULL, sce->ics.swb_sizes[i],
452 sce->sf_idx[w*16 + i],
453 sce->band_type[w*16 + i],
455 sce->ics.window_clipping[w]);
457 start += sce->ics.swb_sizes[i];
463 * Downscale spectral coefficients for near-clipping windows to avoid artifacts
465 static void avoid_clipping(AACEncContext *s, SingleChannelElement *sce)
469 if (sce->ics.clip_avoidance_factor < 1.0f) {
470 for (w = 0; w < sce->ics.num_windows; w++) {
472 for (i = 0; i < sce->ics.max_sfb; i++) {
473 float *swb_coeffs = &sce->coeffs[start + w*128];
474 for (j = 0; j < sce->ics.swb_sizes[i]; j++)
475 swb_coeffs[j] *= sce->ics.clip_avoidance_factor;
476 start += sce->ics.swb_sizes[i];
483 * Encode one channel of audio data.
485 static int encode_individual_channel(AVCodecContext *avctx, AACEncContext *s,
486 SingleChannelElement *sce,
489 put_bits(&s->pb, 8, sce->sf_idx[0]);
490 if (!common_window) {
491 put_ics_info(s, &sce->ics);
492 if (s->coder->encode_main_pred)
493 s->coder->encode_main_pred(s, sce);
494 if (s->coder->encode_ltp_info)
495 s->coder->encode_ltp_info(s, sce, 0);
497 encode_band_info(s, sce);
498 encode_scale_factors(avctx, s, sce);
499 encode_pulses(s, &sce->pulse);
500 put_bits(&s->pb, 1, !!sce->tns.present);
501 if (s->coder->encode_tns_info)
502 s->coder->encode_tns_info(s, sce);
503 put_bits(&s->pb, 1, 0); //ssr
504 encode_spectral_coeffs(s, sce);
509 * Write some auxiliary information about the created AAC file.
511 static void put_bitstream_info(AACEncContext *s, const char *name)
513 int i, namelen, padbits;
515 namelen = strlen(name) + 2;
516 put_bits(&s->pb, 3, TYPE_FIL);
517 put_bits(&s->pb, 4, FFMIN(namelen, 15));
519 put_bits(&s->pb, 8, namelen - 14);
520 put_bits(&s->pb, 4, 0); //extension type - filler
521 padbits = -put_bits_count(&s->pb) & 7;
522 align_put_bits(&s->pb);
523 for (i = 0; i < namelen - 2; i++)
524 put_bits(&s->pb, 8, name[i]);
525 put_bits(&s->pb, 12 - padbits, 0);
529 * Copy input samples.
530 * Channels are reordered from libavcodec's default order to AAC order.
532 static void copy_input_samples(AACEncContext *s, const AVFrame *frame)
535 int end = 2048 + (frame ? frame->nb_samples : 0);
536 const uint8_t *channel_map = s->reorder_map;
538 /* copy and remap input samples */
539 for (ch = 0; ch < s->channels; ch++) {
540 /* copy last 1024 samples of previous frame to the start of the current frame */
541 memcpy(&s->planar_samples[ch][1024], &s->planar_samples[ch][2048], 1024 * sizeof(s->planar_samples[0][0]));
543 /* copy new samples and zero any remaining samples */
545 memcpy(&s->planar_samples[ch][2048],
546 frame->extended_data[channel_map[ch]],
547 frame->nb_samples * sizeof(s->planar_samples[0][0]));
549 memset(&s->planar_samples[ch][end], 0,
550 (3072 - end) * sizeof(s->planar_samples[0][0]));
554 static int aac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
555 const AVFrame *frame, int *got_packet_ptr)
557 AACEncContext *s = avctx->priv_data;
558 float **samples = s->planar_samples, *samples2, *la, *overlap;
560 SingleChannelElement *sce;
561 IndividualChannelStream *ics;
562 int i, its, ch, w, chans, tag, start_ch, ret, frame_bits;
563 int target_bits, rate_bits, too_many_bits, too_few_bits;
564 int ms_mode = 0, is_mode = 0, tns_mode = 0, pred_mode = 0;
565 int chan_el_counter[4];
566 FFPsyWindowInfo windows[AAC_MAX_CHANNELS];
568 /* add current frame to queue */
570 if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
573 if (!s->afq.remaining_samples || (!s->afq.frame_alloc && !s->afq.frame_count))
577 copy_input_samples(s, frame);
579 ff_psy_preprocess(s->psypp, s->planar_samples, s->channels);
581 if (!avctx->frame_number)
585 for (i = 0; i < s->chan_map[0]; i++) {
586 FFPsyWindowInfo* wi = windows + start_ch;
587 tag = s->chan_map[i+1];
588 chans = tag == TYPE_CPE ? 2 : 1;
590 for (ch = 0; ch < chans; ch++) {
592 float clip_avoidance_factor;
595 s->cur_channel = start_ch + ch;
596 overlap = &samples[s->cur_channel][0];
597 samples2 = overlap + 1024;
598 la = samples2 + (448+64);
601 if (tag == TYPE_LFE) {
602 wi[ch].window_type[0] = wi[ch].window_type[1] = ONLY_LONG_SEQUENCE;
603 wi[ch].window_shape = 0;
604 wi[ch].num_windows = 1;
605 wi[ch].grouping[0] = 1;
606 wi[ch].clipping[0] = 0;
608 /* Only the lowest 12 coefficients are used in a LFE channel.
609 * The expression below results in only the bottom 8 coefficients
610 * being used for 11.025kHz to 16kHz sample rates.
612 ics->num_swb = s->samplerate_index >= 8 ? 1 : 3;
614 wi[ch] = s->psy.model->window(&s->psy, samples2, la, s->cur_channel,
615 ics->window_sequence[0]);
617 ics->window_sequence[1] = ics->window_sequence[0];
618 ics->window_sequence[0] = wi[ch].window_type[0];
619 ics->use_kb_window[1] = ics->use_kb_window[0];
620 ics->use_kb_window[0] = wi[ch].window_shape;
621 ics->num_windows = wi[ch].num_windows;
622 ics->swb_sizes = s->psy.bands [ics->num_windows == 8];
623 ics->num_swb = tag == TYPE_LFE ? ics->num_swb : s->psy.num_bands[ics->num_windows == 8];
624 ics->max_sfb = FFMIN(ics->max_sfb, ics->num_swb);
625 ics->swb_offset = wi[ch].window_type[0] == EIGHT_SHORT_SEQUENCE ?
626 ff_swb_offset_128 [s->samplerate_index]:
627 ff_swb_offset_1024[s->samplerate_index];
628 ics->tns_max_bands = wi[ch].window_type[0] == EIGHT_SHORT_SEQUENCE ?
629 ff_tns_max_bands_128 [s->samplerate_index]:
630 ff_tns_max_bands_1024[s->samplerate_index];
632 for (w = 0; w < ics->num_windows; w++)
633 ics->group_len[w] = wi[ch].grouping[w];
635 /* Calculate input sample maximums and evaluate clipping risk */
636 clip_avoidance_factor = 0.0f;
637 for (w = 0; w < ics->num_windows; w++) {
638 const float *wbuf = overlap + w * 128;
639 const int wlen = 2048 / ics->num_windows;
642 /* mdct input is 2 * output */
643 for (j = 0; j < wlen; j++)
644 max = FFMAX(max, fabsf(wbuf[j]));
645 wi[ch].clipping[w] = max;
647 for (w = 0; w < ics->num_windows; w++) {
648 if (wi[ch].clipping[w] > CLIP_AVOIDANCE_FACTOR) {
649 ics->window_clipping[w] = 1;
650 clip_avoidance_factor = FFMAX(clip_avoidance_factor, wi[ch].clipping[w]);
652 ics->window_clipping[w] = 0;
655 if (clip_avoidance_factor > CLIP_AVOIDANCE_FACTOR) {
656 ics->clip_avoidance_factor = CLIP_AVOIDANCE_FACTOR / clip_avoidance_factor;
658 ics->clip_avoidance_factor = 1.0f;
661 apply_window_and_mdct(s, sce, overlap);
663 if (s->options.ltp && s->coder->update_ltp) {
664 s->coder->update_ltp(s, sce);
665 apply_window[sce->ics.window_sequence[0]](s->fdsp, sce, &sce->ltp_state[0]);
666 s->mdct1024.mdct_calc(&s->mdct1024, sce->lcoeffs, sce->ret_buf);
669 for (k = 0; k < 1024; k++) {
670 if (!(fabs(cpe->ch[ch].coeffs[k]) < 1E16)) { // Ensure headroom for energy calculation
671 av_log(avctx, AV_LOG_ERROR, "Input contains (near) NaN/+-Inf\n");
672 return AVERROR(EINVAL);
675 avoid_clipping(s, sce);
679 if ((ret = ff_alloc_packet2(avctx, avpkt, 8192 * s->channels, 0)) < 0)
681 frame_bits = its = 0;
683 init_put_bits(&s->pb, avpkt->data, avpkt->size);
685 if ((avctx->frame_number & 0xFF)==1 && !(avctx->flags & AV_CODEC_FLAG_BITEXACT))
686 put_bitstream_info(s, LIBAVCODEC_IDENT);
689 memset(chan_el_counter, 0, sizeof(chan_el_counter));
690 for (i = 0; i < s->chan_map[0]; i++) {
691 FFPsyWindowInfo* wi = windows + start_ch;
692 const float *coeffs[2];
693 tag = s->chan_map[i+1];
694 chans = tag == TYPE_CPE ? 2 : 1;
696 cpe->common_window = 0;
697 memset(cpe->is_mask, 0, sizeof(cpe->is_mask));
698 memset(cpe->ms_mask, 0, sizeof(cpe->ms_mask));
699 put_bits(&s->pb, 3, tag);
700 put_bits(&s->pb, 4, chan_el_counter[tag]++);
701 for (ch = 0; ch < chans; ch++) {
703 coeffs[ch] = sce->coeffs;
704 sce->ics.predictor_present = 0;
705 sce->ics.ltp.present = 0;
706 memset(sce->ics.ltp.used, 0, sizeof(sce->ics.ltp.used));
707 memset(sce->ics.prediction_used, 0, sizeof(sce->ics.prediction_used));
708 memset(&sce->tns, 0, sizeof(TemporalNoiseShaping));
709 for (w = 0; w < 128; w++)
710 if (sce->band_type[w] > RESERVED_BT)
711 sce->band_type[w] = 0;
713 s->psy.bitres.alloc = -1;
714 s->psy.bitres.bits = s->last_frame_pb_count / s->channels;
715 s->psy.model->analyze(&s->psy, start_ch, coeffs, wi);
716 if (s->psy.bitres.alloc > 0) {
717 /* Lambda unused here on purpose, we need to take psy's unscaled allocation */
718 target_bits += s->psy.bitres.alloc
719 * (s->lambda / (avctx->global_quality ? avctx->global_quality : 120));
720 s->psy.bitres.alloc /= chans;
723 for (ch = 0; ch < chans; ch++) {
724 s->cur_channel = start_ch + ch;
725 if (s->options.pns && s->coder->mark_pns)
726 s->coder->mark_pns(s, avctx, &cpe->ch[ch]);
727 s->coder->search_for_quantizers(avctx, s, &cpe->ch[ch], s->lambda);
730 && wi[0].window_type[0] == wi[1].window_type[0]
731 && wi[0].window_shape == wi[1].window_shape) {
733 cpe->common_window = 1;
734 for (w = 0; w < wi[0].num_windows; w++) {
735 if (wi[0].grouping[w] != wi[1].grouping[w]) {
736 cpe->common_window = 0;
741 for (ch = 0; ch < chans; ch++) { /* TNS and PNS */
743 s->cur_channel = start_ch + ch;
744 if (s->options.tns && s->coder->search_for_tns)
745 s->coder->search_for_tns(s, sce);
746 if (s->options.tns && s->coder->apply_tns_filt)
747 s->coder->apply_tns_filt(s, sce);
748 if (sce->tns.present)
750 if (s->options.pns && s->coder->search_for_pns)
751 s->coder->search_for_pns(s, avctx, sce);
753 s->cur_channel = start_ch;
754 if (s->options.intensity_stereo) { /* Intensity Stereo */
755 if (s->coder->search_for_is)
756 s->coder->search_for_is(s, avctx, cpe);
757 if (cpe->is_mode) is_mode = 1;
758 apply_intensity_stereo(cpe);
760 if (s->options.pred) { /* Prediction */
761 for (ch = 0; ch < chans; ch++) {
763 s->cur_channel = start_ch + ch;
764 if (s->options.pred && s->coder->search_for_pred)
765 s->coder->search_for_pred(s, sce);
766 if (cpe->ch[ch].ics.predictor_present) pred_mode = 1;
768 if (s->coder->adjust_common_pred)
769 s->coder->adjust_common_pred(s, cpe);
770 for (ch = 0; ch < chans; ch++) {
772 s->cur_channel = start_ch + ch;
773 if (s->options.pred && s->coder->apply_main_pred)
774 s->coder->apply_main_pred(s, sce);
776 s->cur_channel = start_ch;
778 if (s->options.mid_side) { /* Mid/Side stereo */
779 if (s->options.mid_side == -1 && s->coder->search_for_ms)
780 s->coder->search_for_ms(s, cpe);
781 else if (cpe->common_window)
782 memset(cpe->ms_mask, 1, sizeof(cpe->ms_mask));
783 apply_mid_side_stereo(cpe);
785 adjust_frame_information(cpe, chans);
786 if (s->options.ltp) { /* LTP */
787 for (ch = 0; ch < chans; ch++) {
789 s->cur_channel = start_ch + ch;
790 if (s->coder->search_for_ltp)
791 s->coder->search_for_ltp(s, sce, cpe->common_window);
792 if (sce->ics.ltp.present) pred_mode = 1;
794 s->cur_channel = start_ch;
795 if (s->coder->adjust_common_ltp)
796 s->coder->adjust_common_ltp(s, cpe);
799 put_bits(&s->pb, 1, cpe->common_window);
800 if (cpe->common_window) {
801 put_ics_info(s, &cpe->ch[0].ics);
802 if (s->coder->encode_main_pred)
803 s->coder->encode_main_pred(s, &cpe->ch[0]);
804 if (s->coder->encode_ltp_info)
805 s->coder->encode_ltp_info(s, &cpe->ch[0], 1);
806 encode_ms_info(&s->pb, cpe);
807 if (cpe->ms_mode) ms_mode = 1;
810 for (ch = 0; ch < chans; ch++) {
811 s->cur_channel = start_ch + ch;
812 encode_individual_channel(avctx, s, &cpe->ch[ch], cpe->common_window);
817 if (avctx->flags & AV_CODEC_FLAG_QSCALE) {
818 /* When using a constant Q-scale, don't mess with lambda */
822 /* rate control stuff
823 * allow between the nominal bitrate, and what psy's bit reservoir says to target
824 * but drift towards the nominal bitrate always
826 frame_bits = put_bits_count(&s->pb);
827 rate_bits = avctx->bit_rate * 1024 / avctx->sample_rate;
828 rate_bits = FFMIN(rate_bits, 6144 * s->channels - 3);
829 too_many_bits = FFMAX(target_bits, rate_bits);
830 too_many_bits = FFMIN(too_many_bits, 6144 * s->channels - 3);
831 too_few_bits = FFMIN(FFMAX(rate_bits - rate_bits/4, target_bits), too_many_bits);
833 /* When using ABR, be strict (but only for increasing) */
834 too_few_bits = too_few_bits - too_few_bits/8;
835 too_many_bits = too_many_bits + too_many_bits/2;
837 if ( its == 0 /* for steady-state Q-scale tracking */
838 || (its < 5 && (frame_bits < too_few_bits || frame_bits > too_many_bits))
839 || frame_bits >= 6144 * s->channels - 3 )
841 float ratio = ((float)rate_bits) / frame_bits;
843 if (frame_bits >= too_few_bits && frame_bits <= too_many_bits) {
845 * This path is for steady-state Q-scale tracking
846 * When frame bits fall within the stable range, we still need to adjust
847 * lambda to maintain it like so in a stable fashion (large jumps in lambda
848 * create artifacts and should be avoided), but slowly
850 ratio = sqrtf(sqrtf(ratio));
851 ratio = av_clipf(ratio, 0.9f, 1.1f);
853 /* Not so fast though */
854 ratio = sqrtf(ratio);
856 s->lambda = FFMIN(s->lambda * ratio, 65536.f);
858 /* Keep iterating if we must reduce and lambda is in the sky */
859 if (ratio > 0.9f && ratio < 1.1f) {
862 if (is_mode || ms_mode || tns_mode || pred_mode) {
863 for (i = 0; i < s->chan_map[0]; i++) {
864 // Must restore coeffs
865 chans = tag == TYPE_CPE ? 2 : 1;
867 for (ch = 0; ch < chans; ch++)
868 memcpy(cpe->ch[ch].coeffs, cpe->ch[ch].pcoeffs, sizeof(cpe->ch[ch].coeffs));
878 if (s->options.ltp && s->coder->ltp_insert_new_frame)
879 s->coder->ltp_insert_new_frame(s);
881 put_bits(&s->pb, 3, TYPE_END);
882 flush_put_bits(&s->pb);
884 s->last_frame_pb_count = put_bits_count(&s->pb);
886 s->lambda_sum += s->lambda;
889 ff_af_queue_remove(&s->afq, avctx->frame_size, &avpkt->pts,
892 avpkt->size = put_bits_count(&s->pb) >> 3;
897 static av_cold int aac_encode_end(AVCodecContext *avctx)
899 AACEncContext *s = avctx->priv_data;
901 av_log(avctx, AV_LOG_INFO, "Qavg: %.3f\n", s->lambda_sum / s->lambda_count);
903 ff_mdct_end(&s->mdct1024);
904 ff_mdct_end(&s->mdct128);
908 ff_psy_preprocess_end(s->psypp);
909 av_freep(&s->buffer.samples);
912 ff_af_queue_close(&s->afq);
916 static av_cold int dsp_init(AVCodecContext *avctx, AACEncContext *s)
920 s->fdsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);
922 return AVERROR(ENOMEM);
925 ff_kbd_window_init(ff_aac_kbd_long_1024, 4.0, 1024);
926 ff_kbd_window_init(ff_aac_kbd_short_128, 6.0, 128);
927 ff_init_ff_sine_windows(10);
928 ff_init_ff_sine_windows(7);
930 if ((ret = ff_mdct_init(&s->mdct1024, 11, 0, 32768.0)) < 0)
932 if ((ret = ff_mdct_init(&s->mdct128, 8, 0, 32768.0)) < 0)
938 static av_cold int alloc_buffers(AVCodecContext *avctx, AACEncContext *s)
941 if (!FF_ALLOCZ_TYPED_ARRAY(s->buffer.samples, s->channels * 3 * 1024) ||
942 !FF_ALLOCZ_TYPED_ARRAY(s->cpe, s->chan_map[0]))
943 return AVERROR(ENOMEM);
945 for(ch = 0; ch < s->channels; ch++)
946 s->planar_samples[ch] = s->buffer.samples + 3 * 1024 * ch;
951 static av_cold int aac_encode_init(AVCodecContext *avctx)
953 AACEncContext *s = avctx->priv_data;
955 const uint8_t *sizes[2];
956 uint8_t grouping[AAC_MAX_CHANNELS];
960 s->last_frame_pb_count = 0;
961 avctx->frame_size = 1024;
962 avctx->initial_padding = 1024;
963 s->lambda = avctx->global_quality > 0 ? avctx->global_quality : 120;
965 /* Channel map and unspecified bitrate guessing */
966 s->channels = avctx->channels;
969 for (i = 0; i < FF_ARRAY_ELEMS(aac_normal_chan_layouts); i++) {
970 if (avctx->channel_layout == aac_normal_chan_layouts[i]) {
971 s->needs_pce = s->options.pce;
978 for (i = 0; i < FF_ARRAY_ELEMS(aac_pce_configs); i++)
979 if (avctx->channel_layout == aac_pce_configs[i].layout)
981 av_get_channel_layout_string(buf, sizeof(buf), -1, avctx->channel_layout);
982 ERROR_IF(i == FF_ARRAY_ELEMS(aac_pce_configs), "Unsupported channel layout \"%s\"\n", buf);
983 av_log(avctx, AV_LOG_INFO, "Using a PCE to encode channel layout \"%s\"\n", buf);
984 s->pce = aac_pce_configs[i];
985 s->reorder_map = s->pce.reorder_map;
986 s->chan_map = s->pce.config_map;
988 s->reorder_map = aac_chan_maps[s->channels - 1];
989 s->chan_map = aac_chan_configs[s->channels - 1];
992 if (!avctx->bit_rate) {
993 for (i = 1; i <= s->chan_map[0]; i++) {
994 avctx->bit_rate += s->chan_map[i] == TYPE_CPE ? 128000 : /* Pair */
995 s->chan_map[i] == TYPE_LFE ? 16000 : /* LFE */
1001 for (i = 0; i < 16; i++)
1002 if (avctx->sample_rate == avpriv_mpeg4audio_sample_rates[i])
1004 s->samplerate_index = i;
1005 ERROR_IF(s->samplerate_index == 16 ||
1006 s->samplerate_index >= ff_aac_swb_size_1024_len ||
1007 s->samplerate_index >= ff_aac_swb_size_128_len,
1008 "Unsupported sample rate %d\n", avctx->sample_rate);
1010 /* Bitrate limiting */
1011 WARN_IF(1024.0 * avctx->bit_rate / avctx->sample_rate > 6144 * s->channels,
1012 "Too many bits %f > %d per frame requested, clamping to max\n",
1013 1024.0 * avctx->bit_rate / avctx->sample_rate,
1014 6144 * s->channels);
1015 avctx->bit_rate = (int64_t)FFMIN(6144 * s->channels / 1024.0 * avctx->sample_rate,
1018 /* Profile and option setting */
1019 avctx->profile = avctx->profile == FF_PROFILE_UNKNOWN ? FF_PROFILE_AAC_LOW :
1021 for (i = 0; i < FF_ARRAY_ELEMS(aacenc_profiles); i++)
1022 if (avctx->profile == aacenc_profiles[i])
1024 if (avctx->profile == FF_PROFILE_MPEG2_AAC_LOW) {
1025 avctx->profile = FF_PROFILE_AAC_LOW;
1026 ERROR_IF(s->options.pred,
1027 "Main prediction unavailable in the \"mpeg2_aac_low\" profile\n");
1028 ERROR_IF(s->options.ltp,
1029 "LTP prediction unavailable in the \"mpeg2_aac_low\" profile\n");
1030 WARN_IF(s->options.pns,
1031 "PNS unavailable in the \"mpeg2_aac_low\" profile, turning off\n");
1033 } else if (avctx->profile == FF_PROFILE_AAC_LTP) {
1035 ERROR_IF(s->options.pred,
1036 "Main prediction unavailable in the \"aac_ltp\" profile\n");
1037 } else if (avctx->profile == FF_PROFILE_AAC_MAIN) {
1038 s->options.pred = 1;
1039 ERROR_IF(s->options.ltp,
1040 "LTP prediction unavailable in the \"aac_main\" profile\n");
1041 } else if (s->options.ltp) {
1042 avctx->profile = FF_PROFILE_AAC_LTP;
1044 "Chainging profile to \"aac_ltp\"\n");
1045 ERROR_IF(s->options.pred,
1046 "Main prediction unavailable in the \"aac_ltp\" profile\n");
1047 } else if (s->options.pred) {
1048 avctx->profile = FF_PROFILE_AAC_MAIN;
1050 "Chainging profile to \"aac_main\"\n");
1051 ERROR_IF(s->options.ltp,
1052 "LTP prediction unavailable in the \"aac_main\" profile\n");
1054 s->profile = avctx->profile;
1056 /* Coder limitations */
1057 s->coder = &ff_aac_coders[s->options.coder];
1058 if (s->options.coder == AAC_CODER_ANMR) {
1059 ERROR_IF(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL,
1060 "The ANMR coder is considered experimental, add -strict -2 to enable!\n");
1061 s->options.intensity_stereo = 0;
1064 ERROR_IF(s->options.ltp && avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL,
1065 "The LPT profile requires experimental compliance, add -strict -2 to enable!\n");
1067 /* M/S introduces horrible artifacts with multichannel files, this is temporary */
1068 if (s->channels > 3)
1069 s->options.mid_side = 0;
1071 if ((ret = dsp_init(avctx, s)) < 0)
1074 if ((ret = alloc_buffers(avctx, s)) < 0)
1077 if ((ret = put_audio_specific_config(avctx)))
1080 sizes[0] = ff_aac_swb_size_1024[s->samplerate_index];
1081 sizes[1] = ff_aac_swb_size_128[s->samplerate_index];
1082 lengths[0] = ff_aac_num_swb_1024[s->samplerate_index];
1083 lengths[1] = ff_aac_num_swb_128[s->samplerate_index];
1084 for (i = 0; i < s->chan_map[0]; i++)
1085 grouping[i] = s->chan_map[i + 1] == TYPE_CPE;
1086 if ((ret = ff_psy_init(&s->psy, avctx, 2, sizes, lengths,
1087 s->chan_map[0], grouping)) < 0)
1089 s->psypp = ff_psy_preprocess_init(avctx);
1090 ff_lpc_init(&s->lpc, 2*avctx->frame_size, TNS_MAX_ORDER, FF_LPC_TYPE_LEVINSON);
1091 s->random_state = 0x1f2e3d4c;
1093 s->abs_pow34 = abs_pow34_v;
1094 s->quant_bands = quantize_bands;
1097 ff_aac_dsp_init_x86(s);
1100 ff_aac_coder_init_mips(s);
1102 ff_af_queue_init(avctx, &s->afq);
1108 #define AACENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
1109 static const AVOption aacenc_options[] = {
1110 {"aac_coder", "Coding algorithm", offsetof(AACEncContext, options.coder), AV_OPT_TYPE_INT, {.i64 = AAC_CODER_FAST}, 0, AAC_CODER_NB-1, AACENC_FLAGS, "coder"},
1111 {"anmr", "ANMR method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_ANMR}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
1112 {"twoloop", "Two loop searching method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_TWOLOOP}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
1113 {"fast", "Default fast search", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_FAST}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
1114 {"aac_ms", "Force M/S stereo coding", offsetof(AACEncContext, options.mid_side), AV_OPT_TYPE_BOOL, {.i64 = -1}, -1, 1, AACENC_FLAGS},
1115 {"aac_is", "Intensity stereo coding", offsetof(AACEncContext, options.intensity_stereo), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1116 {"aac_pns", "Perceptual noise substitution", offsetof(AACEncContext, options.pns), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1117 {"aac_tns", "Temporal noise shaping", offsetof(AACEncContext, options.tns), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1118 {"aac_ltp", "Long term prediction", offsetof(AACEncContext, options.ltp), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
1119 {"aac_pred", "AAC-Main prediction", offsetof(AACEncContext, options.pred), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
1120 {"aac_pce", "Forces the use of PCEs", offsetof(AACEncContext, options.pce), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
1125 static const AVClass aacenc_class = {
1126 .class_name = "AAC encoder",
1127 .item_name = av_default_item_name,
1128 .option = aacenc_options,
1129 .version = LIBAVUTIL_VERSION_INT,
1132 static const AVCodecDefault aac_encode_defaults[] = {
1137 AVCodec ff_aac_encoder = {
1139 .long_name = NULL_IF_CONFIG_SMALL("AAC (Advanced Audio Coding)"),
1140 .type = AVMEDIA_TYPE_AUDIO,
1141 .id = AV_CODEC_ID_AAC,
1142 .priv_data_size = sizeof(AACEncContext),
1143 .init = aac_encode_init,
1144 .encode2 = aac_encode_frame,
1145 .close = aac_encode_end,
1146 .defaults = aac_encode_defaults,
1147 .supported_samplerates = mpeg4audio_sample_rates,
1148 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE | FF_CODEC_CAP_INIT_CLEANUP,
1149 .capabilities = AV_CODEC_CAP_SMALL_LAST_FRAME | AV_CODEC_CAP_DELAY,
1150 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
1151 AV_SAMPLE_FMT_NONE },
1152 .priv_class = &aacenc_class,