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;
54 * Make AAC audio config object.
55 * @see 1.6.2.1 "Syntax - AudioSpecificConfig"
57 static void put_audio_specific_config(AVCodecContext *avctx)
60 AACEncContext *s = avctx->priv_data;
61 int channels = s->channels - (s->channels == 8 ? 1 : 0);
63 init_put_bits(&pb, avctx->extradata, avctx->extradata_size);
64 put_bits(&pb, 5, s->profile+1); //profile
65 put_bits(&pb, 4, s->samplerate_index); //sample rate index
66 put_bits(&pb, 4, channels);
68 put_bits(&pb, 1, 0); //frame length - 1024 samples
69 put_bits(&pb, 1, 0); //does not depend on core coder
70 put_bits(&pb, 1, 0); //is not extension
72 //Explicitly Mark SBR absent
73 put_bits(&pb, 11, 0x2b7); //sync extension
74 put_bits(&pb, 5, AOT_SBR);
79 void ff_quantize_band_cost_cache_init(struct AACEncContext *s)
81 ++s->quantize_band_cost_cache_generation;
82 if (s->quantize_band_cost_cache_generation == 0) {
83 memset(s->quantize_band_cost_cache, 0, sizeof(s->quantize_band_cost_cache));
84 s->quantize_band_cost_cache_generation = 1;
88 #define WINDOW_FUNC(type) \
89 static void apply_ ##type ##_window(AVFloatDSPContext *fdsp, \
90 SingleChannelElement *sce, \
93 WINDOW_FUNC(only_long)
95 const float *lwindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
96 const float *pwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
97 float *out = sce->ret_buf;
99 fdsp->vector_fmul (out, audio, lwindow, 1024);
100 fdsp->vector_fmul_reverse(out + 1024, audio + 1024, pwindow, 1024);
103 WINDOW_FUNC(long_start)
105 const float *lwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_long_1024 : ff_sine_1024;
106 const float *swindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
107 float *out = sce->ret_buf;
109 fdsp->vector_fmul(out, audio, lwindow, 1024);
110 memcpy(out + 1024, audio + 1024, sizeof(out[0]) * 448);
111 fdsp->vector_fmul_reverse(out + 1024 + 448, audio + 1024 + 448, swindow, 128);
112 memset(out + 1024 + 576, 0, sizeof(out[0]) * 448);
115 WINDOW_FUNC(long_stop)
117 const float *lwindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_long_1024 : ff_sine_1024;
118 const float *swindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;
119 float *out = sce->ret_buf;
121 memset(out, 0, sizeof(out[0]) * 448);
122 fdsp->vector_fmul(out + 448, audio + 448, swindow, 128);
123 memcpy(out + 576, audio + 576, sizeof(out[0]) * 448);
124 fdsp->vector_fmul_reverse(out + 1024, audio + 1024, lwindow, 1024);
127 WINDOW_FUNC(eight_short)
129 const float *swindow = sce->ics.use_kb_window[0] ? ff_aac_kbd_short_128 : ff_sine_128;
130 const float *pwindow = sce->ics.use_kb_window[1] ? ff_aac_kbd_short_128 : ff_sine_128;
131 const float *in = audio + 448;
132 float *out = sce->ret_buf;
135 for (w = 0; w < 8; w++) {
136 fdsp->vector_fmul (out, in, w ? pwindow : swindow, 128);
139 fdsp->vector_fmul_reverse(out, in, swindow, 128);
144 static void (*const apply_window[4])(AVFloatDSPContext *fdsp,
145 SingleChannelElement *sce,
146 const float *audio) = {
147 [ONLY_LONG_SEQUENCE] = apply_only_long_window,
148 [LONG_START_SEQUENCE] = apply_long_start_window,
149 [EIGHT_SHORT_SEQUENCE] = apply_eight_short_window,
150 [LONG_STOP_SEQUENCE] = apply_long_stop_window
153 static void apply_window_and_mdct(AACEncContext *s, SingleChannelElement *sce,
157 const float *output = sce->ret_buf;
159 apply_window[sce->ics.window_sequence[0]](s->fdsp, sce, audio);
161 if (sce->ics.window_sequence[0] != EIGHT_SHORT_SEQUENCE)
162 s->mdct1024.mdct_calc(&s->mdct1024, sce->coeffs, output);
164 for (i = 0; i < 1024; i += 128)
165 s->mdct128.mdct_calc(&s->mdct128, &sce->coeffs[i], output + i*2);
166 memcpy(audio, audio + 1024, sizeof(audio[0]) * 1024);
167 memcpy(sce->pcoeffs, sce->coeffs, sizeof(sce->pcoeffs));
171 * Encode ics_info element.
172 * @see Table 4.6 (syntax of ics_info)
174 static void put_ics_info(AACEncContext *s, IndividualChannelStream *info)
178 put_bits(&s->pb, 1, 0); // ics_reserved bit
179 put_bits(&s->pb, 2, info->window_sequence[0]);
180 put_bits(&s->pb, 1, info->use_kb_window[0]);
181 if (info->window_sequence[0] != EIGHT_SHORT_SEQUENCE) {
182 put_bits(&s->pb, 6, info->max_sfb);
183 put_bits(&s->pb, 1, !!info->predictor_present);
185 put_bits(&s->pb, 4, info->max_sfb);
186 for (w = 1; w < 8; w++)
187 put_bits(&s->pb, 1, !info->group_len[w]);
193 * @see 4.6.8.1 "Joint Coding - M/S Stereo"
195 static void encode_ms_info(PutBitContext *pb, ChannelElement *cpe)
199 put_bits(pb, 2, cpe->ms_mode);
200 if (cpe->ms_mode == 1)
201 for (w = 0; w < cpe->ch[0].ics.num_windows; w += cpe->ch[0].ics.group_len[w])
202 for (i = 0; i < cpe->ch[0].ics.max_sfb; i++)
203 put_bits(pb, 1, cpe->ms_mask[w*16 + i]);
207 * Produce integer coefficients from scalefactors provided by the model.
209 static void adjust_frame_information(ChannelElement *cpe, int chans)
214 for (ch = 0; ch < chans; ch++) {
215 IndividualChannelStream *ics = &cpe->ch[ch].ics;
217 cpe->ch[ch].pulse.num_pulse = 0;
218 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
219 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
220 for (cmaxsfb = ics->num_swb; cmaxsfb > 0 && cpe->ch[ch].zeroes[w*16+cmaxsfb-1]; cmaxsfb--)
222 maxsfb = FFMAX(maxsfb, cmaxsfb);
225 ics->max_sfb = maxsfb;
227 //adjust zero bands for window groups
228 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
229 for (g = 0; g < ics->max_sfb; g++) {
231 for (w2 = w; w2 < w + ics->group_len[w]; w2++) {
232 if (!cpe->ch[ch].zeroes[w2*16 + g]) {
237 cpe->ch[ch].zeroes[w*16 + g] = i;
242 if (chans > 1 && cpe->common_window) {
243 IndividualChannelStream *ics0 = &cpe->ch[0].ics;
244 IndividualChannelStream *ics1 = &cpe->ch[1].ics;
246 ics0->max_sfb = FFMAX(ics0->max_sfb, ics1->max_sfb);
247 ics1->max_sfb = ics0->max_sfb;
248 for (w = 0; w < ics0->num_windows*16; w += 16)
249 for (i = 0; i < ics0->max_sfb; i++)
250 if (cpe->ms_mask[w+i])
252 if (msc == 0 || ics0->max_sfb == 0)
255 cpe->ms_mode = msc < ics0->max_sfb * ics0->num_windows ? 1 : 2;
259 static void apply_intensity_stereo(ChannelElement *cpe)
262 IndividualChannelStream *ics = &cpe->ch[0].ics;
263 if (!cpe->common_window)
265 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
266 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
267 int start = (w+w2) * 128;
268 for (g = 0; g < ics->num_swb; g++) {
269 int p = -1 + 2 * (cpe->ch[1].band_type[w*16+g] - 14);
270 float scale = cpe->ch[0].is_ener[w*16+g];
271 if (!cpe->is_mask[w*16 + g]) {
272 start += ics->swb_sizes[g];
275 if (cpe->ms_mask[w*16 + g])
277 for (i = 0; i < ics->swb_sizes[g]; i++) {
278 float sum = (cpe->ch[0].coeffs[start+i] + p*cpe->ch[1].coeffs[start+i])*scale;
279 cpe->ch[0].coeffs[start+i] = sum;
280 cpe->ch[1].coeffs[start+i] = 0.0f;
282 start += ics->swb_sizes[g];
288 static void apply_mid_side_stereo(ChannelElement *cpe)
291 IndividualChannelStream *ics = &cpe->ch[0].ics;
292 if (!cpe->common_window)
294 for (w = 0; w < ics->num_windows; w += ics->group_len[w]) {
295 for (w2 = 0; w2 < ics->group_len[w]; w2++) {
296 int start = (w+w2) * 128;
297 for (g = 0; g < ics->num_swb; g++) {
298 /* ms_mask can be used for other purposes in PNS and I/S,
299 * so must not apply M/S if any band uses either, even if
302 if (!cpe->ms_mask[w*16 + g] || cpe->is_mask[w*16 + g]
303 || cpe->ch[0].band_type[w*16 + g] >= NOISE_BT
304 || cpe->ch[1].band_type[w*16 + g] >= NOISE_BT) {
305 start += ics->swb_sizes[g];
308 for (i = 0; i < ics->swb_sizes[g]; i++) {
309 float L = (cpe->ch[0].coeffs[start+i] + cpe->ch[1].coeffs[start+i]) * 0.5f;
310 float R = L - cpe->ch[1].coeffs[start+i];
311 cpe->ch[0].coeffs[start+i] = L;
312 cpe->ch[1].coeffs[start+i] = R;
314 start += ics->swb_sizes[g];
321 * Encode scalefactor band coding type.
323 static void encode_band_info(AACEncContext *s, SingleChannelElement *sce)
327 if (s->coder->set_special_band_scalefactors)
328 s->coder->set_special_band_scalefactors(s, sce);
330 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
331 s->coder->encode_window_bands_info(s, sce, w, sce->ics.group_len[w], s->lambda);
335 * Encode scalefactors.
337 static void encode_scale_factors(AVCodecContext *avctx, AACEncContext *s,
338 SingleChannelElement *sce)
340 int diff, off_sf = sce->sf_idx[0], off_pns = sce->sf_idx[0] - NOISE_OFFSET;
341 int off_is = 0, noise_flag = 1;
344 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
345 for (i = 0; i < sce->ics.max_sfb; i++) {
346 if (!sce->zeroes[w*16 + i]) {
347 if (sce->band_type[w*16 + i] == NOISE_BT) {
348 diff = sce->sf_idx[w*16 + i] - off_pns;
349 off_pns = sce->sf_idx[w*16 + i];
350 if (noise_flag-- > 0) {
351 put_bits(&s->pb, NOISE_PRE_BITS, diff + NOISE_PRE);
354 } else if (sce->band_type[w*16 + i] == INTENSITY_BT ||
355 sce->band_type[w*16 + i] == INTENSITY_BT2) {
356 diff = sce->sf_idx[w*16 + i] - off_is;
357 off_is = sce->sf_idx[w*16 + i];
359 diff = sce->sf_idx[w*16 + i] - off_sf;
360 off_sf = sce->sf_idx[w*16 + i];
362 diff += SCALE_DIFF_ZERO;
363 av_assert0(diff >= 0 && diff <= 120);
364 put_bits(&s->pb, ff_aac_scalefactor_bits[diff], ff_aac_scalefactor_code[diff]);
373 static void encode_pulses(AACEncContext *s, Pulse *pulse)
377 put_bits(&s->pb, 1, !!pulse->num_pulse);
378 if (!pulse->num_pulse)
381 put_bits(&s->pb, 2, pulse->num_pulse - 1);
382 put_bits(&s->pb, 6, pulse->start);
383 for (i = 0; i < pulse->num_pulse; i++) {
384 put_bits(&s->pb, 5, pulse->pos[i]);
385 put_bits(&s->pb, 4, pulse->amp[i]);
390 * Encode spectral coefficients processed by psychoacoustic model.
392 static void encode_spectral_coeffs(AACEncContext *s, SingleChannelElement *sce)
396 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
398 for (i = 0; i < sce->ics.max_sfb; i++) {
399 if (sce->zeroes[w*16 + i]) {
400 start += sce->ics.swb_sizes[i];
403 for (w2 = w; w2 < w + sce->ics.group_len[w]; w2++) {
404 s->coder->quantize_and_encode_band(s, &s->pb,
405 &sce->coeffs[start + w2*128],
406 NULL, sce->ics.swb_sizes[i],
407 sce->sf_idx[w*16 + i],
408 sce->band_type[w*16 + i],
410 sce->ics.window_clipping[w]);
412 start += sce->ics.swb_sizes[i];
418 * Downscale spectral coefficients for near-clipping windows to avoid artifacts
420 static void avoid_clipping(AACEncContext *s, SingleChannelElement *sce)
424 if (sce->ics.clip_avoidance_factor < 1.0f) {
425 for (w = 0; w < sce->ics.num_windows; w++) {
427 for (i = 0; i < sce->ics.max_sfb; i++) {
428 float *swb_coeffs = &sce->coeffs[start + w*128];
429 for (j = 0; j < sce->ics.swb_sizes[i]; j++)
430 swb_coeffs[j] *= sce->ics.clip_avoidance_factor;
431 start += sce->ics.swb_sizes[i];
438 * Encode one channel of audio data.
440 static int encode_individual_channel(AVCodecContext *avctx, AACEncContext *s,
441 SingleChannelElement *sce,
444 put_bits(&s->pb, 8, sce->sf_idx[0]);
445 if (!common_window) {
446 put_ics_info(s, &sce->ics);
447 if (s->coder->encode_main_pred)
448 s->coder->encode_main_pred(s, sce);
449 if (s->coder->encode_ltp_info)
450 s->coder->encode_ltp_info(s, sce, 0);
452 encode_band_info(s, sce);
453 encode_scale_factors(avctx, s, sce);
454 encode_pulses(s, &sce->pulse);
455 put_bits(&s->pb, 1, !!sce->tns.present);
456 if (s->coder->encode_tns_info)
457 s->coder->encode_tns_info(s, sce);
458 put_bits(&s->pb, 1, 0); //ssr
459 encode_spectral_coeffs(s, sce);
464 * Write some auxiliary information about the created AAC file.
466 static void put_bitstream_info(AACEncContext *s, const char *name)
468 int i, namelen, padbits;
470 namelen = strlen(name) + 2;
471 put_bits(&s->pb, 3, TYPE_FIL);
472 put_bits(&s->pb, 4, FFMIN(namelen, 15));
474 put_bits(&s->pb, 8, namelen - 14);
475 put_bits(&s->pb, 4, 0); //extension type - filler
476 padbits = -put_bits_count(&s->pb) & 7;
477 avpriv_align_put_bits(&s->pb);
478 for (i = 0; i < namelen - 2; i++)
479 put_bits(&s->pb, 8, name[i]);
480 put_bits(&s->pb, 12 - padbits, 0);
484 * Copy input samples.
485 * Channels are reordered from libavcodec's default order to AAC order.
487 static void copy_input_samples(AACEncContext *s, const AVFrame *frame)
490 int end = 2048 + (frame ? frame->nb_samples : 0);
491 const uint8_t *channel_map = aac_chan_maps[s->channels - 1];
493 /* copy and remap input samples */
494 for (ch = 0; ch < s->channels; ch++) {
495 /* copy last 1024 samples of previous frame to the start of the current frame */
496 memcpy(&s->planar_samples[ch][1024], &s->planar_samples[ch][2048], 1024 * sizeof(s->planar_samples[0][0]));
498 /* copy new samples and zero any remaining samples */
500 memcpy(&s->planar_samples[ch][2048],
501 frame->extended_data[channel_map[ch]],
502 frame->nb_samples * sizeof(s->planar_samples[0][0]));
504 memset(&s->planar_samples[ch][end], 0,
505 (3072 - end) * sizeof(s->planar_samples[0][0]));
509 static int aac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
510 const AVFrame *frame, int *got_packet_ptr)
512 AACEncContext *s = avctx->priv_data;
513 float **samples = s->planar_samples, *samples2, *la, *overlap;
515 SingleChannelElement *sce;
516 IndividualChannelStream *ics;
517 int i, its, ch, w, chans, tag, start_ch, ret, frame_bits;
518 int target_bits, rate_bits, too_many_bits, too_few_bits;
519 int ms_mode = 0, is_mode = 0, tns_mode = 0, pred_mode = 0;
520 int chan_el_counter[4];
521 FFPsyWindowInfo windows[AAC_MAX_CHANNELS];
523 /* add current frame to queue */
525 if ((ret = ff_af_queue_add(&s->afq, frame)) < 0)
528 if (!s->afq.remaining_samples || (!s->afq.frame_alloc && !s->afq.frame_count))
532 copy_input_samples(s, frame);
534 ff_psy_preprocess(s->psypp, s->planar_samples, s->channels);
536 if (!avctx->frame_number)
540 for (i = 0; i < s->chan_map[0]; i++) {
541 FFPsyWindowInfo* wi = windows + start_ch;
542 tag = s->chan_map[i+1];
543 chans = tag == TYPE_CPE ? 2 : 1;
545 for (ch = 0; ch < chans; ch++) {
547 float clip_avoidance_factor;
550 s->cur_channel = start_ch + ch;
551 overlap = &samples[s->cur_channel][0];
552 samples2 = overlap + 1024;
553 la = samples2 + (448+64);
556 if (tag == TYPE_LFE) {
557 wi[ch].window_type[0] = wi[ch].window_type[1] = ONLY_LONG_SEQUENCE;
558 wi[ch].window_shape = 0;
559 wi[ch].num_windows = 1;
560 wi[ch].grouping[0] = 1;
561 wi[ch].clipping[0] = 0;
563 /* Only the lowest 12 coefficients are used in a LFE channel.
564 * The expression below results in only the bottom 8 coefficients
565 * being used for 11.025kHz to 16kHz sample rates.
567 ics->num_swb = s->samplerate_index >= 8 ? 1 : 3;
569 wi[ch] = s->psy.model->window(&s->psy, samples2, la, s->cur_channel,
570 ics->window_sequence[0]);
572 ics->window_sequence[1] = ics->window_sequence[0];
573 ics->window_sequence[0] = wi[ch].window_type[0];
574 ics->use_kb_window[1] = ics->use_kb_window[0];
575 ics->use_kb_window[0] = wi[ch].window_shape;
576 ics->num_windows = wi[ch].num_windows;
577 ics->swb_sizes = s->psy.bands [ics->num_windows == 8];
578 ics->num_swb = tag == TYPE_LFE ? ics->num_swb : s->psy.num_bands[ics->num_windows == 8];
579 ics->max_sfb = FFMIN(ics->max_sfb, ics->num_swb);
580 ics->swb_offset = wi[ch].window_type[0] == EIGHT_SHORT_SEQUENCE ?
581 ff_swb_offset_128 [s->samplerate_index]:
582 ff_swb_offset_1024[s->samplerate_index];
583 ics->tns_max_bands = wi[ch].window_type[0] == EIGHT_SHORT_SEQUENCE ?
584 ff_tns_max_bands_128 [s->samplerate_index]:
585 ff_tns_max_bands_1024[s->samplerate_index];
587 for (w = 0; w < ics->num_windows; w++)
588 ics->group_len[w] = wi[ch].grouping[w];
590 /* Calculate input sample maximums and evaluate clipping risk */
591 clip_avoidance_factor = 0.0f;
592 for (w = 0; w < ics->num_windows; w++) {
593 const float *wbuf = overlap + w * 128;
594 const int wlen = 2048 / ics->num_windows;
597 /* mdct input is 2 * output */
598 for (j = 0; j < wlen; j++)
599 max = FFMAX(max, fabsf(wbuf[j]));
600 wi[ch].clipping[w] = max;
602 for (w = 0; w < ics->num_windows; w++) {
603 if (wi[ch].clipping[w] > CLIP_AVOIDANCE_FACTOR) {
604 ics->window_clipping[w] = 1;
605 clip_avoidance_factor = FFMAX(clip_avoidance_factor, wi[ch].clipping[w]);
607 ics->window_clipping[w] = 0;
610 if (clip_avoidance_factor > CLIP_AVOIDANCE_FACTOR) {
611 ics->clip_avoidance_factor = CLIP_AVOIDANCE_FACTOR / clip_avoidance_factor;
613 ics->clip_avoidance_factor = 1.0f;
616 apply_window_and_mdct(s, sce, overlap);
618 if (s->options.ltp && s->coder->update_ltp) {
619 s->coder->update_ltp(s, sce);
620 apply_window[sce->ics.window_sequence[0]](s->fdsp, sce, &sce->ltp_state[0]);
621 s->mdct1024.mdct_calc(&s->mdct1024, sce->lcoeffs, sce->ret_buf);
624 for (k = 0; k < 1024; k++) {
625 if (!(fabs(cpe->ch[ch].coeffs[k]) < 1E16)) { // Ensure headroom for energy calculation
626 av_log(avctx, AV_LOG_ERROR, "Input contains (near) NaN/+-Inf\n");
627 return AVERROR(EINVAL);
630 avoid_clipping(s, sce);
634 if ((ret = ff_alloc_packet2(avctx, avpkt, 8192 * s->channels, 0)) < 0)
636 frame_bits = its = 0;
638 init_put_bits(&s->pb, avpkt->data, avpkt->size);
640 if ((avctx->frame_number & 0xFF)==1 && !(avctx->flags & AV_CODEC_FLAG_BITEXACT))
641 put_bitstream_info(s, LIBAVCODEC_IDENT);
644 memset(chan_el_counter, 0, sizeof(chan_el_counter));
645 for (i = 0; i < s->chan_map[0]; i++) {
646 FFPsyWindowInfo* wi = windows + start_ch;
647 const float *coeffs[2];
648 tag = s->chan_map[i+1];
649 chans = tag == TYPE_CPE ? 2 : 1;
651 cpe->common_window = 0;
652 memset(cpe->is_mask, 0, sizeof(cpe->is_mask));
653 memset(cpe->ms_mask, 0, sizeof(cpe->ms_mask));
654 put_bits(&s->pb, 3, tag);
655 put_bits(&s->pb, 4, chan_el_counter[tag]++);
656 for (ch = 0; ch < chans; ch++) {
658 coeffs[ch] = sce->coeffs;
659 sce->ics.predictor_present = 0;
660 sce->ics.ltp.present = 0;
661 memset(sce->ics.ltp.used, 0, sizeof(sce->ics.ltp.used));
662 memset(sce->ics.prediction_used, 0, sizeof(sce->ics.prediction_used));
663 memset(&sce->tns, 0, sizeof(TemporalNoiseShaping));
664 for (w = 0; w < 128; w++)
665 if (sce->band_type[w] > RESERVED_BT)
666 sce->band_type[w] = 0;
668 s->psy.bitres.alloc = -1;
669 s->psy.bitres.bits = s->last_frame_pb_count / s->channels;
670 s->psy.model->analyze(&s->psy, start_ch, coeffs, wi);
671 if (s->psy.bitres.alloc > 0) {
672 /* Lambda unused here on purpose, we need to take psy's unscaled allocation */
673 target_bits += s->psy.bitres.alloc
674 * (s->lambda / (avctx->global_quality ? avctx->global_quality : 120));
675 s->psy.bitres.alloc /= chans;
678 for (ch = 0; ch < chans; ch++) {
679 s->cur_channel = start_ch + ch;
680 if (s->options.pns && s->coder->mark_pns)
681 s->coder->mark_pns(s, avctx, &cpe->ch[ch]);
682 s->coder->search_for_quantizers(avctx, s, &cpe->ch[ch], s->lambda);
685 && wi[0].window_type[0] == wi[1].window_type[0]
686 && wi[0].window_shape == wi[1].window_shape) {
688 cpe->common_window = 1;
689 for (w = 0; w < wi[0].num_windows; w++) {
690 if (wi[0].grouping[w] != wi[1].grouping[w]) {
691 cpe->common_window = 0;
696 for (ch = 0; ch < chans; ch++) { /* TNS and PNS */
698 s->cur_channel = start_ch + ch;
699 if (s->options.tns && s->coder->search_for_tns)
700 s->coder->search_for_tns(s, sce);
701 if (s->options.tns && s->coder->apply_tns_filt)
702 s->coder->apply_tns_filt(s, sce);
703 if (sce->tns.present)
705 if (s->options.pns && s->coder->search_for_pns)
706 s->coder->search_for_pns(s, avctx, sce);
708 s->cur_channel = start_ch;
709 if (s->options.intensity_stereo) { /* Intensity Stereo */
710 if (s->coder->search_for_is)
711 s->coder->search_for_is(s, avctx, cpe);
712 if (cpe->is_mode) is_mode = 1;
713 apply_intensity_stereo(cpe);
715 if (s->options.pred) { /* Prediction */
716 for (ch = 0; ch < chans; ch++) {
718 s->cur_channel = start_ch + ch;
719 if (s->options.pred && s->coder->search_for_pred)
720 s->coder->search_for_pred(s, sce);
721 if (cpe->ch[ch].ics.predictor_present) pred_mode = 1;
723 if (s->coder->adjust_common_pred)
724 s->coder->adjust_common_pred(s, cpe);
725 for (ch = 0; ch < chans; ch++) {
727 s->cur_channel = start_ch + ch;
728 if (s->options.pred && s->coder->apply_main_pred)
729 s->coder->apply_main_pred(s, sce);
731 s->cur_channel = start_ch;
733 if (s->options.mid_side) { /* Mid/Side stereo */
734 if (s->options.mid_side == -1 && s->coder->search_for_ms)
735 s->coder->search_for_ms(s, cpe);
736 else if (cpe->common_window)
737 memset(cpe->ms_mask, 1, sizeof(cpe->ms_mask));
738 apply_mid_side_stereo(cpe);
740 adjust_frame_information(cpe, chans);
741 if (s->options.ltp) { /* LTP */
742 for (ch = 0; ch < chans; ch++) {
744 s->cur_channel = start_ch + ch;
745 if (s->coder->search_for_ltp)
746 s->coder->search_for_ltp(s, sce, cpe->common_window);
747 if (sce->ics.ltp.present) pred_mode = 1;
749 s->cur_channel = start_ch;
750 if (s->coder->adjust_common_ltp)
751 s->coder->adjust_common_ltp(s, cpe);
754 put_bits(&s->pb, 1, cpe->common_window);
755 if (cpe->common_window) {
756 put_ics_info(s, &cpe->ch[0].ics);
757 if (s->coder->encode_main_pred)
758 s->coder->encode_main_pred(s, &cpe->ch[0]);
759 if (s->coder->encode_ltp_info)
760 s->coder->encode_ltp_info(s, &cpe->ch[0], 1);
761 encode_ms_info(&s->pb, cpe);
762 if (cpe->ms_mode) ms_mode = 1;
765 for (ch = 0; ch < chans; ch++) {
766 s->cur_channel = start_ch + ch;
767 encode_individual_channel(avctx, s, &cpe->ch[ch], cpe->common_window);
772 if (avctx->flags & CODEC_FLAG_QSCALE) {
773 /* When using a constant Q-scale, don't mess with lambda */
777 /* rate control stuff
778 * allow between the nominal bitrate, and what psy's bit reservoir says to target
779 * but drift towards the nominal bitrate always
781 frame_bits = put_bits_count(&s->pb);
782 rate_bits = avctx->bit_rate * 1024 / avctx->sample_rate;
783 rate_bits = FFMIN(rate_bits, 6144 * s->channels - 3);
784 too_many_bits = FFMAX(target_bits, rate_bits);
785 too_many_bits = FFMIN(too_many_bits, 6144 * s->channels - 3);
786 too_few_bits = FFMIN(FFMAX(rate_bits - rate_bits/4, target_bits), too_many_bits);
788 /* When using ABR, be strict (but only for increasing) */
789 too_few_bits = too_few_bits - too_few_bits/8;
790 too_many_bits = too_many_bits + too_many_bits/2;
792 if ( its == 0 /* for steady-state Q-scale tracking */
793 || (its < 5 && (frame_bits < too_few_bits || frame_bits > too_many_bits))
794 || frame_bits >= 6144 * s->channels - 3 )
796 float ratio = ((float)rate_bits) / frame_bits;
798 if (frame_bits >= too_few_bits && frame_bits <= too_many_bits) {
800 * This path is for steady-state Q-scale tracking
801 * When frame bits fall within the stable range, we still need to adjust
802 * lambda to maintain it like so in a stable fashion (large jumps in lambda
803 * create artifacts and should be avoided), but slowly
805 ratio = sqrtf(sqrtf(ratio));
806 ratio = av_clipf(ratio, 0.9f, 1.1f);
808 /* Not so fast though */
809 ratio = sqrtf(ratio);
811 s->lambda = FFMIN(s->lambda * ratio, 65536.f);
813 /* Keep iterating if we must reduce and lambda is in the sky */
814 if (ratio > 0.9f && ratio < 1.1f) {
817 if (is_mode || ms_mode || tns_mode || pred_mode) {
818 for (i = 0; i < s->chan_map[0]; i++) {
819 // Must restore coeffs
820 chans = tag == TYPE_CPE ? 2 : 1;
822 for (ch = 0; ch < chans; ch++)
823 memcpy(cpe->ch[ch].coeffs, cpe->ch[ch].pcoeffs, sizeof(cpe->ch[ch].coeffs));
833 if (s->options.ltp && s->coder->ltp_insert_new_frame)
834 s->coder->ltp_insert_new_frame(s);
836 put_bits(&s->pb, 3, TYPE_END);
837 flush_put_bits(&s->pb);
839 s->last_frame_pb_count = put_bits_count(&s->pb);
841 s->lambda_sum += s->lambda;
844 ff_af_queue_remove(&s->afq, avctx->frame_size, &avpkt->pts,
847 avpkt->size = put_bits_count(&s->pb) >> 3;
852 static av_cold int aac_encode_end(AVCodecContext *avctx)
854 AACEncContext *s = avctx->priv_data;
856 av_log(avctx, AV_LOG_INFO, "Qavg: %.3f\n", s->lambda_sum / s->lambda_count);
858 ff_mdct_end(&s->mdct1024);
859 ff_mdct_end(&s->mdct128);
863 ff_psy_preprocess_end(s->psypp);
864 av_freep(&s->buffer.samples);
867 ff_af_queue_close(&s->afq);
871 static av_cold int dsp_init(AVCodecContext *avctx, AACEncContext *s)
875 s->fdsp = avpriv_float_dsp_alloc(avctx->flags & AV_CODEC_FLAG_BITEXACT);
877 return AVERROR(ENOMEM);
880 ff_kbd_window_init(ff_aac_kbd_long_1024, 4.0, 1024);
881 ff_kbd_window_init(ff_aac_kbd_short_128, 6.0, 128);
882 ff_init_ff_sine_windows(10);
883 ff_init_ff_sine_windows(7);
885 if ((ret = ff_mdct_init(&s->mdct1024, 11, 0, 32768.0)) < 0)
887 if ((ret = ff_mdct_init(&s->mdct128, 8, 0, 32768.0)) < 0)
893 static av_cold int alloc_buffers(AVCodecContext *avctx, AACEncContext *s)
896 FF_ALLOCZ_ARRAY_OR_GOTO(avctx, s->buffer.samples, s->channels, 3 * 1024 * sizeof(s->buffer.samples[0]), alloc_fail);
897 FF_ALLOCZ_ARRAY_OR_GOTO(avctx, s->cpe, s->chan_map[0], sizeof(ChannelElement), alloc_fail);
898 FF_ALLOCZ_OR_GOTO(avctx, avctx->extradata, 5 + AV_INPUT_BUFFER_PADDING_SIZE, alloc_fail);
900 for(ch = 0; ch < s->channels; ch++)
901 s->planar_samples[ch] = s->buffer.samples + 3 * 1024 * ch;
905 return AVERROR(ENOMEM);
908 static av_cold void aac_encode_init_tables(void)
913 static av_cold int aac_encode_init(AVCodecContext *avctx)
915 AACEncContext *s = avctx->priv_data;
917 const uint8_t *sizes[2];
918 uint8_t grouping[AAC_MAX_CHANNELS];
922 s->last_frame_pb_count = 0;
923 avctx->extradata_size = 5;
924 avctx->frame_size = 1024;
925 avctx->initial_padding = 1024;
926 s->lambda = avctx->global_quality > 0 ? avctx->global_quality : 120;
928 /* Channel map and unspecified bitrate guessing */
929 s->channels = avctx->channels;
930 ERROR_IF(s->channels > AAC_MAX_CHANNELS || s->channels == 7,
931 "Unsupported number of channels: %d\n", s->channels);
932 s->chan_map = aac_chan_configs[s->channels-1];
933 if (!avctx->bit_rate) {
934 for (i = 1; i <= s->chan_map[0]; i++) {
935 avctx->bit_rate += s->chan_map[i] == TYPE_CPE ? 128000 : /* Pair */
936 s->chan_map[i] == TYPE_LFE ? 16000 : /* LFE */
942 for (i = 0; i < 16; i++)
943 if (avctx->sample_rate == avpriv_mpeg4audio_sample_rates[i])
945 s->samplerate_index = i;
946 ERROR_IF(s->samplerate_index == 16 ||
947 s->samplerate_index >= ff_aac_swb_size_1024_len ||
948 s->samplerate_index >= ff_aac_swb_size_128_len,
949 "Unsupported sample rate %d\n", avctx->sample_rate);
951 /* Bitrate limiting */
952 WARN_IF(1024.0 * avctx->bit_rate / avctx->sample_rate > 6144 * s->channels,
953 "Too many bits %f > %d per frame requested, clamping to max\n",
954 1024.0 * avctx->bit_rate / avctx->sample_rate,
956 avctx->bit_rate = (int64_t)FFMIN(6144 * s->channels / 1024.0 * avctx->sample_rate,
959 /* Profile and option setting */
960 avctx->profile = avctx->profile == FF_PROFILE_UNKNOWN ? FF_PROFILE_AAC_LOW :
962 for (i = 0; i < FF_ARRAY_ELEMS(aacenc_profiles); i++)
963 if (avctx->profile == aacenc_profiles[i])
965 if (avctx->profile == FF_PROFILE_MPEG2_AAC_LOW) {
966 avctx->profile = FF_PROFILE_AAC_LOW;
967 ERROR_IF(s->options.pred,
968 "Main prediction unavailable in the \"mpeg2_aac_low\" profile\n");
969 ERROR_IF(s->options.ltp,
970 "LTP prediction unavailable in the \"mpeg2_aac_low\" profile\n");
971 WARN_IF(s->options.pns,
972 "PNS unavailable in the \"mpeg2_aac_low\" profile, turning off\n");
974 } else if (avctx->profile == FF_PROFILE_AAC_LTP) {
976 ERROR_IF(s->options.pred,
977 "Main prediction unavailable in the \"aac_ltp\" profile\n");
978 } else if (avctx->profile == FF_PROFILE_AAC_MAIN) {
980 ERROR_IF(s->options.ltp,
981 "LTP prediction unavailable in the \"aac_main\" profile\n");
982 } else if (s->options.ltp) {
983 avctx->profile = FF_PROFILE_AAC_LTP;
985 "Chainging profile to \"aac_ltp\"\n");
986 ERROR_IF(s->options.pred,
987 "Main prediction unavailable in the \"aac_ltp\" profile\n");
988 } else if (s->options.pred) {
989 avctx->profile = FF_PROFILE_AAC_MAIN;
991 "Chainging profile to \"aac_main\"\n");
992 ERROR_IF(s->options.ltp,
993 "LTP prediction unavailable in the \"aac_main\" profile\n");
995 s->profile = avctx->profile;
997 /* Coder limitations */
998 s->coder = &ff_aac_coders[s->options.coder];
999 if (s->options.coder == AAC_CODER_ANMR) {
1000 ERROR_IF(avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL,
1001 "The ANMR coder is considered experimental, add -strict -2 to enable!\n");
1002 s->options.intensity_stereo = 0;
1005 ERROR_IF(s->options.ltp && avctx->strict_std_compliance > FF_COMPLIANCE_EXPERIMENTAL,
1006 "The LPT profile requires experimental compliance, add -strict -2 to enable!\n");
1008 /* M/S introduces horrible artifacts with multichannel files, this is temporary */
1009 if (s->channels > 3)
1010 s->options.mid_side = 0;
1012 if ((ret = dsp_init(avctx, s)) < 0)
1015 if ((ret = alloc_buffers(avctx, s)) < 0)
1018 put_audio_specific_config(avctx);
1020 sizes[0] = ff_aac_swb_size_1024[s->samplerate_index];
1021 sizes[1] = ff_aac_swb_size_128[s->samplerate_index];
1022 lengths[0] = ff_aac_num_swb_1024[s->samplerate_index];
1023 lengths[1] = ff_aac_num_swb_128[s->samplerate_index];
1024 for (i = 0; i < s->chan_map[0]; i++)
1025 grouping[i] = s->chan_map[i + 1] == TYPE_CPE;
1026 if ((ret = ff_psy_init(&s->psy, avctx, 2, sizes, lengths,
1027 s->chan_map[0], grouping)) < 0)
1029 s->psypp = ff_psy_preprocess_init(avctx);
1030 ff_lpc_init(&s->lpc, 2*avctx->frame_size, TNS_MAX_ORDER, FF_LPC_TYPE_LEVINSON);
1031 s->random_state = 0x1f2e3d4c;
1033 s->abs_pow34 = abs_pow34_v;
1034 s->quant_bands = quantize_bands;
1037 ff_aac_dsp_init_x86(s);
1040 ff_aac_coder_init_mips(s);
1042 if ((ret = ff_thread_once(&aac_table_init, &aac_encode_init_tables)) != 0)
1043 return AVERROR_UNKNOWN;
1045 ff_af_queue_init(avctx, &s->afq);
1049 aac_encode_end(avctx);
1053 #define AACENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
1054 static const AVOption aacenc_options[] = {
1055 {"aac_coder", "Coding algorithm", offsetof(AACEncContext, options.coder), AV_OPT_TYPE_INT, {.i64 = AAC_CODER_TWOLOOP}, 0, AAC_CODER_NB-1, AACENC_FLAGS, "coder"},
1056 {"anmr", "ANMR method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_ANMR}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
1057 {"twoloop", "Two loop searching method", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_TWOLOOP}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
1058 {"fast", "Constant quantizer", 0, AV_OPT_TYPE_CONST, {.i64 = AAC_CODER_FAST}, INT_MIN, INT_MAX, AACENC_FLAGS, "coder"},
1059 {"aac_ms", "Force M/S stereo coding", offsetof(AACEncContext, options.mid_side), AV_OPT_TYPE_BOOL, {.i64 = -1}, -1, 1, AACENC_FLAGS},
1060 {"aac_is", "Intensity stereo coding", offsetof(AACEncContext, options.intensity_stereo), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1061 {"aac_pns", "Perceptual noise substitution", offsetof(AACEncContext, options.pns), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1062 {"aac_tns", "Temporal noise shaping", offsetof(AACEncContext, options.tns), AV_OPT_TYPE_BOOL, {.i64 = 1}, -1, 1, AACENC_FLAGS},
1063 {"aac_ltp", "Long term prediction", offsetof(AACEncContext, options.ltp), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
1064 {"aac_pred", "AAC-Main prediction", offsetof(AACEncContext, options.pred), AV_OPT_TYPE_BOOL, {.i64 = 0}, -1, 1, AACENC_FLAGS},
1068 static const AVClass aacenc_class = {
1070 av_default_item_name,
1072 LIBAVUTIL_VERSION_INT,
1075 static const AVCodecDefault aac_encode_defaults[] = {
1080 AVCodec ff_aac_encoder = {
1082 .long_name = NULL_IF_CONFIG_SMALL("AAC (Advanced Audio Coding)"),
1083 .type = AVMEDIA_TYPE_AUDIO,
1084 .id = AV_CODEC_ID_AAC,
1085 .priv_data_size = sizeof(AACEncContext),
1086 .init = aac_encode_init,
1087 .encode2 = aac_encode_frame,
1088 .close = aac_encode_end,
1089 .defaults = aac_encode_defaults,
1090 .supported_samplerates = mpeg4audio_sample_rates,
1091 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE,
1092 .capabilities = AV_CODEC_CAP_SMALL_LAST_FRAME | AV_CODEC_CAP_DELAY,
1093 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLTP,
1094 AV_SAMPLE_FMT_NONE },
1095 .priv_class = &aacenc_class,