2 * AAC encoder psychoacoustic model
3 * Copyright (C) 2008 Konstantin Shishkov
5 * This file is part of Libav.
7 * Libav 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 * Libav 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 Libav; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * AAC encoder psychoacoustic model
31 /***********************************
33 * thresholds linearization after their modifications for attaining given bitrate
34 * try other bitrate controlling mechanism (maybe use ratecontrol.c?)
35 * control quality for quality-based output
36 **********************************/
39 * constants for 3GPP AAC psychoacoustic model
42 #define PSY_3GPP_THR_SPREAD_HI 1.5f // spreading factor for low-to-hi threshold spreading (15 dB/Bark)
43 #define PSY_3GPP_THR_SPREAD_LOW 3.0f // spreading factor for hi-to-low threshold spreading (30 dB/Bark)
45 #define PSY_3GPP_RPEMIN 0.01f
46 #define PSY_3GPP_RPELEV 2.0f
48 /* LAME psy model constants */
49 #define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order
50 #define AAC_BLOCK_SIZE_LONG 1024 ///< long block size
51 #define AAC_BLOCK_SIZE_SHORT 128 ///< short block size
52 #define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence
53 #define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block
60 * information for single band used by 3GPP TS26.403-inspired psychoacoustic model
62 typedef struct AacPsyBand{
63 float energy; ///< band energy
64 float thr; ///< energy threshold
65 float thr_quiet; ///< threshold in quiet
69 * single/pair channel context for psychoacoustic model
71 typedef struct AacPsyChannel{
72 AacPsyBand band[128]; ///< bands information
73 AacPsyBand prev_band[128]; ///< bands information from the previous frame
75 float win_energy; ///< sliding average of channel energy
76 float iir_state[2]; ///< hi-pass IIR filter state
77 uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence)
78 enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame
79 /* LAME psy model specific members */
80 float attack_threshold; ///< attack threshold for this channel
81 float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS];
82 int prev_attack; ///< attack value for the last short block in the previous sequence
86 * psychoacoustic model frame type-dependent coefficients
88 typedef struct AacPsyCoeffs{
89 float ath; ///< absolute threshold of hearing per bands
90 float barks; ///< Bark value for each spectral band in long frame
91 float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame
92 float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame
93 float min_snr; ///< minimal SNR
97 * 3GPP TS26.403-inspired psychoacoustic model specific data
99 typedef struct AacPsyContext{
100 AacPsyCoeffs psy_coef[2][64];
105 * LAME psy model preset struct
108 int quality; ///< Quality to map the rest of the vaules to.
109 /* This is overloaded to be both kbps per channel in ABR mode, and
110 * requested quality in constant quality mode.
112 float st_lrm; ///< short threshold for L, R, and M channels
116 * LAME psy model preset table for ABR
118 static const PsyLamePreset psy_abr_map[] = {
119 /* TODO: Tuning. These were taken from LAME. */
137 * LAME psy model preset table for constant quality
139 static const PsyLamePreset psy_vbr_map[] = {
155 * LAME psy model FIR coefficient table
157 static const float psy_fir_coeffs[] = {
158 -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
159 -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
160 -5.52212e-17 * 2, -0.313819 * 2
164 * calculates the attack threshold for ABR from the above table for the LAME psy model
166 static float lame_calc_attack_threshold(int bitrate)
168 /* Assume max bitrate to start with */
169 int lower_range = 12, upper_range = 12;
170 int lower_range_kbps = psy_abr_map[12].quality;
171 int upper_range_kbps = psy_abr_map[12].quality;
174 /* Determine which bitrates the value specified falls between.
175 * If the loop ends without breaking our above assumption of 320kbps was correct.
177 for (i = 1; i < 13; i++) {
178 if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) {
180 upper_range_kbps = psy_abr_map[i ].quality;
182 lower_range_kbps = psy_abr_map[i - 1].quality;
183 break; /* Upper range found */
187 /* Determine which range the value specified is closer to */
188 if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps))
189 return psy_abr_map[lower_range].st_lrm;
190 return psy_abr_map[upper_range].st_lrm;
194 * LAME psy model specific initialization
196 static void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx) {
199 for (i = 0; i < avctx->channels; i++) {
200 AacPsyChannel *pch = &ctx->ch[i];
202 if (avctx->flags & CODEC_FLAG_QSCALE)
203 pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm;
205 pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000);
207 for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++)
208 pch->prev_energy_subshort[j] = 10.0f;
213 * Calculate Bark value for given line.
215 static av_cold float calc_bark(float f)
217 return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f));
222 * Calculate ATH value for given frequency.
223 * Borrowed from Lame.
225 static av_cold float ath(float f, float add)
228 return 3.64 * pow(f, -0.8)
229 - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4))
230 + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7))
231 + (0.6 + 0.04 * add) * 0.001 * f * f * f * f;
234 static av_cold int psy_3gpp_init(FFPsyContext *ctx) {
238 float prev, minscale, minath;
240 ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
241 pctx = (AacPsyContext*) ctx->model_priv_data;
243 minath = ath(3410, ATH_ADD);
244 for (j = 0; j < 2; j++) {
245 AacPsyCoeffs *coeffs = pctx->psy_coef[j];
246 const uint8_t *band_sizes = ctx->bands[j];
247 float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
250 for (g = 0; g < ctx->num_bands[j]; g++) {
252 bark = calc_bark((i-1) * line_to_frequency);
253 coeffs[g].barks = (bark + prev) / 2.0;
256 for (g = 0; g < ctx->num_bands[j] - 1; g++) {
257 AacPsyCoeffs *coeff = &coeffs[g];
258 float bark_width = coeffs[g+1].barks - coeffs->barks;
259 coeff->spread_low[0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_LOW);
260 coeff->spread_hi [0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_HI);
263 for (g = 0; g < ctx->num_bands[j]; g++) {
264 minscale = ath(start * line_to_frequency, ATH_ADD);
265 for (i = 1; i < band_sizes[g]; i++)
266 minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD));
267 coeffs[g].ath = minscale - minath;
268 start += band_sizes[g];
272 pctx->ch = av_mallocz(sizeof(AacPsyChannel) * ctx->avctx->channels);
274 lame_window_init(pctx, ctx->avctx);
280 * IIR filter used in block switching decision
282 static float iir_filter(int in, float state[2])
286 ret = 0.7548f * (in - state[0]) + 0.5095f * state[1];
293 * window grouping information stored as bits (0 - new group, 1 - group continues)
295 static const uint8_t window_grouping[9] = {
296 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
300 * Tell encoder which window types to use.
301 * @see 3GPP TS26.403 5.4.1 "Blockswitching"
303 static FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx,
304 const int16_t *audio, const int16_t *la,
305 int channel, int prev_type)
308 int br = ctx->avctx->bit_rate / ctx->avctx->channels;
309 int attack_ratio = br <= 16000 ? 18 : 10;
310 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
311 AacPsyChannel *pch = &pctx->ch[channel];
312 uint8_t grouping = 0;
313 int next_type = pch->next_window_seq;
316 memset(&wi, 0, sizeof(wi));
319 int switch_to_eight = 0;
320 float sum = 0.0, sum2 = 0.0;
323 for (i = 0; i < 8; i++) {
324 for (j = 0; j < 128; j++) {
325 v = iir_filter(la[(i*128+j)*ctx->avctx->channels], pch->iir_state);
331 for (i = 0; i < 8; i++) {
332 if (s[i] > pch->win_energy * attack_ratio) {
338 pch->win_energy = pch->win_energy*7/8 + sum2/64;
340 wi.window_type[1] = prev_type;
342 case ONLY_LONG_SEQUENCE:
343 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
344 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
346 case LONG_START_SEQUENCE:
347 wi.window_type[0] = EIGHT_SHORT_SEQUENCE;
348 grouping = pch->next_grouping;
349 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
351 case LONG_STOP_SEQUENCE:
352 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
353 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
355 case EIGHT_SHORT_SEQUENCE:
356 stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight;
357 wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
358 grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0;
359 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
363 pch->next_grouping = window_grouping[attack_n];
364 pch->next_window_seq = next_type;
366 for (i = 0; i < 3; i++)
367 wi.window_type[i] = prev_type;
368 grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0;
372 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
378 for (i = 0; i < 8; i++) {
379 if (!((grouping >> i) & 1))
381 wi.grouping[lastgrp]++;
389 * Calculate band thresholds as suggested in 3GPP TS26.403
391 static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
392 const float *coefs, const FFPsyWindowInfo *wi)
394 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
395 AacPsyChannel *pch = &pctx->ch[channel];
398 const int num_bands = ctx->num_bands[wi->num_windows == 8];
399 const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8];
400 AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
402 //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
403 for (w = 0; w < wi->num_windows*16; w += 16) {
404 for (g = 0; g < num_bands; g++) {
405 AacPsyBand *band = &pch->band[w+g];
407 for (i = 0; i < band_sizes[g]; i++)
408 band->energy += coefs[start+i] * coefs[start+i];
409 band->thr = band->energy * 0.001258925f;
410 start += band_sizes[g];
413 //modify thresholds and energies - spread, threshold in quiet, pre-echo control
414 for (w = 0; w < wi->num_windows*16; w += 16) {
415 AacPsyBand *bands = &pch->band[w];
416 //5.4.2.3 "Spreading" & 5.4.3 "Spreaded Energy Calculation"
417 for (g = 1; g < num_bands; g++)
418 bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]);
419 for (g = num_bands - 2; g >= 0; g--)
420 bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]);
421 //5.4.2.4 "Threshold in quiet"
422 for (g = 0; g < num_bands; g++) {
423 AacPsyBand *band = &bands[g];
424 band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath);
425 //5.4.2.5 "Pre-echo control"
426 if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (wi->window_type[1] == LONG_START_SEQUENCE && !w)))
427 band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr,
428 PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
432 for (w = 0; w < wi->num_windows*16; w += 16) {
433 for (g = 0; g < num_bands; g++) {
434 AacPsyBand *band = &pch->band[w+g];
435 FFPsyBand *psy_band = &ctx->psy_bands[channel*PSY_MAX_BANDS+w+g];
437 psy_band->threshold = band->thr;
438 psy_band->energy = band->energy;
442 memcpy(pch->prev_band, pch->band, sizeof(pch->band));
445 static av_cold void psy_3gpp_end(FFPsyContext *apc)
447 AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data;
449 av_freep(&apc->model_priv_data);
452 static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
454 int blocktype = ONLY_LONG_SEQUENCE;
456 if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE)
457 blocktype = LONG_STOP_SEQUENCE;
459 blocktype = EIGHT_SHORT_SEQUENCE;
460 if (ctx->next_window_seq == ONLY_LONG_SEQUENCE)
461 ctx->next_window_seq = LONG_START_SEQUENCE;
462 if (ctx->next_window_seq == LONG_STOP_SEQUENCE)
463 ctx->next_window_seq = EIGHT_SHORT_SEQUENCE;
466 wi->window_type[0] = ctx->next_window_seq;
467 ctx->next_window_seq = blocktype;
470 static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx,
471 const int16_t *audio, const int16_t *la,
472 int channel, int prev_type)
474 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
475 AacPsyChannel *pch = &pctx->ch[channel];
477 int uselongblock = 1;
478 int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
482 memset(&wi, 0, sizeof(wi));
484 float hpfsmpl[AAC_BLOCK_SIZE_LONG];
485 float const *pf = hpfsmpl;
486 float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
487 float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
488 float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
489 int chans = ctx->avctx->channels;
490 const int16_t *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN) * chans;
493 /* LAME comment: apply high pass filter of fs/4 */
494 for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
496 sum1 = firbuf[(i + ((PSY_LAME_FIR_LEN - 1) / 2)) * chans];
498 for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
499 sum1 += psy_fir_coeffs[j] * (firbuf[(i + j) * chans] + firbuf[(i + PSY_LAME_FIR_LEN - j) * chans]);
500 sum2 += psy_fir_coeffs[j + 1] * (firbuf[(i + j + 1) * chans] + firbuf[(i + PSY_LAME_FIR_LEN - j - 1) * chans]);
502 hpfsmpl[i] = sum1 + sum2;
505 /* Calculate the energies of each sub-shortblock */
506 for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) {
507 energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)];
508 assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0);
509 attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)];
510 energy_short[0] += energy_subshort[i];
513 for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) {
514 float const *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS);
516 for (; pf < pfe; pf++)
519 pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
520 energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
521 /* FIXME: The indexes below are [i + 3 - 2] in the LAME source.
522 * Obviously the 3 and 2 have some significance, or this would be just [i + 1]
523 * (which is what we use here). What the 3 stands for is ambigious, as it is both
524 * number of short blocks, and the number of sub-short blocks.
525 * It seems that LAME is comparing each sub-block to sub-block + 1 in the
528 if (p > energy_subshort[i + 1])
529 p = p / energy_subshort[i + 1];
530 else if (energy_subshort[i + 1] > p * 10.0f)
531 p = energy_subshort[i + 1] / (p * 10.0f);
534 attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
537 /* compare energy between sub-short blocks */
538 for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++)
539 if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS])
540 if (attack_intensity[i] > pch->attack_threshold)
541 attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1;
543 /* should have energy change between short blocks, in order to avoid periodic signals */
544 /* Good samples to show the effect are Trumpet test songs */
545 /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */
546 /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */
547 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) {
548 float const u = energy_short[i - 1];
549 float const v = energy_short[i];
550 float const m = FFMAX(u, v);
551 if (m < 40000) { /* (2) */
552 if (u < 1.7f * v && v < 1.7f * u) { /* (1) */
553 if (i == 1 && attacks[0] < attacks[i])
558 att_sum += attacks[i];
561 if (attacks[0] <= pch->prev_attack)
564 att_sum += attacks[0];
565 /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */
566 if (pch->prev_attack == 3 || att_sum) {
569 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++)
570 if (attacks[i] && attacks[i-1])
574 /* We have no lookahead info, so just use same type as the previous sequence. */
575 uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE);
578 lame_apply_block_type(pch, &wi, uselongblock);
580 wi.window_type[1] = prev_type;
581 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
584 if (wi.window_type[0] == LONG_START_SEQUENCE)
593 for (i = 0; i < 8; i++) {
594 if (!((pch->next_grouping >> i) & 1))
596 wi.grouping[lastgrp]++;
600 /* Determine grouping, based on the location of the first attack, and save for
602 * FIXME: Move this to analysis.
603 * TODO: Tune groupings depending on attack location
604 * TODO: Handle more than one attack in a group
606 for (i = 0; i < 9; i++) {
612 pch->next_grouping = window_grouping[grouping];
614 pch->prev_attack = attacks[8];
619 const FFPsyModel ff_aac_psy_model =
621 .name = "3GPP TS 26.403-inspired model",
622 .init = psy_3gpp_init,
623 .window = psy_lame_window,
624 .analyze = psy_3gpp_analyze,