2 * AAC encoder psychoacoustic model
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
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_SPREAD_LOW 1.5f // spreading factor for ascending threshold spreading (15 dB/Bark)
43 #define PSY_3GPP_SPREAD_HI 3.0f // spreading factor for descending 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 ffac; ///< form factor
65 float thr; ///< energy threshold
66 float min_snr; ///< minimal SNR
67 float thr_quiet; ///< threshold in quiet
71 * single/pair channel context for psychoacoustic model
73 typedef struct AacPsyChannel{
74 AacPsyBand band[128]; ///< bands information
75 AacPsyBand prev_band[128]; ///< bands information from the previous frame
77 float win_energy; ///< sliding average of channel energy
78 float iir_state[2]; ///< hi-pass IIR filter state
79 uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence)
80 enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame
81 /* LAME psy model specific members */
82 float attack_threshold; ///< attack threshold for this channel
83 float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS];
84 int prev_attack; ///< attack value for the last short block in the previous sequence
88 * psychoacoustic model frame type-dependent coefficients
90 typedef struct AacPsyCoeffs{
91 float ath [64]; ///< absolute threshold of hearing per bands
92 float barks [64]; ///< Bark value for each spectral band in long frame
93 float spread_low[64]; ///< spreading factor for low-to-high threshold spreading in long frame
94 float spread_hi [64]; ///< spreading factor for high-to-low threshold spreading in long frame
98 * 3GPP TS26.403-inspired psychoacoustic model specific data
100 typedef struct AacPsyContext{
101 AacPsyCoeffs psy_coef[2];
106 * LAME psy model preset struct
109 int quality; ///< Quality to map the rest of the vaules to.
110 /* This is overloaded to be both kbps per channel in ABR mode, and
111 * requested quality in constant quality mode.
113 float st_lrm; ///< short threshold for L, R, and M channels
117 * LAME psy model preset table for ABR
119 static const PsyLamePreset psy_abr_map[] = {
120 /* TODO: Tuning. These were taken from LAME. */
138 * LAME psy model preset table for constant quality
140 static const PsyLamePreset psy_vbr_map[] = {
156 * LAME psy model FIR coefficient table
158 static const float psy_fir_coeffs[] = {
159 -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
160 -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
161 -5.52212e-17 * 2, -0.313819 * 2
165 * calculates the attack threshold for ABR from the above table for the LAME psy model
167 static float lame_calc_attack_threshold(int bitrate)
169 /* Assume max bitrate to start with */
170 int lower_range = 12, upper_range = 12;
171 int lower_range_kbps = psy_abr_map[12].quality;
172 int upper_range_kbps = psy_abr_map[12].quality;
175 /* Determine which bitrates the value specified falls between.
176 * If the loop ends without breaking our above assumption of 320kbps was correct.
178 for (i = 1; i < 13; i++) {
179 if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) {
181 upper_range_kbps = psy_abr_map[i ].quality;
183 lower_range_kbps = psy_abr_map[i - 1].quality;
184 break; /* Upper range found */
188 /* Determine which range the value specified is closer to */
189 if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps))
190 return psy_abr_map[lower_range].st_lrm;
191 return psy_abr_map[upper_range].st_lrm;
195 * LAME psy model specific initialization
197 static void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx) {
200 for (i = 0; i < avctx->channels; i++) {
201 AacPsyChannel *pch = &ctx->ch[i];
203 if (avctx->flags & CODEC_FLAG_QSCALE)
204 pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm;
206 pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000);
208 for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++)
209 pch->prev_energy_subshort[i] = 10.0f;
214 * Calculate Bark value for given line.
216 static av_cold float calc_bark(float f)
218 return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f));
223 * Calculate ATH value for given frequency.
224 * Borrowed from Lame.
226 static av_cold float ath(float f, float add)
229 return 3.64 * pow(f, -0.8)
230 - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4))
231 + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7))
232 + (0.6 + 0.04 * add) * 0.001 * f * f * f * f;
235 static av_cold int psy_3gpp_init(FFPsyContext *ctx) {
239 float prev, minscale, minath;
241 ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
242 pctx = (AacPsyContext*) ctx->model_priv_data;
244 minath = ath(3410, ATH_ADD);
245 for (j = 0; j < 2; j++) {
246 AacPsyCoeffs *coeffs = &pctx->psy_coef[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++) {
251 i += ctx->bands[j][g];
252 bark = calc_bark((i-1) * line_to_frequency);
253 coeffs->barks[g] = (bark + prev) / 2.0;
256 for (g = 0; g < ctx->num_bands[j] - 1; g++) {
257 coeffs->spread_low[g] = pow(10.0, -(coeffs->barks[g+1] - coeffs->barks[g]) * PSY_3GPP_SPREAD_LOW);
258 coeffs->spread_hi [g] = pow(10.0, -(coeffs->barks[g+1] - coeffs->barks[g]) * PSY_3GPP_SPREAD_HI);
261 for (g = 0; g < ctx->num_bands[j]; g++) {
262 minscale = ath(start * line_to_frequency, ATH_ADD);
263 for (i = 1; i < ctx->bands[j][g]; i++)
264 minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD));
265 coeffs->ath[g] = minscale - minath;
266 start += ctx->bands[j][g];
270 pctx->ch = av_mallocz(sizeof(AacPsyChannel) * ctx->avctx->channels);
272 lame_window_init(pctx, ctx->avctx);
278 * IIR filter used in block switching decision
280 static float iir_filter(int in, float state[2])
284 ret = 0.7548f * (in - state[0]) + 0.5095f * state[1];
291 * window grouping information stored as bits (0 - new group, 1 - group continues)
293 static const uint8_t window_grouping[9] = {
294 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
298 * Tell encoder which window types to use.
299 * @see 3GPP TS26.403 5.4.1 "Blockswitching"
301 static FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx,
302 const int16_t *audio, const int16_t *la,
303 int channel, int prev_type)
306 int br = ctx->avctx->bit_rate / ctx->avctx->channels;
307 int attack_ratio = br <= 16000 ? 18 : 10;
308 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
309 AacPsyChannel *pch = &pctx->ch[channel];
310 uint8_t grouping = 0;
311 int next_type = pch->next_window_seq;
314 memset(&wi, 0, sizeof(wi));
317 int switch_to_eight = 0;
318 float sum = 0.0, sum2 = 0.0;
321 for (i = 0; i < 8; i++) {
322 for (j = 0; j < 128; j++) {
323 v = iir_filter(la[(i*128+j)*ctx->avctx->channels], pch->iir_state);
329 for (i = 0; i < 8; i++) {
330 if (s[i] > pch->win_energy * attack_ratio) {
336 pch->win_energy = pch->win_energy*7/8 + sum2/64;
338 wi.window_type[1] = prev_type;
340 case ONLY_LONG_SEQUENCE:
341 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
342 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
344 case LONG_START_SEQUENCE:
345 wi.window_type[0] = EIGHT_SHORT_SEQUENCE;
346 grouping = pch->next_grouping;
347 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
349 case LONG_STOP_SEQUENCE:
350 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
351 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
353 case EIGHT_SHORT_SEQUENCE:
354 stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight;
355 wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
356 grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0;
357 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
361 pch->next_grouping = window_grouping[attack_n];
362 pch->next_window_seq = next_type;
364 for (i = 0; i < 3; i++)
365 wi.window_type[i] = prev_type;
366 grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0;
370 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
376 for (i = 0; i < 8; i++) {
377 if (!((grouping >> i) & 1))
379 wi.grouping[lastgrp]++;
387 * Calculate band thresholds as suggested in 3GPP TS26.403
389 static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
390 const float *coefs, const FFPsyWindowInfo *wi)
392 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
393 AacPsyChannel *pch = &pctx->ch[channel];
396 const int num_bands = ctx->num_bands[wi->num_windows == 8];
397 const uint8_t* band_sizes = ctx->bands[wi->num_windows == 8];
398 AacPsyCoeffs *coeffs = &pctx->psy_coef[wi->num_windows == 8];
400 //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
401 for (w = 0; w < wi->num_windows*16; w += 16) {
402 for (g = 0; g < num_bands; g++) {
403 AacPsyBand *band = &pch->band[w+g];
405 for (i = 0; i < band_sizes[g]; i++)
406 band->energy += coefs[start+i] * coefs[start+i];
407 band->energy *= 1.0f / (512*512);
408 band->thr = band->energy * 0.001258925f;
409 start += band_sizes[g];
411 ctx->psy_bands[channel*PSY_MAX_BANDS+w+g].energy = band->energy;
414 //modify thresholds - spread, threshold in quiet - 5.4.3 "Spreaded Energy Calculation"
415 for (w = 0; w < wi->num_windows*16; w += 16) {
416 AacPsyBand *band = &pch->band[w];
417 for (g = 1; g < num_bands; g++)
418 band[g].thr = FFMAX(band[g].thr, band[g-1].thr * coeffs->spread_low[g-1]);
419 for (g = num_bands - 2; g >= 0; g--)
420 band[g].thr = FFMAX(band[g].thr, band[g+1].thr * coeffs->spread_hi [g]);
421 for (g = 0; g < num_bands; g++) {
422 band[g].thr_quiet = FFMAX(band[g].thr, coeffs->ath[g]);
423 if (wi->num_windows != 8 && wi->window_type[1] != EIGHT_SHORT_SEQUENCE)
424 band[g].thr_quiet = FFMAX(PSY_3GPP_RPEMIN*band[g].thr_quiet,
425 FFMIN(band[g].thr_quiet,
426 PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
427 band[g].thr = FFMAX(band[g].thr, band[g].thr_quiet * 0.25);
429 ctx->psy_bands[channel*PSY_MAX_BANDS+w+g].threshold = band[g].thr;
432 memcpy(pch->prev_band, pch->band, sizeof(pch->band));
435 static av_cold void psy_3gpp_end(FFPsyContext *apc)
437 AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data;
439 av_freep(&apc->model_priv_data);
442 static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
444 int blocktype = ONLY_LONG_SEQUENCE;
446 if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE)
447 blocktype = LONG_STOP_SEQUENCE;
449 blocktype = EIGHT_SHORT_SEQUENCE;
450 if (ctx->next_window_seq == ONLY_LONG_SEQUENCE)
451 ctx->next_window_seq = LONG_START_SEQUENCE;
452 if (ctx->next_window_seq == LONG_STOP_SEQUENCE)
453 ctx->next_window_seq = EIGHT_SHORT_SEQUENCE;
456 wi->window_type[0] = ctx->next_window_seq;
457 ctx->next_window_seq = blocktype;
460 static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx,
461 const int16_t *audio, const int16_t *la,
462 int channel, int prev_type)
464 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
465 AacPsyChannel *pch = &pctx->ch[channel];
467 int uselongblock = 1;
468 int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
472 memset(&wi, 0, sizeof(wi));
474 float hpfsmpl[AAC_BLOCK_SIZE_LONG];
475 float const *pf = hpfsmpl;
476 float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
477 float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
478 float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
479 int chans = ctx->avctx->channels;
480 const int16_t *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN) * chans;
483 /* LAME comment: apply high pass filter of fs/4 */
484 for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
486 sum1 = firbuf[(i + ((PSY_LAME_FIR_LEN - 1) / 2)) * chans];
488 for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
489 sum1 += psy_fir_coeffs[j] * (firbuf[(i + j) * chans] + firbuf[(i + PSY_LAME_FIR_LEN - j) * chans]);
490 sum2 += psy_fir_coeffs[j + 1] * (firbuf[(i + j + 1) * chans] + firbuf[(i + PSY_LAME_FIR_LEN - j - 1) * chans]);
492 hpfsmpl[i] = sum1 + sum2;
495 /* Calculate the energies of each sub-shortblock */
496 for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) {
497 energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)];
498 assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0);
499 attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)];
500 energy_short[0] += energy_subshort[i];
503 for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) {
504 float const *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS);
506 for (; pf < pfe; pf++)
509 pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
510 energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
511 /* FIXME: The indexes below are [i + 3 - 2] in the LAME source.
512 * Obviously the 3 and 2 have some significance, or this would be just [i + 1]
513 * (which is what we use here). What the 3 stands for is ambigious, as it is both
514 * number of short blocks, and the number of sub-short blocks.
515 * It seems that LAME is comparing each sub-block to sub-block + 1 in the
518 if (p > energy_subshort[i + 1])
519 p = p / energy_subshort[i + 1];
520 else if (energy_subshort[i + 1] > p * 10.0f)
521 p = energy_subshort[i + 1] / (p * 10.0f);
524 attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
527 /* compare energy between sub-short blocks */
528 for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++)
529 if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS])
530 if (attack_intensity[i] > pch->attack_threshold)
531 attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1;
533 /* should have energy change between short blocks, in order to avoid periodic signals */
534 /* Good samples to show the effect are Trumpet test songs */
535 /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */
536 /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */
537 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) {
538 float const u = energy_short[i - 1];
539 float const v = energy_short[i];
540 float const m = FFMAX(u, v);
541 if (m < 40000) { /* (2) */
542 if (u < 1.7f * v && v < 1.7f * u) { /* (1) */
543 if (i == 1 && attacks[0] < attacks[i])
548 att_sum += attacks[i];
551 if (attacks[0] <= pch->prev_attack)
554 att_sum += attacks[0];
555 /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */
556 if (pch->prev_attack == 3 || att_sum) {
559 if (attacks[1] && attacks[0])
561 if (attacks[2] && attacks[1])
563 if (attacks[3] && attacks[2])
565 if (attacks[4] && attacks[3])
567 if (attacks[5] && attacks[4])
569 if (attacks[6] && attacks[5])
571 if (attacks[7] && attacks[6])
573 if (attacks[8] && attacks[7])
577 /* We have no lookahead info, so just use same type as the previous sequence. */
578 uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE);
581 lame_apply_block_type(pch, &wi, uselongblock);
583 wi.window_type[1] = prev_type;
584 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
587 if (wi.window_type[0] == LONG_START_SEQUENCE)
596 for (i = 0; i < 8; i++) {
597 if (!((pch->next_grouping >> i) & 1))
599 wi.grouping[lastgrp]++;
603 /* Determine grouping, based on the location of the first attack, and save for
605 * FIXME: Move this to analysis.
606 * TODO: Tune groupings depending on attack location
607 * TODO: Handle more than one attack in a group
609 for (i = 0; i < 9; i++) {
615 pch->next_grouping = window_grouping[grouping];
617 pch->prev_attack = attacks[8];
622 const FFPsyModel ff_aac_psy_model =
624 .name = "3GPP TS 26.403-inspired model",
625 .init = psy_3gpp_init,
626 .window = psy_lame_window,
627 .analyze = psy_3gpp_analyze,