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
27 #include "libavutil/attributes.h"
32 /***********************************
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)
44 /* spreading factor for low-to-hi energy spreading, long block, > 22kbps/channel (20dB/Bark) */
45 #define PSY_3GPP_EN_SPREAD_HI_L1 2.0f
46 /* spreading factor for low-to-hi energy spreading, long block, <= 22kbps/channel (15dB/Bark) */
47 #define PSY_3GPP_EN_SPREAD_HI_L2 1.5f
48 /* spreading factor for low-to-hi energy spreading, short block (15 dB/Bark) */
49 #define PSY_3GPP_EN_SPREAD_HI_S 1.5f
50 /* spreading factor for hi-to-low energy spreading, long block (30dB/Bark) */
51 #define PSY_3GPP_EN_SPREAD_LOW_L 3.0f
52 /* spreading factor for hi-to-low energy spreading, short block (20dB/Bark) */
53 #define PSY_3GPP_EN_SPREAD_LOW_S 2.0f
55 #define PSY_3GPP_RPEMIN 0.01f
56 #define PSY_3GPP_RPELEV 2.0f
58 #define PSY_3GPP_C1 3.0f /* log2(8) */
59 #define PSY_3GPP_C2 1.3219281f /* log2(2.5) */
60 #define PSY_3GPP_C3 0.55935729f /* 1 - C2 / C1 */
62 #define PSY_SNR_1DB 7.9432821e-1f /* -1dB */
63 #define PSY_SNR_25DB 3.1622776e-3f /* -25dB */
65 #define PSY_3GPP_SAVE_SLOPE_L -0.46666667f
66 #define PSY_3GPP_SAVE_SLOPE_S -0.36363637f
67 #define PSY_3GPP_SAVE_ADD_L -0.84285712f
68 #define PSY_3GPP_SAVE_ADD_S -0.75f
69 #define PSY_3GPP_SPEND_SLOPE_L 0.66666669f
70 #define PSY_3GPP_SPEND_SLOPE_S 0.81818181f
71 #define PSY_3GPP_SPEND_ADD_L -0.35f
72 #define PSY_3GPP_SPEND_ADD_S -0.26111111f
73 #define PSY_3GPP_CLIP_LO_L 0.2f
74 #define PSY_3GPP_CLIP_LO_S 0.2f
75 #define PSY_3GPP_CLIP_HI_L 0.95f
76 #define PSY_3GPP_CLIP_HI_S 0.75f
78 #define PSY_3GPP_AH_THR_LONG 0.5f
79 #define PSY_3GPP_AH_THR_SHORT 0.63f
87 #define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f)
89 /* LAME psy model constants */
90 #define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order
91 #define AAC_BLOCK_SIZE_LONG 1024 ///< long block size
92 #define AAC_BLOCK_SIZE_SHORT 128 ///< short block size
93 #define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence
94 #define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block
101 * information for single band used by 3GPP TS26.403-inspired psychoacoustic model
103 typedef struct AacPsyBand{
104 float energy; ///< band energy
105 float thr; ///< energy threshold
106 float thr_quiet; ///< threshold in quiet
107 float nz_lines; ///< number of non-zero spectral lines
108 float active_lines; ///< number of active spectral lines
109 float pe; ///< perceptual entropy
110 float pe_const; ///< constant part of the PE calculation
111 float norm_fac; ///< normalization factor for linearization
112 int avoid_holes; ///< hole avoidance flag
116 * single/pair channel context for psychoacoustic model
118 typedef struct AacPsyChannel{
119 AacPsyBand band[128]; ///< bands information
120 AacPsyBand prev_band[128]; ///< bands information from the previous frame
122 float win_energy; ///< sliding average of channel energy
123 float iir_state[2]; ///< hi-pass IIR filter state
124 uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence)
125 enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame
126 /* LAME psy model specific members */
127 float attack_threshold; ///< attack threshold for this channel
128 float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS];
129 int prev_attack; ///< attack value for the last short block in the previous sequence
133 * psychoacoustic model frame type-dependent coefficients
135 typedef struct AacPsyCoeffs{
136 float ath; ///< absolute threshold of hearing per bands
137 float barks; ///< Bark value for each spectral band in long frame
138 float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame
139 float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame
140 float min_snr; ///< minimal SNR
144 * 3GPP TS26.403-inspired psychoacoustic model specific data
146 typedef struct AacPsyContext{
147 int chan_bitrate; ///< bitrate per channel
148 int frame_bits; ///< average bits per frame
149 int fill_level; ///< bit reservoir fill level
151 float min; ///< minimum allowed PE for bit factor calculation
152 float max; ///< maximum allowed PE for bit factor calculation
153 float previous; ///< allowed PE of the previous frame
154 float correction; ///< PE correction factor
156 AacPsyCoeffs psy_coef[2][64];
161 * LAME psy model preset struct
163 typedef struct PsyLamePreset {
164 int quality; ///< Quality to map the rest of the vaules to.
165 /* This is overloaded to be both kbps per channel in ABR mode, and
166 * requested quality in constant quality mode.
168 float st_lrm; ///< short threshold for L, R, and M channels
172 * LAME psy model preset table for ABR
174 static const PsyLamePreset psy_abr_map[] = {
175 /* TODO: Tuning. These were taken from LAME. */
193 * LAME psy model preset table for constant quality
195 static const PsyLamePreset psy_vbr_map[] = {
211 * LAME psy model FIR coefficient table
213 static const float psy_fir_coeffs[] = {
214 -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
215 -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
216 -5.52212e-17 * 2, -0.313819 * 2
220 * Calculate the ABR attack threshold from the above LAME psymodel table.
222 static float lame_calc_attack_threshold(int bitrate)
224 /* Assume max bitrate to start with */
225 int lower_range = 12, upper_range = 12;
226 int lower_range_kbps = psy_abr_map[12].quality;
227 int upper_range_kbps = psy_abr_map[12].quality;
230 /* Determine which bitrates the value specified falls between.
231 * If the loop ends without breaking our above assumption of 320kbps was correct.
233 for (i = 1; i < 13; i++) {
234 if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) {
236 upper_range_kbps = psy_abr_map[i ].quality;
238 lower_range_kbps = psy_abr_map[i - 1].quality;
239 break; /* Upper range found */
243 /* Determine which range the value specified is closer to */
244 if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps))
245 return psy_abr_map[lower_range].st_lrm;
246 return psy_abr_map[upper_range].st_lrm;
250 * LAME psy model specific initialization
252 static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx)
256 for (i = 0; i < avctx->channels; i++) {
257 AacPsyChannel *pch = &ctx->ch[i];
259 if (avctx->flags & CODEC_FLAG_QSCALE)
260 pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm;
262 pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000);
264 for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++)
265 pch->prev_energy_subshort[j] = 10.0f;
270 * Calculate Bark value for given line.
272 static av_cold float calc_bark(float f)
274 return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f));
279 * Calculate ATH value for given frequency.
280 * Borrowed from Lame.
282 static av_cold float ath(float f, float add)
285 return 3.64 * pow(f, -0.8)
286 - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4))
287 + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7))
288 + (0.6 + 0.04 * add) * 0.001 * f * f * f * f;
291 static av_cold int psy_3gpp_init(FFPsyContext *ctx) {
295 float prev, minscale, minath, minsnr, pe_min;
296 const int chan_bitrate = ctx->avctx->bit_rate / ctx->avctx->channels;
297 const int bandwidth = ctx->avctx->cutoff ? ctx->avctx->cutoff : ctx->avctx->sample_rate / 2;
298 const float num_bark = calc_bark((float)bandwidth);
300 ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
301 pctx = (AacPsyContext*) ctx->model_priv_data;
303 pctx->chan_bitrate = chan_bitrate;
304 pctx->frame_bits = chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate;
305 pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
306 pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
307 ctx->bitres.size = 6144 - pctx->frame_bits;
308 ctx->bitres.size -= ctx->bitres.size % 8;
309 pctx->fill_level = ctx->bitres.size;
310 minath = ath(3410 - 0.733 * ATH_ADD, ATH_ADD);
311 for (j = 0; j < 2; j++) {
312 AacPsyCoeffs *coeffs = pctx->psy_coef[j];
313 const uint8_t *band_sizes = ctx->bands[j];
314 float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
315 float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) / ctx->avctx->sample_rate;
316 /* reference encoder uses 2.4% here instead of 60% like the spec says */
317 float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark;
318 float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L;
319 /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */
320 float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1;
324 for (g = 0; g < ctx->num_bands[j]; g++) {
326 bark = calc_bark((i-1) * line_to_frequency);
327 coeffs[g].barks = (bark + prev) / 2.0;
330 for (g = 0; g < ctx->num_bands[j] - 1; g++) {
331 AacPsyCoeffs *coeff = &coeffs[g];
332 float bark_width = coeffs[g+1].barks - coeffs->barks;
333 coeff->spread_low[0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_LOW);
334 coeff->spread_hi [0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_HI);
335 coeff->spread_low[1] = pow(10.0, -bark_width * en_spread_low);
336 coeff->spread_hi [1] = pow(10.0, -bark_width * en_spread_hi);
337 pe_min = bark_pe * bark_width;
338 minsnr = pow(2.0f, pe_min / band_sizes[g]) - 1.5f;
339 coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB);
342 for (g = 0; g < ctx->num_bands[j]; g++) {
343 minscale = ath(start * line_to_frequency, ATH_ADD);
344 for (i = 1; i < band_sizes[g]; i++)
345 minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD));
346 coeffs[g].ath = minscale - minath;
347 start += band_sizes[g];
351 pctx->ch = av_mallocz(sizeof(AacPsyChannel) * ctx->avctx->channels);
353 lame_window_init(pctx, ctx->avctx);
359 * IIR filter used in block switching decision
361 static float iir_filter(int in, float state[2])
365 ret = 0.7548f * (in - state[0]) + 0.5095f * state[1];
372 * window grouping information stored as bits (0 - new group, 1 - group continues)
374 static const uint8_t window_grouping[9] = {
375 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
379 * Tell encoder which window types to use.
380 * @see 3GPP TS26.403 5.4.1 "Blockswitching"
382 static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx,
383 const int16_t *audio,
385 int channel, int prev_type)
388 int br = ctx->avctx->bit_rate / ctx->avctx->channels;
389 int attack_ratio = br <= 16000 ? 18 : 10;
390 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
391 AacPsyChannel *pch = &pctx->ch[channel];
392 uint8_t grouping = 0;
393 int next_type = pch->next_window_seq;
394 FFPsyWindowInfo wi = { { 0 } };
398 int switch_to_eight = 0;
399 float sum = 0.0, sum2 = 0.0;
402 for (i = 0; i < 8; i++) {
403 for (j = 0; j < 128; j++) {
404 v = iir_filter(la[i*128+j], pch->iir_state);
410 for (i = 0; i < 8; i++) {
411 if (s[i] > pch->win_energy * attack_ratio) {
417 pch->win_energy = pch->win_energy*7/8 + sum2/64;
419 wi.window_type[1] = prev_type;
421 case ONLY_LONG_SEQUENCE:
422 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
423 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
425 case LONG_START_SEQUENCE:
426 wi.window_type[0] = EIGHT_SHORT_SEQUENCE;
427 grouping = pch->next_grouping;
428 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
430 case LONG_STOP_SEQUENCE:
431 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
432 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
434 case EIGHT_SHORT_SEQUENCE:
435 stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight;
436 wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
437 grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0;
438 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
442 pch->next_grouping = window_grouping[attack_n];
443 pch->next_window_seq = next_type;
445 for (i = 0; i < 3; i++)
446 wi.window_type[i] = prev_type;
447 grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0;
451 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
457 for (i = 0; i < 8; i++) {
458 if (!((grouping >> i) & 1))
460 wi.grouping[lastgrp]++;
467 /* 5.6.1.2 "Calculation of Bit Demand" */
468 static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size,
471 const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L;
472 const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L;
473 const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L;
474 const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L;
475 const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L;
476 const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L;
477 float clipped_pe, bit_save, bit_spend, bit_factor, fill_level;
479 ctx->fill_level += ctx->frame_bits - bits;
480 ctx->fill_level = av_clip(ctx->fill_level, 0, size);
481 fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high);
482 clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max);
483 bit_save = (fill_level + bitsave_add) * bitsave_slope;
484 assert(bit_save <= 0.3f && bit_save >= -0.05000001f);
485 bit_spend = (fill_level + bitspend_add) * bitspend_slope;
486 assert(bit_spend <= 0.5f && bit_spend >= -0.1f);
487 /* The bit factor graph in the spec is obviously incorrect.
488 * bit_spend + ((bit_spend - bit_spend))...
489 * The reference encoder subtracts everything from 1, but also seems incorrect.
490 * 1 - bit_save + ((bit_spend + bit_save))...
491 * Hopefully below is correct.
493 bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min);
494 /* NOTE: The reference encoder attempts to center pe max/min around the current pe. */
495 ctx->pe.max = FFMAX(pe, ctx->pe.max);
496 ctx->pe.min = FFMIN(pe, ctx->pe.min);
498 return FFMIN(ctx->frame_bits * bit_factor, ctx->frame_bits + size - bits);
501 static float calc_pe_3gpp(AacPsyBand *band)
506 band->pe_const = 0.0f;
507 band->active_lines = 0.0f;
508 if (band->energy > band->thr) {
509 a = log2f(band->energy);
510 pe = a - log2f(band->thr);
511 band->active_lines = band->nz_lines;
512 if (pe < PSY_3GPP_C1) {
513 pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2;
514 a = a * PSY_3GPP_C3 + PSY_3GPP_C2;
515 band->active_lines *= PSY_3GPP_C3;
517 band->pe = pe * band->nz_lines;
518 band->pe_const = a * band->nz_lines;
524 static float calc_reduction_3gpp(float a, float desired_pe, float pe,
527 float thr_avg, reduction;
529 thr_avg = powf(2.0f, (a - pe) / (4.0f * active_lines));
530 reduction = powf(2.0f, (a - desired_pe) / (4.0f * active_lines)) - thr_avg;
532 return FFMAX(reduction, 0.0f);
535 static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr,
538 float thr = band->thr;
540 if (band->energy > thr) {
541 thr = powf(thr, 0.25f) + reduction;
542 thr = powf(thr, 4.0f);
544 /* This deviates from the 3GPP spec to match the reference encoder.
545 * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands
546 * that have hole avoidance on (active or inactive). It always reduces the
547 * threshold of bands with hole avoidance off.
549 if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) {
550 thr = FFMAX(band->thr, band->energy * min_snr);
551 band->avoid_holes = PSY_3GPP_AH_ACTIVE;
559 * Calculate band thresholds as suggested in 3GPP TS26.403
561 static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel,
562 const float *coefs, const FFPsyWindowInfo *wi)
564 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
565 AacPsyChannel *pch = &pctx->ch[channel];
568 float desired_bits, desired_pe, delta_pe, reduction, spread_en[128] = {0};
569 float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
570 float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f);
571 const int num_bands = ctx->num_bands[wi->num_windows == 8];
572 const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8];
573 AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
574 const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG;
576 //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
577 for (w = 0; w < wi->num_windows*16; w += 16) {
578 for (g = 0; g < num_bands; g++) {
579 AacPsyBand *band = &pch->band[w+g];
581 float form_factor = 0.0f;
583 for (i = 0; i < band_sizes[g]; i++) {
584 band->energy += coefs[start+i] * coefs[start+i];
585 form_factor += sqrtf(fabs(coefs[start+i]));
587 band->thr = band->energy * 0.001258925f;
588 band->nz_lines = form_factor / powf(band->energy / band_sizes[g], 0.25f);
590 start += band_sizes[g];
593 //modify thresholds and energies - spread, threshold in quiet, pre-echo control
594 for (w = 0; w < wi->num_windows*16; w += 16) {
595 AacPsyBand *bands = &pch->band[w];
597 /* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */
598 spread_en[0] = bands[0].energy;
599 for (g = 1; g < num_bands; g++) {
600 bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]);
601 spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]);
603 for (g = num_bands - 2; g >= 0; g--) {
604 bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]);
605 spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]);
607 //5.4.2.4 "Threshold in quiet"
608 for (g = 0; g < num_bands; g++) {
609 AacPsyBand *band = &bands[g];
611 band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath);
612 //5.4.2.5 "Pre-echo control"
613 if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (wi->window_type[1] == LONG_START_SEQUENCE && !w)))
614 band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr,
615 PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
617 /* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */
618 pe += calc_pe_3gpp(band);
620 active_lines += band->active_lines;
622 /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */
623 if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f)
624 band->avoid_holes = PSY_3GPP_AH_NONE;
626 band->avoid_holes = PSY_3GPP_AH_INACTIVE;
630 /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */
631 ctx->ch[channel].entropy = pe;
632 desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8);
633 desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits);
634 /* NOTE: PE correction is kept simple. During initial testing it had very
635 * little effect on the final bitrate. Probably a good idea to come
636 * back and do more testing later.
638 if (ctx->bitres.bits > 0)
639 desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits),
641 pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits);
643 if (desired_pe < pe) {
644 /* 5.6.1.3.4 "First Estimation of the reduction value" */
645 for (w = 0; w < wi->num_windows*16; w += 16) {
646 reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines);
650 for (g = 0; g < num_bands; g++) {
651 AacPsyBand *band = &pch->band[w+g];
653 band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
655 pe += calc_pe_3gpp(band);
657 active_lines += band->active_lines;
661 /* 5.6.1.3.5 "Second Estimation of the reduction value" */
662 for (i = 0; i < 2; i++) {
663 float pe_no_ah = 0.0f, desired_pe_no_ah;
664 active_lines = a = 0.0f;
665 for (w = 0; w < wi->num_windows*16; w += 16) {
666 for (g = 0; g < num_bands; g++) {
667 AacPsyBand *band = &pch->band[w+g];
669 if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) {
670 pe_no_ah += band->pe;
672 active_lines += band->active_lines;
676 desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f);
677 if (active_lines > 0.0f)
678 reduction += calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines);
681 for (w = 0; w < wi->num_windows*16; w += 16) {
682 for (g = 0; g < num_bands; g++) {
683 AacPsyBand *band = &pch->band[w+g];
685 if (active_lines > 0.0f)
686 band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
687 pe += calc_pe_3gpp(band);
688 band->norm_fac = band->active_lines / band->thr;
689 norm_fac += band->norm_fac;
692 delta_pe = desired_pe - pe;
693 if (fabs(delta_pe) > 0.05f * desired_pe)
697 if (pe < 1.15f * desired_pe) {
698 /* 6.6.1.3.6 "Final threshold modification by linearization" */
699 norm_fac = 1.0f / norm_fac;
700 for (w = 0; w < wi->num_windows*16; w += 16) {
701 for (g = 0; g < num_bands; g++) {
702 AacPsyBand *band = &pch->band[w+g];
704 if (band->active_lines > 0.5f) {
705 float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe;
706 float thr = band->thr;
708 thr *= powf(2.0f, delta_sfb_pe / band->active_lines);
709 if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE)
710 thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy);
716 /* 5.6.1.3.7 "Further perceptual entropy reduction" */
718 while (pe > desired_pe && g--) {
719 for (w = 0; w < wi->num_windows*16; w+= 16) {
720 AacPsyBand *band = &pch->band[w+g];
721 if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) {
722 coeffs[g].min_snr = PSY_SNR_1DB;
723 band->thr = band->energy * PSY_SNR_1DB;
724 pe += band->active_lines * 1.5f - band->pe;
728 /* TODO: allow more holes (unused without mid/side) */
732 for (w = 0; w < wi->num_windows*16; w += 16) {
733 for (g = 0; g < num_bands; g++) {
734 AacPsyBand *band = &pch->band[w+g];
735 FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g];
737 psy_band->threshold = band->thr;
738 psy_band->energy = band->energy;
742 memcpy(pch->prev_band, pch->band, sizeof(pch->band));
745 static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
746 const float **coeffs, const FFPsyWindowInfo *wi)
749 FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel);
751 for (ch = 0; ch < group->num_ch; ch++)
752 psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]);
755 static av_cold void psy_3gpp_end(FFPsyContext *apc)
757 AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data;
759 av_freep(&apc->model_priv_data);
762 static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
764 int blocktype = ONLY_LONG_SEQUENCE;
766 if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE)
767 blocktype = LONG_STOP_SEQUENCE;
769 blocktype = EIGHT_SHORT_SEQUENCE;
770 if (ctx->next_window_seq == ONLY_LONG_SEQUENCE)
771 ctx->next_window_seq = LONG_START_SEQUENCE;
772 if (ctx->next_window_seq == LONG_STOP_SEQUENCE)
773 ctx->next_window_seq = EIGHT_SHORT_SEQUENCE;
776 wi->window_type[0] = ctx->next_window_seq;
777 ctx->next_window_seq = blocktype;
780 static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio,
781 const float *la, int channel, int prev_type)
783 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
784 AacPsyChannel *pch = &pctx->ch[channel];
786 int uselongblock = 1;
787 int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
789 FFPsyWindowInfo wi = { { 0 } };
792 float hpfsmpl[AAC_BLOCK_SIZE_LONG];
793 float const *pf = hpfsmpl;
794 float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
795 float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
796 float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
797 const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN);
800 /* LAME comment: apply high pass filter of fs/4 */
801 for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
803 sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2];
805 for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
806 sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]);
807 sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]);
809 /* NOTE: The LAME psymodel expects its input in the range -32768 to
810 * 32768. Tuning this for normalized floats would be difficult. */
811 hpfsmpl[i] = (sum1 + sum2) * 32768.0f;
814 /* Calculate the energies of each sub-shortblock */
815 for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) {
816 energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)];
817 assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0);
818 attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)];
819 energy_short[0] += energy_subshort[i];
822 for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) {
823 float const *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS);
825 for (; pf < pfe; pf++)
826 p = FFMAX(p, fabsf(*pf));
827 pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
828 energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
829 /* NOTE: The indexes below are [i + 3 - 2] in the LAME source.
830 * Obviously the 3 and 2 have some significance, or this would be just [i + 1]
831 * (which is what we use here). What the 3 stands for is ambiguous, as it is both
832 * number of short blocks, and the number of sub-short blocks.
833 * It seems that LAME is comparing each sub-block to sub-block + 1 in the
836 if (p > energy_subshort[i + 1])
837 p = p / energy_subshort[i + 1];
838 else if (energy_subshort[i + 1] > p * 10.0f)
839 p = energy_subshort[i + 1] / (p * 10.0f);
842 attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
845 /* compare energy between sub-short blocks */
846 for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++)
847 if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS])
848 if (attack_intensity[i] > pch->attack_threshold)
849 attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1;
851 /* should have energy change between short blocks, in order to avoid periodic signals */
852 /* Good samples to show the effect are Trumpet test songs */
853 /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */
854 /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */
855 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) {
856 float const u = energy_short[i - 1];
857 float const v = energy_short[i];
858 float const m = FFMAX(u, v);
859 if (m < 40000) { /* (2) */
860 if (u < 1.7f * v && v < 1.7f * u) { /* (1) */
861 if (i == 1 && attacks[0] < attacks[i])
866 att_sum += attacks[i];
869 if (attacks[0] <= pch->prev_attack)
872 att_sum += attacks[0];
873 /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */
874 if (pch->prev_attack == 3 || att_sum) {
877 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++)
878 if (attacks[i] && attacks[i-1])
882 /* We have no lookahead info, so just use same type as the previous sequence. */
883 uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE);
886 lame_apply_block_type(pch, &wi, uselongblock);
888 wi.window_type[1] = prev_type;
889 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
892 if (wi.window_type[0] == LONG_START_SEQUENCE)
901 for (i = 0; i < 8; i++) {
902 if (!((pch->next_grouping >> i) & 1))
904 wi.grouping[lastgrp]++;
908 /* Determine grouping, based on the location of the first attack, and save for
910 * FIXME: Move this to analysis.
911 * TODO: Tune groupings depending on attack location
912 * TODO: Handle more than one attack in a group
914 for (i = 0; i < 9; i++) {
920 pch->next_grouping = window_grouping[grouping];
922 pch->prev_attack = attacks[8];
927 const FFPsyModel ff_aac_psy_model =
929 .name = "3GPP TS 26.403-inspired model",
930 .init = psy_3gpp_init,
931 .window = psy_lame_window,
932 .analyze = psy_3gpp_analyze,