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 * try other bitrate controlling mechanism (maybe use ratecontrol.c?)
34 * control quality for quality-based output
35 **********************************/
38 * constants for 3GPP AAC psychoacoustic model
41 #define PSY_3GPP_THR_SPREAD_HI 1.5f // spreading factor for low-to-hi threshold spreading (15 dB/Bark)
42 #define PSY_3GPP_THR_SPREAD_LOW 3.0f // spreading factor for hi-to-low threshold spreading (30 dB/Bark)
43 /* spreading factor for low-to-hi energy spreading, long block, > 22kbps/channel (20dB/Bark) */
44 #define PSY_3GPP_EN_SPREAD_HI_L1 2.0f
45 /* spreading factor for low-to-hi energy spreading, long block, <= 22kbps/channel (15dB/Bark) */
46 #define PSY_3GPP_EN_SPREAD_HI_L2 1.5f
47 /* spreading factor for low-to-hi energy spreading, short block (15 dB/Bark) */
48 #define PSY_3GPP_EN_SPREAD_HI_S 1.5f
49 /* spreading factor for hi-to-low energy spreading, long block (30dB/Bark) */
50 #define PSY_3GPP_EN_SPREAD_LOW_L 3.0f
51 /* spreading factor for hi-to-low energy spreading, short block (20dB/Bark) */
52 #define PSY_3GPP_EN_SPREAD_LOW_S 2.0f
54 #define PSY_3GPP_RPEMIN 0.01f
55 #define PSY_3GPP_RPELEV 2.0f
57 #define PSY_3GPP_C1 3.0f /* log2(8) */
58 #define PSY_3GPP_C2 1.3219281f /* log2(2.5) */
59 #define PSY_3GPP_C3 0.55935729f /* 1 - C2 / C1 */
61 #define PSY_SNR_1DB 7.9432821e-1f /* -1dB */
62 #define PSY_SNR_25DB 3.1622776e-3f /* -25dB */
64 #define PSY_3GPP_SAVE_SLOPE_L -0.46666667f
65 #define PSY_3GPP_SAVE_SLOPE_S -0.36363637f
66 #define PSY_3GPP_SAVE_ADD_L -0.84285712f
67 #define PSY_3GPP_SAVE_ADD_S -0.75f
68 #define PSY_3GPP_SPEND_SLOPE_L 0.66666669f
69 #define PSY_3GPP_SPEND_SLOPE_S 0.81818181f
70 #define PSY_3GPP_SPEND_ADD_L -0.35f
71 #define PSY_3GPP_SPEND_ADD_S -0.26111111f
72 #define PSY_3GPP_CLIP_LO_L 0.2f
73 #define PSY_3GPP_CLIP_LO_S 0.2f
74 #define PSY_3GPP_CLIP_HI_L 0.95f
75 #define PSY_3GPP_CLIP_HI_S 0.75f
77 #define PSY_3GPP_AH_THR_LONG 0.5f
78 #define PSY_3GPP_AH_THR_SHORT 0.63f
86 #define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f)
88 /* LAME psy model constants */
89 #define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order
90 #define AAC_BLOCK_SIZE_LONG 1024 ///< long block size
91 #define AAC_BLOCK_SIZE_SHORT 128 ///< short block size
92 #define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence
93 #define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block
100 * information for single band used by 3GPP TS26.403-inspired psychoacoustic model
102 typedef struct AacPsyBand{
103 float energy; ///< band energy
104 float thr; ///< energy threshold
105 float thr_quiet; ///< threshold in quiet
106 float nz_lines; ///< number of non-zero spectral lines
107 float active_lines; ///< number of active spectral lines
108 float pe; ///< perceptual entropy
109 float pe_const; ///< constant part of the PE calculation
110 float norm_fac; ///< normalization factor for linearization
111 int avoid_holes; ///< hole avoidance flag
115 * single/pair channel context for psychoacoustic model
117 typedef struct AacPsyChannel{
118 AacPsyBand band[128]; ///< bands information
119 AacPsyBand prev_band[128]; ///< bands information from the previous frame
121 float win_energy; ///< sliding average of channel energy
122 float iir_state[2]; ///< hi-pass IIR filter state
123 uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence)
124 enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame
125 /* LAME psy model specific members */
126 float attack_threshold; ///< attack threshold for this channel
127 float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS];
128 int prev_attack; ///< attack value for the last short block in the previous sequence
132 * psychoacoustic model frame type-dependent coefficients
134 typedef struct AacPsyCoeffs{
135 float ath; ///< absolute threshold of hearing per bands
136 float barks; ///< Bark value for each spectral band in long frame
137 float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame
138 float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame
139 float min_snr; ///< minimal SNR
143 * 3GPP TS26.403-inspired psychoacoustic model specific data
145 typedef struct AacPsyContext{
146 int chan_bitrate; ///< bitrate per channel
147 int frame_bits; ///< average bits per frame
148 int fill_level; ///< bit reservoir fill level
150 float min; ///< minimum allowed PE for bit factor calculation
151 float max; ///< maximum allowed PE for bit factor calculation
152 float previous; ///< allowed PE of the previous frame
153 float correction; ///< PE correction factor
155 AacPsyCoeffs psy_coef[2][64];
160 * LAME psy model preset struct
163 int quality; ///< Quality to map the rest of the vaules to.
164 /* This is overloaded to be both kbps per channel in ABR mode, and
165 * requested quality in constant quality mode.
167 float st_lrm; ///< short threshold for L, R, and M channels
171 * LAME psy model preset table for ABR
173 static const PsyLamePreset psy_abr_map[] = {
174 /* TODO: Tuning. These were taken from LAME. */
192 * LAME psy model preset table for constant quality
194 static const PsyLamePreset psy_vbr_map[] = {
210 * LAME psy model FIR coefficient table
212 static const float psy_fir_coeffs[] = {
213 -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
214 -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
215 -5.52212e-17 * 2, -0.313819 * 2
219 * calculates the attack threshold for ABR from the above table for the LAME psy model
221 static float lame_calc_attack_threshold(int bitrate)
223 /* Assume max bitrate to start with */
224 int lower_range = 12, upper_range = 12;
225 int lower_range_kbps = psy_abr_map[12].quality;
226 int upper_range_kbps = psy_abr_map[12].quality;
229 /* Determine which bitrates the value specified falls between.
230 * If the loop ends without breaking our above assumption of 320kbps was correct.
232 for (i = 1; i < 13; i++) {
233 if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) {
235 upper_range_kbps = psy_abr_map[i ].quality;
237 lower_range_kbps = psy_abr_map[i - 1].quality;
238 break; /* Upper range found */
242 /* Determine which range the value specified is closer to */
243 if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps))
244 return psy_abr_map[lower_range].st_lrm;
245 return psy_abr_map[upper_range].st_lrm;
249 * LAME psy model specific initialization
251 static void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx) {
254 for (i = 0; i < avctx->channels; i++) {
255 AacPsyChannel *pch = &ctx->ch[i];
257 if (avctx->flags & CODEC_FLAG_QSCALE)
258 pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm;
260 pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000);
262 for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++)
263 pch->prev_energy_subshort[j] = 10.0f;
268 * Calculate Bark value for given line.
270 static av_cold float calc_bark(float f)
272 return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f));
277 * Calculate ATH value for given frequency.
278 * Borrowed from Lame.
280 static av_cold float ath(float f, float add)
283 return 3.64 * pow(f, -0.8)
284 - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4))
285 + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7))
286 + (0.6 + 0.04 * add) * 0.001 * f * f * f * f;
289 static av_cold int psy_3gpp_init(FFPsyContext *ctx) {
293 float prev, minscale, minath, minsnr, pe_min;
294 const int chan_bitrate = ctx->avctx->bit_rate / ctx->avctx->channels;
295 const int bandwidth = ctx->avctx->cutoff ? ctx->avctx->cutoff : ctx->avctx->sample_rate / 2;
296 const float num_bark = calc_bark((float)bandwidth);
298 ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
299 pctx = (AacPsyContext*) ctx->model_priv_data;
301 pctx->chan_bitrate = chan_bitrate;
302 pctx->frame_bits = chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate;
303 pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
304 pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
305 ctx->bitres.size = 6144 - pctx->frame_bits;
306 ctx->bitres.size -= ctx->bitres.size % 8;
307 pctx->fill_level = ctx->bitres.size;
308 minath = ath(3410, ATH_ADD);
309 for (j = 0; j < 2; j++) {
310 AacPsyCoeffs *coeffs = pctx->psy_coef[j];
311 const uint8_t *band_sizes = ctx->bands[j];
312 float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
313 float avg_chan_bits = chan_bitrate / ctx->avctx->sample_rate * (j ? 128.0f : 1024.0f);
314 /* reference encoder uses 2.4% here instead of 60% like the spec says */
315 float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark;
316 float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L;
317 /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */
318 float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1;
322 for (g = 0; g < ctx->num_bands[j]; g++) {
324 bark = calc_bark((i-1) * line_to_frequency);
325 coeffs[g].barks = (bark + prev) / 2.0;
328 for (g = 0; g < ctx->num_bands[j] - 1; g++) {
329 AacPsyCoeffs *coeff = &coeffs[g];
330 float bark_width = coeffs[g+1].barks - coeffs->barks;
331 coeff->spread_low[0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_LOW);
332 coeff->spread_hi [0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_HI);
333 coeff->spread_low[1] = pow(10.0, -bark_width * en_spread_low);
334 coeff->spread_hi [1] = pow(10.0, -bark_width * en_spread_hi);
335 pe_min = bark_pe * bark_width;
336 minsnr = pow(2.0f, pe_min / band_sizes[g]) - 1.5f;
337 coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB);
340 for (g = 0; g < ctx->num_bands[j]; g++) {
341 minscale = ath(start * line_to_frequency, ATH_ADD);
342 for (i = 1; i < band_sizes[g]; i++)
343 minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD));
344 coeffs[g].ath = minscale - minath;
345 start += band_sizes[g];
349 pctx->ch = av_mallocz(sizeof(AacPsyChannel) * ctx->avctx->channels);
351 lame_window_init(pctx, ctx->avctx);
357 * IIR filter used in block switching decision
359 static float iir_filter(int in, float state[2])
363 ret = 0.7548f * (in - state[0]) + 0.5095f * state[1];
370 * window grouping information stored as bits (0 - new group, 1 - group continues)
372 static const uint8_t window_grouping[9] = {
373 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
377 * Tell encoder which window types to use.
378 * @see 3GPP TS26.403 5.4.1 "Blockswitching"
380 static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx,
381 const int16_t *audio,
383 int channel, int prev_type)
386 int br = ctx->avctx->bit_rate / ctx->avctx->channels;
387 int attack_ratio = br <= 16000 ? 18 : 10;
388 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
389 AacPsyChannel *pch = &pctx->ch[channel];
390 uint8_t grouping = 0;
391 int next_type = pch->next_window_seq;
394 memset(&wi, 0, sizeof(wi));
397 int switch_to_eight = 0;
398 float sum = 0.0, sum2 = 0.0;
401 for (i = 0; i < 8; i++) {
402 for (j = 0; j < 128; j++) {
403 v = iir_filter(la[(i*128+j)*ctx->avctx->channels], pch->iir_state);
409 for (i = 0; i < 8; i++) {
410 if (s[i] > pch->win_energy * attack_ratio) {
416 pch->win_energy = pch->win_energy*7/8 + sum2/64;
418 wi.window_type[1] = prev_type;
420 case ONLY_LONG_SEQUENCE:
421 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
422 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
424 case LONG_START_SEQUENCE:
425 wi.window_type[0] = EIGHT_SHORT_SEQUENCE;
426 grouping = pch->next_grouping;
427 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
429 case LONG_STOP_SEQUENCE:
430 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
431 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
433 case EIGHT_SHORT_SEQUENCE:
434 stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight;
435 wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
436 grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0;
437 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
441 pch->next_grouping = window_grouping[attack_n];
442 pch->next_window_seq = next_type;
444 for (i = 0; i < 3; i++)
445 wi.window_type[i] = prev_type;
446 grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0;
450 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
456 for (i = 0; i < 8; i++) {
457 if (!((grouping >> i) & 1))
459 wi.grouping[lastgrp]++;
466 /* 5.6.1.2 "Calculation of Bit Demand" */
467 static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size,
470 const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L;
471 const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L;
472 const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L;
473 const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L;
474 const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L;
475 const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L;
476 float clipped_pe, bit_save, bit_spend, bit_factor, fill_level;
478 ctx->fill_level += ctx->frame_bits - bits;
479 ctx->fill_level = av_clip(ctx->fill_level, 0, size);
480 fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high);
481 clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max);
482 bit_save = (fill_level + bitsave_add) * bitsave_slope;
483 assert(bit_save <= 0.3f && bit_save >= -0.05000001f);
484 bit_spend = (fill_level + bitspend_add) * bitspend_slope;
485 assert(bit_spend <= 0.5f && bit_spend >= -0.1f);
486 /* The bit factor graph in the spec is obviously incorrect.
487 * bit_spend + ((bit_spend - bit_spend))...
488 * The reference encoder subtracts everything from 1, but also seems incorrect.
489 * 1 - bit_save + ((bit_spend + bit_save))...
490 * Hopefully below is correct.
492 bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min);
493 /* NOTE: The reference encoder attempts to center pe max/min around the current pe. */
494 ctx->pe.max = FFMAX(pe, ctx->pe.max);
495 ctx->pe.min = FFMIN(pe, ctx->pe.min);
497 return FFMIN(ctx->frame_bits * bit_factor, ctx->frame_bits + size - bits);
500 static float calc_pe_3gpp(AacPsyBand *band)
505 band->pe_const = 0.0f;
506 band->active_lines = 0.0f;
507 if (band->energy > band->thr) {
508 a = log2f(band->energy);
509 pe = a - log2f(band->thr);
510 band->active_lines = band->nz_lines;
511 if (pe < PSY_3GPP_C1) {
512 pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2;
513 a = a * PSY_3GPP_C3 + PSY_3GPP_C2;
514 band->active_lines *= PSY_3GPP_C3;
516 band->pe = pe * band->nz_lines;
517 band->pe_const = a * band->nz_lines;
523 static float calc_reduction_3gpp(float a, float desired_pe, float pe,
526 float thr_avg, reduction;
528 thr_avg = powf(2.0f, (a - pe) / (4.0f * active_lines));
529 reduction = powf(2.0f, (a - desired_pe) / (4.0f * active_lines)) - thr_avg;
531 return FFMAX(reduction, 0.0f);
534 static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr,
537 float thr = band->thr;
539 if (band->energy > thr) {
540 thr = powf(thr, 0.25f) + reduction;
541 thr = powf(thr, 4.0f);
543 /* This deviates from the 3GPP spec to match the reference encoder.
544 * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands
545 * that have hole avoidance on (active or inactive). It always reduces the
546 * threshold of bands with hole avoidance off.
548 if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) {
549 thr = FFMAX(band->thr, band->energy * min_snr);
550 band->avoid_holes = PSY_3GPP_AH_ACTIVE;
558 * Calculate band thresholds as suggested in 3GPP TS26.403
560 static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel,
561 const float *coefs, const FFPsyWindowInfo *wi)
563 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
564 AacPsyChannel *pch = &pctx->ch[channel];
567 float desired_bits, desired_pe, delta_pe, reduction, spread_en[128] = {0};
568 float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
569 float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f);
570 const int num_bands = ctx->num_bands[wi->num_windows == 8];
571 const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8];
572 AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
573 const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG;
575 //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
576 for (w = 0; w < wi->num_windows*16; w += 16) {
577 for (g = 0; g < num_bands; g++) {
578 AacPsyBand *band = &pch->band[w+g];
580 float form_factor = 0.0f;
582 for (i = 0; i < band_sizes[g]; i++) {
583 band->energy += coefs[start+i] * coefs[start+i];
584 form_factor += sqrtf(fabs(coefs[start+i]));
586 band->thr = band->energy * 0.001258925f;
587 band->nz_lines = form_factor / powf(band->energy / band_sizes[g], 0.25f);
589 start += band_sizes[g];
592 //modify thresholds and energies - spread, threshold in quiet, pre-echo control
593 for (w = 0; w < wi->num_windows*16; w += 16) {
594 AacPsyBand *bands = &pch->band[w];
596 //5.4.2.3 "Spreading" & 5.4.3 "Spreaded Energy Calculation"
597 spread_en[0] = bands[0].energy;
598 for (g = 1; g < num_bands; g++) {
599 bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]);
600 spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]);
602 for (g = num_bands - 2; g >= 0; g--) {
603 bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]);
604 spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]);
606 //5.4.2.4 "Threshold in quiet"
607 for (g = 0; g < num_bands; g++) {
608 AacPsyBand *band = &bands[g];
610 band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath);
611 //5.4.2.5 "Pre-echo control"
612 if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (wi->window_type[1] == LONG_START_SEQUENCE && !w)))
613 band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr,
614 PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
616 /* 5.6.1.3.1 "Prepatory steps of the perceptual entropy calculation" */
617 pe += calc_pe_3gpp(band);
619 active_lines += band->active_lines;
621 /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */
622 if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f)
623 band->avoid_holes = PSY_3GPP_AH_NONE;
625 band->avoid_holes = PSY_3GPP_AH_INACTIVE;
629 /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */
630 ctx->ch[channel].entropy = pe;
631 desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8);
632 desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits);
633 /* NOTE: PE correction is kept simple. During initial testing it had very
634 * little effect on the final bitrate. Probably a good idea to come
635 * back and do more testing later.
637 if (ctx->bitres.bits > 0)
638 desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits),
640 pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits);
642 if (desired_pe < pe) {
643 /* 5.6.1.3.4 "First Estimation of the reduction value" */
644 for (w = 0; w < wi->num_windows*16; w += 16) {
645 reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines);
649 for (g = 0; g < num_bands; g++) {
650 AacPsyBand *band = &pch->band[w+g];
652 band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
654 pe += calc_pe_3gpp(band);
656 active_lines += band->active_lines;
660 /* 5.6.1.3.5 "Second Estimation of the reduction value" */
661 for (i = 0; i < 2; i++) {
662 float pe_no_ah = 0.0f, desired_pe_no_ah;
663 active_lines = a = 0.0f;
664 for (w = 0; w < wi->num_windows*16; w += 16) {
665 for (g = 0; g < num_bands; g++) {
666 AacPsyBand *band = &pch->band[w+g];
668 if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) {
669 pe_no_ah += band->pe;
671 active_lines += band->active_lines;
675 desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f);
676 if (active_lines > 0.0f)
677 reduction += calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines);
680 for (w = 0; w < wi->num_windows*16; w += 16) {
681 for (g = 0; g < num_bands; g++) {
682 AacPsyBand *band = &pch->band[w+g];
684 if (active_lines > 0.0f)
685 band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
686 pe += calc_pe_3gpp(band);
687 band->norm_fac = band->active_lines / band->thr;
688 norm_fac += band->norm_fac;
691 delta_pe = desired_pe - pe;
692 if (fabs(delta_pe) > 0.05f * desired_pe)
696 if (pe < 1.15f * desired_pe) {
697 /* 6.6.1.3.6 "Final threshold modification by linearization" */
698 norm_fac = 1.0f / norm_fac;
699 for (w = 0; w < wi->num_windows*16; w += 16) {
700 for (g = 0; g < num_bands; g++) {
701 AacPsyBand *band = &pch->band[w+g];
703 if (band->active_lines > 0.5f) {
704 float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe;
705 float thr = band->thr;
707 thr *= powf(2.0f, delta_sfb_pe / band->active_lines);
708 if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE)
709 thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy);
715 /* 5.6.1.3.7 "Further perceptual entropy reduction" */
717 while (pe > desired_pe && g--) {
718 for (w = 0; w < wi->num_windows*16; w+= 16) {
719 AacPsyBand *band = &pch->band[w+g];
720 if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) {
721 coeffs[g].min_snr = PSY_SNR_1DB;
722 band->thr = band->energy * PSY_SNR_1DB;
723 pe += band->active_lines * 1.5f - band->pe;
727 /* TODO: allow more holes (unused without mid/side) */
731 for (w = 0; w < wi->num_windows*16; w += 16) {
732 for (g = 0; g < num_bands; g++) {
733 AacPsyBand *band = &pch->band[w+g];
734 FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g];
736 psy_band->threshold = band->thr;
737 psy_band->energy = band->energy;
741 memcpy(pch->prev_band, pch->band, sizeof(pch->band));
744 static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
745 const float **coeffs, const FFPsyWindowInfo *wi)
748 FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel);
750 for (ch = 0; ch < group->num_ch; ch++)
751 psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]);
754 static av_cold void psy_3gpp_end(FFPsyContext *apc)
756 AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data;
758 av_freep(&apc->model_priv_data);
761 static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
763 int blocktype = ONLY_LONG_SEQUENCE;
765 if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE)
766 blocktype = LONG_STOP_SEQUENCE;
768 blocktype = EIGHT_SHORT_SEQUENCE;
769 if (ctx->next_window_seq == ONLY_LONG_SEQUENCE)
770 ctx->next_window_seq = LONG_START_SEQUENCE;
771 if (ctx->next_window_seq == LONG_STOP_SEQUENCE)
772 ctx->next_window_seq = EIGHT_SHORT_SEQUENCE;
775 wi->window_type[0] = ctx->next_window_seq;
776 ctx->next_window_seq = blocktype;
779 static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx,
780 const int16_t *audio, const int16_t *la,
781 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 };
791 memset(&wi, 0, sizeof(wi));
793 float hpfsmpl[AAC_BLOCK_SIZE_LONG];
794 float const *pf = hpfsmpl;
795 float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
796 float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
797 float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
798 int chans = ctx->avctx->channels;
799 const int16_t *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN) * chans;
802 /* LAME comment: apply high pass filter of fs/4 */
803 for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
805 sum1 = firbuf[(i + ((PSY_LAME_FIR_LEN - 1) / 2)) * chans];
807 for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
808 sum1 += psy_fir_coeffs[j] * (firbuf[(i + j) * chans] + firbuf[(i + PSY_LAME_FIR_LEN - j) * chans]);
809 sum2 += psy_fir_coeffs[j + 1] * (firbuf[(i + j + 1) * chans] + firbuf[(i + PSY_LAME_FIR_LEN - j - 1) * chans]);
811 hpfsmpl[i] = sum1 + sum2;
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++)
828 pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
829 energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
830 /* FIXME: The indexes below are [i + 3 - 2] in the LAME source.
831 * Obviously the 3 and 2 have some significance, or this would be just [i + 1]
832 * (which is what we use here). What the 3 stands for is ambigious, as it is both
833 * number of short blocks, and the number of sub-short blocks.
834 * It seems that LAME is comparing each sub-block to sub-block + 1 in the
837 if (p > energy_subshort[i + 1])
838 p = p / energy_subshort[i + 1];
839 else if (energy_subshort[i + 1] > p * 10.0f)
840 p = energy_subshort[i + 1] / (p * 10.0f);
843 attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
846 /* compare energy between sub-short blocks */
847 for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++)
848 if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS])
849 if (attack_intensity[i] > pch->attack_threshold)
850 attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1;
852 /* should have energy change between short blocks, in order to avoid periodic signals */
853 /* Good samples to show the effect are Trumpet test songs */
854 /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */
855 /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */
856 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) {
857 float const u = energy_short[i - 1];
858 float const v = energy_short[i];
859 float const m = FFMAX(u, v);
860 if (m < 40000) { /* (2) */
861 if (u < 1.7f * v && v < 1.7f * u) { /* (1) */
862 if (i == 1 && attacks[0] < attacks[i])
867 att_sum += attacks[i];
870 if (attacks[0] <= pch->prev_attack)
873 att_sum += attacks[0];
874 /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */
875 if (pch->prev_attack == 3 || att_sum) {
878 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++)
879 if (attacks[i] && attacks[i-1])
883 /* We have no lookahead info, so just use same type as the previous sequence. */
884 uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE);
887 lame_apply_block_type(pch, &wi, uselongblock);
889 wi.window_type[1] = prev_type;
890 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
893 if (wi.window_type[0] == LONG_START_SEQUENCE)
902 for (i = 0; i < 8; i++) {
903 if (!((pch->next_grouping >> i) & 1))
905 wi.grouping[lastgrp]++;
909 /* Determine grouping, based on the location of the first attack, and save for
911 * FIXME: Move this to analysis.
912 * TODO: Tune groupings depending on attack location
913 * TODO: Handle more than one attack in a group
915 for (i = 0; i < 9; i++) {
921 pch->next_grouping = window_grouping[grouping];
923 pch->prev_attack = attacks[8];
928 const FFPsyModel ff_aac_psy_model =
930 .name = "3GPP TS 26.403-inspired model",
931 .init = psy_3gpp_init,
932 .window = psy_lame_window,
933 .analyze = psy_3gpp_analyze,