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
27 #include "libavutil/attributes.h"
28 #include "libavutil/internal.h"
29 #include "libavutil/libm.h"
35 /***********************************
37 * try other bitrate controlling mechanism (maybe use ratecontrol.c?)
38 * control quality for quality-based output
39 **********************************/
42 * constants for 3GPP AAC psychoacoustic model
45 #define PSY_3GPP_THR_SPREAD_HI 1.5f // spreading factor for low-to-hi threshold spreading (15 dB/Bark)
46 #define PSY_3GPP_THR_SPREAD_LOW 3.0f // spreading factor for hi-to-low threshold spreading (30 dB/Bark)
47 /* spreading factor for low-to-hi energy spreading, long block, > 22kbps/channel (20dB/Bark) */
48 #define PSY_3GPP_EN_SPREAD_HI_L1 2.0f
49 /* spreading factor for low-to-hi energy spreading, long block, <= 22kbps/channel (15dB/Bark) */
50 #define PSY_3GPP_EN_SPREAD_HI_L2 1.5f
51 /* spreading factor for low-to-hi energy spreading, short block (15 dB/Bark) */
52 #define PSY_3GPP_EN_SPREAD_HI_S 1.5f
53 /* spreading factor for hi-to-low energy spreading, long block (30dB/Bark) */
54 #define PSY_3GPP_EN_SPREAD_LOW_L 3.0f
55 /* spreading factor for hi-to-low energy spreading, short block (20dB/Bark) */
56 #define PSY_3GPP_EN_SPREAD_LOW_S 2.0f
58 #define PSY_3GPP_RPEMIN 0.01f
59 #define PSY_3GPP_RPELEV 2.0f
61 #define PSY_3GPP_C1 3.0f /* log2(8) */
62 #define PSY_3GPP_C2 1.3219281f /* log2(2.5) */
63 #define PSY_3GPP_C3 0.55935729f /* 1 - C2 / C1 */
65 #define PSY_SNR_1DB 7.9432821e-1f /* -1dB */
66 #define PSY_SNR_25DB 3.1622776e-3f /* -25dB */
68 #define PSY_3GPP_SAVE_SLOPE_L -0.46666667f
69 #define PSY_3GPP_SAVE_SLOPE_S -0.36363637f
70 #define PSY_3GPP_SAVE_ADD_L -0.84285712f
71 #define PSY_3GPP_SAVE_ADD_S -0.75f
72 #define PSY_3GPP_SPEND_SLOPE_L 0.66666669f
73 #define PSY_3GPP_SPEND_SLOPE_S 0.81818181f
74 #define PSY_3GPP_SPEND_ADD_L -0.35f
75 #define PSY_3GPP_SPEND_ADD_S -0.26111111f
76 #define PSY_3GPP_CLIP_LO_L 0.2f
77 #define PSY_3GPP_CLIP_LO_S 0.2f
78 #define PSY_3GPP_CLIP_HI_L 0.95f
79 #define PSY_3GPP_CLIP_HI_S 0.75f
81 #define PSY_3GPP_AH_THR_LONG 0.5f
82 #define PSY_3GPP_AH_THR_SHORT 0.63f
84 #define PSY_PE_FORGET_SLOPE 511
92 #define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f)
93 #define PSY_3GPP_PE_TO_BITS(bits) ((bits) / 1.18f)
95 /* LAME psy model constants */
96 #define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order
97 #define AAC_BLOCK_SIZE_LONG 1024 ///< long block size
98 #define AAC_BLOCK_SIZE_SHORT 128 ///< short block size
99 #define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence
100 #define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block
107 * information for single band used by 3GPP TS26.403-inspired psychoacoustic model
109 typedef struct AacPsyBand{
110 float energy; ///< band energy
111 float thr; ///< energy threshold
112 float thr_quiet; ///< threshold in quiet
113 float nz_lines; ///< number of non-zero spectral lines
114 float active_lines; ///< number of active spectral lines
115 float pe; ///< perceptual entropy
116 float pe_const; ///< constant part of the PE calculation
117 float norm_fac; ///< normalization factor for linearization
118 int avoid_holes; ///< hole avoidance flag
122 * single/pair channel context for psychoacoustic model
124 typedef struct AacPsyChannel{
125 AacPsyBand band[128]; ///< bands information
126 AacPsyBand prev_band[128]; ///< bands information from the previous frame
128 float win_energy; ///< sliding average of channel energy
129 float iir_state[2]; ///< hi-pass IIR filter state
130 uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence)
131 enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame
132 /* LAME psy model specific members */
133 float attack_threshold; ///< attack threshold for this channel
134 float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS];
135 int prev_attack; ///< attack value for the last short block in the previous sequence
139 * psychoacoustic model frame type-dependent coefficients
141 typedef struct AacPsyCoeffs{
142 float ath; ///< absolute threshold of hearing per bands
143 float barks; ///< Bark value for each spectral band in long frame
144 float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame
145 float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame
146 float min_snr; ///< minimal SNR
150 * 3GPP TS26.403-inspired psychoacoustic model specific data
152 typedef struct AacPsyContext{
153 int chan_bitrate; ///< bitrate per channel
154 int frame_bits; ///< average bits per frame
155 int fill_level; ///< bit reservoir fill level
157 float min; ///< minimum allowed PE for bit factor calculation
158 float max; ///< maximum allowed PE for bit factor calculation
159 float previous; ///< allowed PE of the previous frame
160 float correction; ///< PE correction factor
162 AacPsyCoeffs psy_coef[2][64];
164 float global_quality; ///< normalized global quality taken from avctx
168 * LAME psy model preset struct
170 typedef struct PsyLamePreset {
171 int quality; ///< Quality to map the rest of the vaules to.
172 /* This is overloaded to be both kbps per channel in ABR mode, and
173 * requested quality in constant quality mode.
175 float st_lrm; ///< short threshold for L, R, and M channels
179 * LAME psy model preset table for ABR
181 static const PsyLamePreset psy_abr_map[] = {
182 /* TODO: Tuning. These were taken from LAME. */
200 * LAME psy model preset table for constant quality
202 static const PsyLamePreset psy_vbr_map[] = {
218 * LAME psy model FIR coefficient table
220 static const float psy_fir_coeffs[] = {
221 -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
222 -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
223 -5.52212e-17 * 2, -0.313819 * 2
227 # include "mips/aacpsy_mips.h"
228 #endif /* ARCH_MIPS */
231 * Calculate the ABR attack threshold from the above LAME psymodel table.
233 static float lame_calc_attack_threshold(int bitrate)
235 /* Assume max bitrate to start with */
236 int lower_range = 12, upper_range = 12;
237 int lower_range_kbps = psy_abr_map[12].quality;
238 int upper_range_kbps = psy_abr_map[12].quality;
241 /* Determine which bitrates the value specified falls between.
242 * If the loop ends without breaking our above assumption of 320kbps was correct.
244 for (i = 1; i < 13; i++) {
245 if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) {
247 upper_range_kbps = psy_abr_map[i ].quality;
249 lower_range_kbps = psy_abr_map[i - 1].quality;
250 break; /* Upper range found */
254 /* Determine which range the value specified is closer to */
255 if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps))
256 return psy_abr_map[lower_range].st_lrm;
257 return psy_abr_map[upper_range].st_lrm;
261 * LAME psy model specific initialization
263 static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx)
267 for (i = 0; i < avctx->channels; i++) {
268 AacPsyChannel *pch = &ctx->ch[i];
270 if (avctx->flags & AV_CODEC_FLAG_QSCALE)
271 pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm;
273 pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000);
275 for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++)
276 pch->prev_energy_subshort[j] = 10.0f;
281 * Calculate Bark value for given line.
283 static av_cold float calc_bark(float f)
285 return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f));
290 * Calculate ATH value for given frequency.
291 * Borrowed from Lame.
293 static av_cold float ath(float f, float add)
296 return 3.64 * pow(f, -0.8)
297 - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4))
298 + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7))
299 + (0.6 + 0.04 * add) * 0.001 * f * f * f * f;
302 static av_cold int psy_3gpp_init(FFPsyContext *ctx) {
306 float prev, minscale, minath, minsnr, pe_min;
307 int chan_bitrate = ctx->avctx->bit_rate / ((ctx->avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : ctx->avctx->channels);
309 const int bandwidth = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx);
310 const float num_bark = calc_bark((float)bandwidth);
312 ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
313 if (!ctx->model_priv_data)
314 return AVERROR(ENOMEM);
315 pctx = (AacPsyContext*) ctx->model_priv_data;
316 pctx->global_quality = (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) * 0.01f;
318 if (ctx->avctx->flags & CODEC_FLAG_QSCALE) {
319 /* Use the target average bitrate to compute spread parameters */
320 chan_bitrate = (int)(chan_bitrate / 120.0 * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120));
323 pctx->chan_bitrate = chan_bitrate;
324 pctx->frame_bits = FFMIN(2560, chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate);
325 pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
326 pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
327 ctx->bitres.size = 6144 - pctx->frame_bits;
328 ctx->bitres.size -= ctx->bitres.size % 8;
329 pctx->fill_level = ctx->bitres.size;
330 minath = ath(3410 - 0.733 * ATH_ADD, ATH_ADD);
331 for (j = 0; j < 2; j++) {
332 AacPsyCoeffs *coeffs = pctx->psy_coef[j];
333 const uint8_t *band_sizes = ctx->bands[j];
334 float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
335 float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) / ctx->avctx->sample_rate;
336 /* reference encoder uses 2.4% here instead of 60% like the spec says */
337 float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark;
338 float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L;
339 /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */
340 float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1;
344 for (g = 0; g < ctx->num_bands[j]; g++) {
346 bark = calc_bark((i-1) * line_to_frequency);
347 coeffs[g].barks = (bark + prev) / 2.0;
350 for (g = 0; g < ctx->num_bands[j] - 1; g++) {
351 AacPsyCoeffs *coeff = &coeffs[g];
352 float bark_width = coeffs[g+1].barks - coeffs->barks;
353 coeff->spread_low[0] = ff_exp10(-bark_width * PSY_3GPP_THR_SPREAD_LOW);
354 coeff->spread_hi [0] = ff_exp10(-bark_width * PSY_3GPP_THR_SPREAD_HI);
355 coeff->spread_low[1] = ff_exp10(-bark_width * en_spread_low);
356 coeff->spread_hi [1] = ff_exp10(-bark_width * en_spread_hi);
357 pe_min = bark_pe * bark_width;
358 minsnr = exp2(pe_min / band_sizes[g]) - 1.5f;
359 coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB);
362 for (g = 0; g < ctx->num_bands[j]; g++) {
363 minscale = ath(start * line_to_frequency, ATH_ADD);
364 for (i = 1; i < band_sizes[g]; i++)
365 minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD));
366 coeffs[g].ath = minscale - minath;
367 start += band_sizes[g];
371 pctx->ch = av_mallocz_array(ctx->avctx->channels, sizeof(AacPsyChannel));
373 av_freep(&ctx->model_priv_data);
374 return AVERROR(ENOMEM);
377 lame_window_init(pctx, ctx->avctx);
383 * IIR filter used in block switching decision
385 static float iir_filter(int in, float state[2])
389 ret = 0.7548f * (in - state[0]) + 0.5095f * state[1];
396 * window grouping information stored as bits (0 - new group, 1 - group continues)
398 static const uint8_t window_grouping[9] = {
399 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
403 * Tell encoder which window types to use.
404 * @see 3GPP TS26.403 5.4.1 "Blockswitching"
406 static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx,
407 const int16_t *audio,
409 int channel, int prev_type)
412 int br = ((AacPsyContext*)ctx->model_priv_data)->chan_bitrate;
413 int attack_ratio = br <= 16000 ? 18 : 10;
414 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
415 AacPsyChannel *pch = &pctx->ch[channel];
416 uint8_t grouping = 0;
417 int next_type = pch->next_window_seq;
418 FFPsyWindowInfo wi = { { 0 } };
422 int switch_to_eight = 0;
423 float sum = 0.0, sum2 = 0.0;
426 for (i = 0; i < 8; i++) {
427 for (j = 0; j < 128; j++) {
428 v = iir_filter(la[i*128+j], pch->iir_state);
434 for (i = 0; i < 8; i++) {
435 if (s[i] > pch->win_energy * attack_ratio) {
441 pch->win_energy = pch->win_energy*7/8 + sum2/64;
443 wi.window_type[1] = prev_type;
445 case ONLY_LONG_SEQUENCE:
446 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
447 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
449 case LONG_START_SEQUENCE:
450 wi.window_type[0] = EIGHT_SHORT_SEQUENCE;
451 grouping = pch->next_grouping;
452 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
454 case LONG_STOP_SEQUENCE:
455 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
456 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
458 case EIGHT_SHORT_SEQUENCE:
459 stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight;
460 wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
461 grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0;
462 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
466 pch->next_grouping = window_grouping[attack_n];
467 pch->next_window_seq = next_type;
469 for (i = 0; i < 3; i++)
470 wi.window_type[i] = prev_type;
471 grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0;
475 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
481 for (i = 0; i < 8; i++) {
482 if (!((grouping >> i) & 1))
484 wi.grouping[lastgrp]++;
491 /* 5.6.1.2 "Calculation of Bit Demand" */
492 static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size,
495 const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L;
496 const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L;
497 const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L;
498 const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L;
499 const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L;
500 const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L;
501 float clipped_pe, bit_save, bit_spend, bit_factor, fill_level, forgetful_min_pe;
503 ctx->fill_level += ctx->frame_bits - bits;
504 ctx->fill_level = av_clip(ctx->fill_level, 0, size);
505 fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high);
506 clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max);
507 bit_save = (fill_level + bitsave_add) * bitsave_slope;
508 assert(bit_save <= 0.3f && bit_save >= -0.05000001f);
509 bit_spend = (fill_level + bitspend_add) * bitspend_slope;
510 assert(bit_spend <= 0.5f && bit_spend >= -0.1f);
511 /* The bit factor graph in the spec is obviously incorrect.
512 * bit_spend + ((bit_spend - bit_spend))...
513 * The reference encoder subtracts everything from 1, but also seems incorrect.
514 * 1 - bit_save + ((bit_spend + bit_save))...
515 * Hopefully below is correct.
517 bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min);
518 /* NOTE: The reference encoder attempts to center pe max/min around the current pe.
519 * Here we do that by slowly forgetting pe.min when pe stays in a range that makes
520 * it unlikely (ie: above the mean)
522 ctx->pe.max = FFMAX(pe, ctx->pe.max);
523 forgetful_min_pe = ((ctx->pe.min * PSY_PE_FORGET_SLOPE)
524 + FFMAX(ctx->pe.min, pe * (pe / ctx->pe.max))) / (PSY_PE_FORGET_SLOPE + 1);
525 ctx->pe.min = FFMIN(pe, forgetful_min_pe);
527 /* NOTE: allocate a minimum of 1/8th average frame bits, to avoid
528 * reservoir starvation from producing zero-bit frames
531 ctx->frame_bits * bit_factor,
532 FFMAX(ctx->frame_bits + size - bits, ctx->frame_bits / 8));
535 static float calc_pe_3gpp(AacPsyBand *band)
540 band->pe_const = 0.0f;
541 band->active_lines = 0.0f;
542 if (band->energy > band->thr) {
543 a = log2f(band->energy);
544 pe = a - log2f(band->thr);
545 band->active_lines = band->nz_lines;
546 if (pe < PSY_3GPP_C1) {
547 pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2;
548 a = a * PSY_3GPP_C3 + PSY_3GPP_C2;
549 band->active_lines *= PSY_3GPP_C3;
551 band->pe = pe * band->nz_lines;
552 band->pe_const = a * band->nz_lines;
558 static float calc_reduction_3gpp(float a, float desired_pe, float pe,
561 float thr_avg, reduction;
563 if(active_lines == 0.0)
566 thr_avg = exp2f((a - pe) / (4.0f * active_lines));
567 reduction = exp2f((a - desired_pe) / (4.0f * active_lines)) - thr_avg;
569 return FFMAX(reduction, 0.0f);
572 static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr,
575 float thr = band->thr;
577 if (band->energy > thr) {
579 thr = sqrtf(thr) + reduction;
583 /* This deviates from the 3GPP spec to match the reference encoder.
584 * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands
585 * that have hole avoidance on (active or inactive). It always reduces the
586 * threshold of bands with hole avoidance off.
588 if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) {
589 thr = FFMAX(band->thr, band->energy * min_snr);
590 band->avoid_holes = PSY_3GPP_AH_ACTIVE;
597 #ifndef calc_thr_3gpp
598 static void calc_thr_3gpp(const FFPsyWindowInfo *wi, const int num_bands, AacPsyChannel *pch,
599 const uint8_t *band_sizes, const float *coefs, const int cutoff)
602 int start = 0, wstart = 0;
603 for (w = 0; w < wi->num_windows*16; w += 16) {
605 for (g = 0; g < num_bands; g++) {
606 AacPsyBand *band = &pch->band[w+g];
608 float form_factor = 0.0f;
611 if (wstart < cutoff) {
612 for (i = 0; i < band_sizes[g]; i++) {
613 band->energy += coefs[start+i] * coefs[start+i];
614 form_factor += sqrtf(fabs(coefs[start+i]));
617 Temp = band->energy > 0 ? sqrtf((float)band_sizes[g] / band->energy) : 0;
618 band->thr = band->energy * 0.001258925f;
619 band->nz_lines = form_factor * sqrtf(Temp);
621 start += band_sizes[g];
622 wstart += band_sizes[g];
626 #endif /* calc_thr_3gpp */
628 #ifndef psy_hp_filter
629 static void psy_hp_filter(const float *firbuf, float *hpfsmpl, const float *psy_fir_coeffs)
632 for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
634 sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2];
636 for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
637 sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]);
638 sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]);
640 /* NOTE: The LAME psymodel expects it's input in the range -32768 to 32768.
641 * Tuning this for normalized floats would be difficult. */
642 hpfsmpl[i] = (sum1 + sum2) * 32768.0f;
645 #endif /* psy_hp_filter */
648 * Calculate band thresholds as suggested in 3GPP TS26.403
650 static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel,
651 const float *coefs, const FFPsyWindowInfo *wi)
653 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
654 AacPsyChannel *pch = &pctx->ch[channel];
656 float desired_bits, desired_pe, delta_pe, reduction= NAN, spread_en[128] = {0};
657 float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
658 float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f);
659 const int num_bands = ctx->num_bands[wi->num_windows == 8];
660 const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8];
661 AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
662 const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG;
663 const int bandwidth = ctx->cutoff ? ctx->cutoff : AAC_CUTOFF(ctx->avctx);
664 const int cutoff = bandwidth * 2048 / wi->num_windows / ctx->avctx->sample_rate;
666 //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
667 calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs, cutoff);
669 //modify thresholds and energies - spread, threshold in quiet, pre-echo control
670 for (w = 0; w < wi->num_windows*16; w += 16) {
671 AacPsyBand *bands = &pch->band[w];
673 /* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */
674 spread_en[0] = bands[0].energy;
675 for (g = 1; g < num_bands; g++) {
676 bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]);
677 spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]);
679 for (g = num_bands - 2; g >= 0; g--) {
680 bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]);
681 spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]);
683 //5.4.2.4 "Threshold in quiet"
684 for (g = 0; g < num_bands; g++) {
685 AacPsyBand *band = &bands[g];
687 band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath);
688 //5.4.2.5 "Pre-echo control"
689 if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (wi->window_type[1] == LONG_START_SEQUENCE && !w)))
690 band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr,
691 PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
693 /* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */
694 pe += calc_pe_3gpp(band);
696 active_lines += band->active_lines;
698 /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */
699 if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f)
700 band->avoid_holes = PSY_3GPP_AH_NONE;
702 band->avoid_holes = PSY_3GPP_AH_INACTIVE;
706 /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */
707 ctx->ch[channel].entropy = pe;
708 if (ctx->avctx->flags & CODEC_FLAG_QSCALE) {
709 /* (2.5 * 120) achieves almost transparent rate, and we want to give
710 * ample room downwards, so we make that equivalent to QSCALE=2.4
712 desired_pe = pe * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) / (2 * 2.5f * 120.0f);
713 desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe));
714 desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping
716 /* PE slope smoothing */
717 if (ctx->bitres.bits > 0) {
718 desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe));
719 desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping
722 pctx->pe.max = FFMAX(pe, pctx->pe.max);
723 pctx->pe.min = FFMIN(pe, pctx->pe.min);
725 desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8);
726 desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits);
728 /* NOTE: PE correction is kept simple. During initial testing it had very
729 * little effect on the final bitrate. Probably a good idea to come
730 * back and do more testing later.
732 if (ctx->bitres.bits > 0)
733 desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits),
736 pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits);
737 ctx->bitres.alloc = desired_bits;
739 if (desired_pe < pe) {
740 /* 5.6.1.3.4 "First Estimation of the reduction value" */
741 for (w = 0; w < wi->num_windows*16; w += 16) {
742 reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines);
746 for (g = 0; g < num_bands; g++) {
747 AacPsyBand *band = &pch->band[w+g];
749 band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
751 pe += calc_pe_3gpp(band);
753 active_lines += band->active_lines;
757 /* 5.6.1.3.5 "Second Estimation of the reduction value" */
758 for (i = 0; i < 2; i++) {
759 float pe_no_ah = 0.0f, desired_pe_no_ah;
760 active_lines = a = 0.0f;
761 for (w = 0; w < wi->num_windows*16; w += 16) {
762 for (g = 0; g < num_bands; g++) {
763 AacPsyBand *band = &pch->band[w+g];
765 if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) {
766 pe_no_ah += band->pe;
768 active_lines += band->active_lines;
772 desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f);
773 if (active_lines > 0.0f)
774 reduction = calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines);
777 for (w = 0; w < wi->num_windows*16; w += 16) {
778 for (g = 0; g < num_bands; g++) {
779 AacPsyBand *band = &pch->band[w+g];
781 if (active_lines > 0.0f)
782 band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
783 pe += calc_pe_3gpp(band);
784 if (band->thr > 0.0f)
785 band->norm_fac = band->active_lines / band->thr;
787 band->norm_fac = 0.0f;
788 norm_fac += band->norm_fac;
791 delta_pe = desired_pe - pe;
792 if (fabs(delta_pe) > 0.05f * desired_pe)
796 if (pe < 1.15f * desired_pe) {
797 /* 6.6.1.3.6 "Final threshold modification by linearization" */
798 norm_fac = 1.0f / norm_fac;
799 for (w = 0; w < wi->num_windows*16; w += 16) {
800 for (g = 0; g < num_bands; g++) {
801 AacPsyBand *band = &pch->band[w+g];
803 if (band->active_lines > 0.5f) {
804 float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe;
805 float thr = band->thr;
807 thr *= exp2f(delta_sfb_pe / band->active_lines);
808 if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE)
809 thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy);
815 /* 5.6.1.3.7 "Further perceptual entropy reduction" */
817 while (pe > desired_pe && g--) {
818 for (w = 0; w < wi->num_windows*16; w+= 16) {
819 AacPsyBand *band = &pch->band[w+g];
820 if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) {
821 coeffs[g].min_snr = PSY_SNR_1DB;
822 band->thr = band->energy * PSY_SNR_1DB;
823 pe += band->active_lines * 1.5f - band->pe;
827 /* TODO: allow more holes (unused without mid/side) */
831 for (w = 0; w < wi->num_windows*16; w += 16) {
832 for (g = 0; g < num_bands; g++) {
833 AacPsyBand *band = &pch->band[w+g];
834 FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g];
836 psy_band->threshold = band->thr;
837 psy_band->energy = band->energy;
838 psy_band->spread = band->active_lines * 2.0f / band_sizes[g];
839 psy_band->bits = PSY_3GPP_PE_TO_BITS(band->pe);
843 memcpy(pch->prev_band, pch->band, sizeof(pch->band));
846 static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
847 const float **coeffs, const FFPsyWindowInfo *wi)
850 FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel);
852 for (ch = 0; ch < group->num_ch; ch++)
853 psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]);
856 static av_cold void psy_3gpp_end(FFPsyContext *apc)
858 AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data;
860 av_freep(&apc->model_priv_data);
863 static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
865 int blocktype = ONLY_LONG_SEQUENCE;
867 if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE)
868 blocktype = LONG_STOP_SEQUENCE;
870 blocktype = EIGHT_SHORT_SEQUENCE;
871 if (ctx->next_window_seq == ONLY_LONG_SEQUENCE)
872 ctx->next_window_seq = LONG_START_SEQUENCE;
873 if (ctx->next_window_seq == LONG_STOP_SEQUENCE)
874 ctx->next_window_seq = EIGHT_SHORT_SEQUENCE;
877 wi->window_type[0] = ctx->next_window_seq;
878 ctx->next_window_seq = blocktype;
881 static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio,
882 const float *la, int channel, int prev_type)
884 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
885 AacPsyChannel *pch = &pctx->ch[channel];
887 int uselongblock = 1;
888 int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
889 float clippings[AAC_NUM_BLOCKS_SHORT];
891 FFPsyWindowInfo wi = { { 0 } };
894 float hpfsmpl[AAC_BLOCK_SIZE_LONG];
895 float const *pf = hpfsmpl;
896 float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
897 float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
898 float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
899 const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN);
902 /* LAME comment: apply high pass filter of fs/4 */
903 psy_hp_filter(firbuf, hpfsmpl, psy_fir_coeffs);
905 /* Calculate the energies of each sub-shortblock */
906 for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) {
907 energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)];
908 assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0);
909 attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)];
910 energy_short[0] += energy_subshort[i];
913 for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) {
914 float const *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS);
916 for (; pf < pfe; pf++)
917 p = FFMAX(p, fabsf(*pf));
918 pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
919 energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
920 /* NOTE: The indexes below are [i + 3 - 2] in the LAME source.
921 * Obviously the 3 and 2 have some significance, or this would be just [i + 1]
922 * (which is what we use here). What the 3 stands for is ambiguous, as it is both
923 * number of short blocks, and the number of sub-short blocks.
924 * It seems that LAME is comparing each sub-block to sub-block + 1 in the
927 if (p > energy_subshort[i + 1])
928 p = p / energy_subshort[i + 1];
929 else if (energy_subshort[i + 1] > p * 10.0f)
930 p = energy_subshort[i + 1] / (p * 10.0f);
933 attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
936 /* compare energy between sub-short blocks */
937 for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++)
938 if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS])
939 if (attack_intensity[i] > pch->attack_threshold)
940 attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1;
942 /* should have energy change between short blocks, in order to avoid periodic signals */
943 /* Good samples to show the effect are Trumpet test songs */
944 /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */
945 /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */
946 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) {
947 float const u = energy_short[i - 1];
948 float const v = energy_short[i];
949 float const m = FFMAX(u, v);
950 if (m < 40000) { /* (2) */
951 if (u < 1.7f * v && v < 1.7f * u) { /* (1) */
952 if (i == 1 && attacks[0] < attacks[i])
957 att_sum += attacks[i];
960 if (attacks[0] <= pch->prev_attack)
963 att_sum += attacks[0];
964 /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */
965 if (pch->prev_attack == 3 || att_sum) {
968 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++)
969 if (attacks[i] && attacks[i-1])
973 /* We have no lookahead info, so just use same type as the previous sequence. */
974 uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE);
977 lame_apply_block_type(pch, &wi, uselongblock);
979 /* Calculate input sample maximums and evaluate clipping risk */
981 for (i = 0; i < AAC_NUM_BLOCKS_SHORT; i++) {
982 const float *wbuf = audio + i * AAC_BLOCK_SIZE_SHORT;
985 for (j = 0; j < AAC_BLOCK_SIZE_SHORT; j++)
986 max = FFMAX(max, fabsf(wbuf[j]));
990 for (i = 0; i < 8; i++)
994 wi.window_type[1] = prev_type;
995 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
996 float clipping = 0.0f;
1000 if (wi.window_type[0] == LONG_START_SEQUENCE)
1001 wi.window_shape = 0;
1003 wi.window_shape = 1;
1005 for (i = 0; i < 8; i++)
1006 clipping = FFMAX(clipping, clippings[i]);
1007 wi.clipping[0] = clipping;
1012 wi.window_shape = 0;
1013 for (i = 0; i < 8; i++) {
1014 if (!((pch->next_grouping >> i) & 1))
1016 wi.grouping[lastgrp]++;
1019 for (i = 0; i < 8; i += wi.grouping[i]) {
1021 float clipping = 0.0f;
1022 for (w = 0; w < wi.grouping[i] && !clipping; w++)
1023 clipping = FFMAX(clipping, clippings[i+w]);
1024 wi.clipping[i] = clipping;
1028 /* Determine grouping, based on the location of the first attack, and save for
1030 * FIXME: Move this to analysis.
1031 * TODO: Tune groupings depending on attack location
1032 * TODO: Handle more than one attack in a group
1034 for (i = 0; i < 9; i++) {
1040 pch->next_grouping = window_grouping[grouping];
1042 pch->prev_attack = attacks[8];
1047 const FFPsyModel ff_aac_psy_model =
1049 .name = "3GPP TS 26.403-inspired model",
1050 .init = psy_3gpp_init,
1051 .window = psy_lame_window,
1052 .analyze = psy_3gpp_analyze,
1053 .end = psy_3gpp_end,