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/libm.h"
34 /***********************************
36 * try other bitrate controlling mechanism (maybe use ratecontrol.c?)
37 * control quality for quality-based output
38 **********************************/
41 * constants for 3GPP AAC psychoacoustic model
44 #define PSY_3GPP_THR_SPREAD_HI 1.5f // spreading factor for low-to-hi threshold spreading (15 dB/Bark)
45 #define PSY_3GPP_THR_SPREAD_LOW 3.0f // spreading factor for hi-to-low threshold spreading (30 dB/Bark)
46 /* spreading factor for low-to-hi energy spreading, long block, > 22kbps/channel (20dB/Bark) */
47 #define PSY_3GPP_EN_SPREAD_HI_L1 2.0f
48 /* spreading factor for low-to-hi energy spreading, long block, <= 22kbps/channel (15dB/Bark) */
49 #define PSY_3GPP_EN_SPREAD_HI_L2 1.5f
50 /* spreading factor for low-to-hi energy spreading, short block (15 dB/Bark) */
51 #define PSY_3GPP_EN_SPREAD_HI_S 1.5f
52 /* spreading factor for hi-to-low energy spreading, long block (30dB/Bark) */
53 #define PSY_3GPP_EN_SPREAD_LOW_L 3.0f
54 /* spreading factor for hi-to-low energy spreading, short block (20dB/Bark) */
55 #define PSY_3GPP_EN_SPREAD_LOW_S 2.0f
57 #define PSY_3GPP_RPEMIN 0.01f
58 #define PSY_3GPP_RPELEV 2.0f
60 #define PSY_3GPP_C1 3.0f /* log2(8) */
61 #define PSY_3GPP_C2 1.3219281f /* log2(2.5) */
62 #define PSY_3GPP_C3 0.55935729f /* 1 - C2 / C1 */
64 #define PSY_SNR_1DB 7.9432821e-1f /* -1dB */
65 #define PSY_SNR_25DB 3.1622776e-3f /* -25dB */
67 #define PSY_3GPP_SAVE_SLOPE_L -0.46666667f
68 #define PSY_3GPP_SAVE_SLOPE_S -0.36363637f
69 #define PSY_3GPP_SAVE_ADD_L -0.84285712f
70 #define PSY_3GPP_SAVE_ADD_S -0.75f
71 #define PSY_3GPP_SPEND_SLOPE_L 0.66666669f
72 #define PSY_3GPP_SPEND_SLOPE_S 0.81818181f
73 #define PSY_3GPP_SPEND_ADD_L -0.35f
74 #define PSY_3GPP_SPEND_ADD_S -0.26111111f
75 #define PSY_3GPP_CLIP_LO_L 0.2f
76 #define PSY_3GPP_CLIP_LO_S 0.2f
77 #define PSY_3GPP_CLIP_HI_L 0.95f
78 #define PSY_3GPP_CLIP_HI_S 0.75f
80 #define PSY_3GPP_AH_THR_LONG 0.5f
81 #define PSY_3GPP_AH_THR_SHORT 0.63f
89 #define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f)
90 #define PSY_3GPP_PE_TO_BITS(bits) ((bits) / 1.18f)
92 /* LAME psy model constants */
93 #define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order
94 #define AAC_BLOCK_SIZE_LONG 1024 ///< long block size
95 #define AAC_BLOCK_SIZE_SHORT 128 ///< short block size
96 #define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence
97 #define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block
104 * information for single band used by 3GPP TS26.403-inspired psychoacoustic model
106 typedef struct AacPsyBand{
107 float energy; ///< band energy
108 float thr; ///< energy threshold
109 float thr_quiet; ///< threshold in quiet
110 float nz_lines; ///< number of non-zero spectral lines
111 float active_lines; ///< number of active spectral lines
112 float pe; ///< perceptual entropy
113 float pe_const; ///< constant part of the PE calculation
114 float norm_fac; ///< normalization factor for linearization
115 int avoid_holes; ///< hole avoidance flag
119 * single/pair channel context for psychoacoustic model
121 typedef struct AacPsyChannel{
122 AacPsyBand band[128]; ///< bands information
123 AacPsyBand prev_band[128]; ///< bands information from the previous frame
125 float win_energy; ///< sliding average of channel energy
126 float iir_state[2]; ///< hi-pass IIR filter state
127 uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence)
128 enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame
129 /* LAME psy model specific members */
130 float attack_threshold; ///< attack threshold for this channel
131 float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS];
132 int prev_attack; ///< attack value for the last short block in the previous sequence
136 * psychoacoustic model frame type-dependent coefficients
138 typedef struct AacPsyCoeffs{
139 float ath; ///< absolute threshold of hearing per bands
140 float barks; ///< Bark value for each spectral band in long frame
141 float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame
142 float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame
143 float min_snr; ///< minimal SNR
147 * 3GPP TS26.403-inspired psychoacoustic model specific data
149 typedef struct AacPsyContext{
150 int chan_bitrate; ///< bitrate per channel
151 int frame_bits; ///< average bits per frame
152 int fill_level; ///< bit reservoir fill level
154 float min; ///< minimum allowed PE for bit factor calculation
155 float max; ///< maximum allowed PE for bit factor calculation
156 float previous; ///< allowed PE of the previous frame
157 float correction; ///< PE correction factor
159 AacPsyCoeffs psy_coef[2][64];
161 float global_quality; ///< normalized global quality taken from avctx
165 * LAME psy model preset struct
167 typedef struct PsyLamePreset {
168 int quality; ///< Quality to map the rest of the vaules to.
169 /* This is overloaded to be both kbps per channel in ABR mode, and
170 * requested quality in constant quality mode.
172 float st_lrm; ///< short threshold for L, R, and M channels
176 * LAME psy model preset table for ABR
178 static const PsyLamePreset psy_abr_map[] = {
179 /* TODO: Tuning. These were taken from LAME. */
197 * LAME psy model preset table for constant quality
199 static const PsyLamePreset psy_vbr_map[] = {
215 * LAME psy model FIR coefficient table
217 static const float psy_fir_coeffs[] = {
218 -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
219 -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
220 -5.52212e-17 * 2, -0.313819 * 2
224 # include "mips/aacpsy_mips.h"
225 #endif /* ARCH_MIPS */
228 * Calculate the ABR attack threshold from the above LAME psymodel table.
230 static float lame_calc_attack_threshold(int bitrate)
232 /* Assume max bitrate to start with */
233 int lower_range = 12, upper_range = 12;
234 int lower_range_kbps = psy_abr_map[12].quality;
235 int upper_range_kbps = psy_abr_map[12].quality;
238 /* Determine which bitrates the value specified falls between.
239 * If the loop ends without breaking our above assumption of 320kbps was correct.
241 for (i = 1; i < 13; i++) {
242 if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) {
244 upper_range_kbps = psy_abr_map[i ].quality;
246 lower_range_kbps = psy_abr_map[i - 1].quality;
247 break; /* Upper range found */
251 /* Determine which range the value specified is closer to */
252 if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps))
253 return psy_abr_map[lower_range].st_lrm;
254 return psy_abr_map[upper_range].st_lrm;
258 * LAME psy model specific initialization
260 static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx)
264 for (i = 0; i < avctx->channels; i++) {
265 AacPsyChannel *pch = &ctx->ch[i];
267 if (avctx->flags & AV_CODEC_FLAG_QSCALE)
268 pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm;
270 pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000);
272 for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++)
273 pch->prev_energy_subshort[j] = 10.0f;
278 * Calculate Bark value for given line.
280 static av_cold float calc_bark(float f)
282 return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f));
287 * Calculate ATH value for given frequency.
288 * Borrowed from Lame.
290 static av_cold float ath(float f, float add)
293 return 3.64 * pow(f, -0.8)
294 - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4))
295 + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7))
296 + (0.6 + 0.04 * add) * 0.001 * f * f * f * f;
299 static av_cold int psy_3gpp_init(FFPsyContext *ctx) {
303 float prev, minscale, minath, minsnr, pe_min;
304 int chan_bitrate = ctx->avctx->bit_rate / ((ctx->avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : ctx->avctx->channels);
306 const int bandwidth = ctx->avctx->cutoff ? ctx->avctx->cutoff : AAC_CUTOFF(ctx->avctx);
307 const float num_bark = calc_bark((float)bandwidth);
309 ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
310 if (!ctx->model_priv_data)
311 return AVERROR(ENOMEM);
312 pctx = (AacPsyContext*) ctx->model_priv_data;
313 pctx->global_quality = (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) * 0.01f;
315 if (ctx->avctx->flags & CODEC_FLAG_QSCALE) {
316 /* Use the target average bitrate to compute spread parameters */
317 chan_bitrate = (int)(chan_bitrate / 120.0 * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120));
320 pctx->chan_bitrate = chan_bitrate;
321 pctx->frame_bits = FFMIN(2560, chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate);
322 pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
323 pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
324 ctx->bitres.size = 6144 - pctx->frame_bits;
325 ctx->bitres.size -= ctx->bitres.size % 8;
326 pctx->fill_level = ctx->bitres.size;
327 minath = ath(3410 - 0.733 * ATH_ADD, ATH_ADD);
328 for (j = 0; j < 2; j++) {
329 AacPsyCoeffs *coeffs = pctx->psy_coef[j];
330 const uint8_t *band_sizes = ctx->bands[j];
331 float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
332 float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) / ctx->avctx->sample_rate;
333 /* reference encoder uses 2.4% here instead of 60% like the spec says */
334 float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark;
335 float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L;
336 /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */
337 float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1;
341 for (g = 0; g < ctx->num_bands[j]; g++) {
343 bark = calc_bark((i-1) * line_to_frequency);
344 coeffs[g].barks = (bark + prev) / 2.0;
347 for (g = 0; g < ctx->num_bands[j] - 1; g++) {
348 AacPsyCoeffs *coeff = &coeffs[g];
349 float bark_width = coeffs[g+1].barks - coeffs->barks;
350 coeff->spread_low[0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_LOW);
351 coeff->spread_hi [0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_HI);
352 coeff->spread_low[1] = pow(10.0, -bark_width * en_spread_low);
353 coeff->spread_hi [1] = pow(10.0, -bark_width * en_spread_hi);
354 pe_min = bark_pe * bark_width;
355 minsnr = exp2(pe_min / band_sizes[g]) - 1.5f;
356 coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB);
359 for (g = 0; g < ctx->num_bands[j]; g++) {
360 minscale = ath(start * line_to_frequency, ATH_ADD);
361 for (i = 1; i < band_sizes[g]; i++)
362 minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD));
363 coeffs[g].ath = minscale - minath;
364 start += band_sizes[g];
368 pctx->ch = av_mallocz_array(ctx->avctx->channels, sizeof(AacPsyChannel));
370 av_freep(&ctx->model_priv_data);
371 return AVERROR(ENOMEM);
374 lame_window_init(pctx, ctx->avctx);
380 * IIR filter used in block switching decision
382 static float iir_filter(int in, float state[2])
386 ret = 0.7548f * (in - state[0]) + 0.5095f * state[1];
393 * window grouping information stored as bits (0 - new group, 1 - group continues)
395 static const uint8_t window_grouping[9] = {
396 0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
400 * Tell encoder which window types to use.
401 * @see 3GPP TS26.403 5.4.1 "Blockswitching"
403 static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx,
404 const int16_t *audio,
406 int channel, int prev_type)
409 int br = ((AacPsyContext*)ctx->model_priv_data)->chan_bitrate;
410 int attack_ratio = br <= 16000 ? 18 : 10;
411 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
412 AacPsyChannel *pch = &pctx->ch[channel];
413 uint8_t grouping = 0;
414 int next_type = pch->next_window_seq;
415 FFPsyWindowInfo wi = { { 0 } };
419 int switch_to_eight = 0;
420 float sum = 0.0, sum2 = 0.0;
423 for (i = 0; i < 8; i++) {
424 for (j = 0; j < 128; j++) {
425 v = iir_filter(la[i*128+j], pch->iir_state);
431 for (i = 0; i < 8; i++) {
432 if (s[i] > pch->win_energy * attack_ratio) {
438 pch->win_energy = pch->win_energy*7/8 + sum2/64;
440 wi.window_type[1] = prev_type;
442 case ONLY_LONG_SEQUENCE:
443 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
444 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
446 case LONG_START_SEQUENCE:
447 wi.window_type[0] = EIGHT_SHORT_SEQUENCE;
448 grouping = pch->next_grouping;
449 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
451 case LONG_STOP_SEQUENCE:
452 wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
453 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
455 case EIGHT_SHORT_SEQUENCE:
456 stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight;
457 wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
458 grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0;
459 next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
463 pch->next_grouping = window_grouping[attack_n];
464 pch->next_window_seq = next_type;
466 for (i = 0; i < 3; i++)
467 wi.window_type[i] = prev_type;
468 grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0;
472 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
478 for (i = 0; i < 8; i++) {
479 if (!((grouping >> i) & 1))
481 wi.grouping[lastgrp]++;
488 /* 5.6.1.2 "Calculation of Bit Demand" */
489 static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size,
492 const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L;
493 const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L;
494 const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L;
495 const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L;
496 const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L;
497 const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L;
498 float clipped_pe, bit_save, bit_spend, bit_factor, fill_level;
500 ctx->fill_level += ctx->frame_bits - bits;
501 ctx->fill_level = av_clip(ctx->fill_level, 0, size);
502 fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high);
503 clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max);
504 bit_save = (fill_level + bitsave_add) * bitsave_slope;
505 assert(bit_save <= 0.3f && bit_save >= -0.05000001f);
506 bit_spend = (fill_level + bitspend_add) * bitspend_slope;
507 assert(bit_spend <= 0.5f && bit_spend >= -0.1f);
508 /* The bit factor graph in the spec is obviously incorrect.
509 * bit_spend + ((bit_spend - bit_spend))...
510 * The reference encoder subtracts everything from 1, but also seems incorrect.
511 * 1 - bit_save + ((bit_spend + bit_save))...
512 * Hopefully below is correct.
514 bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min);
515 /* NOTE: The reference encoder attempts to center pe max/min around the current pe. */
516 ctx->pe.max = FFMAX(pe, ctx->pe.max);
517 ctx->pe.min = FFMIN(pe, ctx->pe.min);
519 /* NOTE: allocate a minimum of 1/8th average frame bits, to avoid
520 * reservoir starvation from producing zero-bit frames
523 ctx->frame_bits * bit_factor,
524 FFMAX(ctx->frame_bits + size - bits, ctx->frame_bits / 8));
527 static float calc_pe_3gpp(AacPsyBand *band)
532 band->pe_const = 0.0f;
533 band->active_lines = 0.0f;
534 if (band->energy > band->thr) {
535 a = log2f(band->energy);
536 pe = a - log2f(band->thr);
537 band->active_lines = band->nz_lines;
538 if (pe < PSY_3GPP_C1) {
539 pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2;
540 a = a * PSY_3GPP_C3 + PSY_3GPP_C2;
541 band->active_lines *= PSY_3GPP_C3;
543 band->pe = pe * band->nz_lines;
544 band->pe_const = a * band->nz_lines;
550 static float calc_reduction_3gpp(float a, float desired_pe, float pe,
553 float thr_avg, reduction;
555 if(active_lines == 0.0)
558 thr_avg = exp2f((a - pe) / (4.0f * active_lines));
559 reduction = exp2f((a - desired_pe) / (4.0f * active_lines)) - thr_avg;
561 return FFMAX(reduction, 0.0f);
564 static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr,
567 float thr = band->thr;
569 if (band->energy > thr) {
571 thr = sqrtf(thr) + reduction;
575 /* This deviates from the 3GPP spec to match the reference encoder.
576 * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands
577 * that have hole avoidance on (active or inactive). It always reduces the
578 * threshold of bands with hole avoidance off.
580 if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) {
581 thr = FFMAX(band->thr, band->energy * min_snr);
582 band->avoid_holes = PSY_3GPP_AH_ACTIVE;
589 #ifndef calc_thr_3gpp
590 static void calc_thr_3gpp(const FFPsyWindowInfo *wi, const int num_bands, AacPsyChannel *pch,
591 const uint8_t *band_sizes, const float *coefs)
595 for (w = 0; w < wi->num_windows*16; w += 16) {
596 for (g = 0; g < num_bands; g++) {
597 AacPsyBand *band = &pch->band[w+g];
599 float form_factor = 0.0f;
602 for (i = 0; i < band_sizes[g]; i++) {
603 band->energy += coefs[start+i] * coefs[start+i];
604 form_factor += sqrtf(fabs(coefs[start+i]));
606 Temp = band->energy > 0 ? sqrtf((float)band_sizes[g] / band->energy) : 0;
607 band->thr = band->energy * 0.001258925f;
608 band->nz_lines = form_factor * sqrtf(Temp);
610 start += band_sizes[g];
614 #endif /* calc_thr_3gpp */
616 #ifndef psy_hp_filter
617 static void psy_hp_filter(const float *firbuf, float *hpfsmpl, const float *psy_fir_coeffs)
620 for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
622 sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2];
624 for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
625 sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]);
626 sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]);
628 /* NOTE: The LAME psymodel expects it's input in the range -32768 to 32768.
629 * Tuning this for normalized floats would be difficult. */
630 hpfsmpl[i] = (sum1 + sum2) * 32768.0f;
633 #endif /* psy_hp_filter */
636 * Calculate band thresholds as suggested in 3GPP TS26.403
638 static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel,
639 const float *coefs, const FFPsyWindowInfo *wi)
641 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
642 AacPsyChannel *pch = &pctx->ch[channel];
644 float desired_bits, desired_pe, delta_pe, reduction= NAN, spread_en[128] = {0};
645 float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
646 float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f);
647 const int num_bands = ctx->num_bands[wi->num_windows == 8];
648 const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8];
649 AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
650 const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG;
652 //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
653 calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs);
655 //modify thresholds and energies - spread, threshold in quiet, pre-echo control
656 for (w = 0; w < wi->num_windows*16; w += 16) {
657 AacPsyBand *bands = &pch->band[w];
659 /* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */
660 spread_en[0] = bands[0].energy;
661 for (g = 1; g < num_bands; g++) {
662 bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]);
663 spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]);
665 for (g = num_bands - 2; g >= 0; g--) {
666 bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]);
667 spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]);
669 //5.4.2.4 "Threshold in quiet"
670 for (g = 0; g < num_bands; g++) {
671 AacPsyBand *band = &bands[g];
673 band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath);
674 //5.4.2.5 "Pre-echo control"
675 if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (wi->window_type[1] == LONG_START_SEQUENCE && !w)))
676 band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr,
677 PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
679 /* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */
680 pe += calc_pe_3gpp(band);
682 active_lines += band->active_lines;
684 /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */
685 if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f)
686 band->avoid_holes = PSY_3GPP_AH_NONE;
688 band->avoid_holes = PSY_3GPP_AH_INACTIVE;
692 /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */
693 ctx->ch[channel].entropy = pe;
694 if (ctx->avctx->flags & CODEC_FLAG_QSCALE) {
695 /* (2.5 * 120) achieves almost transparent rate, and we want to give
696 * ample room downwards, so we make that equivalent to QSCALE=2.4
698 desired_pe = pe * (ctx->avctx->global_quality ? ctx->avctx->global_quality : 120) / (2 * 2.5f * 120.0f);
699 desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe));
700 desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping
702 /* PE slope smoothing */
703 if (ctx->bitres.bits > 0) {
704 desired_bits = FFMIN(2560, PSY_3GPP_PE_TO_BITS(desired_pe));
705 desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits); // reflect clipping
708 pctx->pe.max = FFMAX(pe, pctx->pe.max);
709 pctx->pe.min = FFMIN(pe, pctx->pe.min);
711 desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8);
712 desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits);
714 /* NOTE: PE correction is kept simple. During initial testing it had very
715 * little effect on the final bitrate. Probably a good idea to come
716 * back and do more testing later.
718 if (ctx->bitres.bits > 0)
719 desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits),
722 pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits);
723 ctx->bitres.alloc = desired_bits;
725 if (desired_pe < pe) {
726 /* 5.6.1.3.4 "First Estimation of the reduction value" */
727 for (w = 0; w < wi->num_windows*16; w += 16) {
728 reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines);
732 for (g = 0; g < num_bands; g++) {
733 AacPsyBand *band = &pch->band[w+g];
735 band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
737 pe += calc_pe_3gpp(band);
739 active_lines += band->active_lines;
743 /* 5.6.1.3.5 "Second Estimation of the reduction value" */
744 for (i = 0; i < 2; i++) {
745 float pe_no_ah = 0.0f, desired_pe_no_ah;
746 active_lines = a = 0.0f;
747 for (w = 0; w < wi->num_windows*16; w += 16) {
748 for (g = 0; g < num_bands; g++) {
749 AacPsyBand *band = &pch->band[w+g];
751 if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) {
752 pe_no_ah += band->pe;
754 active_lines += band->active_lines;
758 desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f);
759 if (active_lines > 0.0f)
760 reduction = calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines);
763 for (w = 0; w < wi->num_windows*16; w += 16) {
764 for (g = 0; g < num_bands; g++) {
765 AacPsyBand *band = &pch->band[w+g];
767 if (active_lines > 0.0f)
768 band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
769 pe += calc_pe_3gpp(band);
770 if (band->thr > 0.0f)
771 band->norm_fac = band->active_lines / band->thr;
773 band->norm_fac = 0.0f;
774 norm_fac += band->norm_fac;
777 delta_pe = desired_pe - pe;
778 if (fabs(delta_pe) > 0.05f * desired_pe)
782 if (pe < 1.15f * desired_pe) {
783 /* 6.6.1.3.6 "Final threshold modification by linearization" */
784 norm_fac = 1.0f / norm_fac;
785 for (w = 0; w < wi->num_windows*16; w += 16) {
786 for (g = 0; g < num_bands; g++) {
787 AacPsyBand *band = &pch->band[w+g];
789 if (band->active_lines > 0.5f) {
790 float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe;
791 float thr = band->thr;
793 thr *= exp2f(delta_sfb_pe / band->active_lines);
794 if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE)
795 thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy);
801 /* 5.6.1.3.7 "Further perceptual entropy reduction" */
803 while (pe > desired_pe && g--) {
804 for (w = 0; w < wi->num_windows*16; w+= 16) {
805 AacPsyBand *band = &pch->band[w+g];
806 if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) {
807 coeffs[g].min_snr = PSY_SNR_1DB;
808 band->thr = band->energy * PSY_SNR_1DB;
809 pe += band->active_lines * 1.5f - band->pe;
813 /* TODO: allow more holes (unused without mid/side) */
817 for (w = 0; w < wi->num_windows*16; w += 16) {
818 for (g = 0; g < num_bands; g++) {
819 AacPsyBand *band = &pch->band[w+g];
820 FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g];
822 psy_band->threshold = band->thr;
823 psy_band->energy = band->energy;
824 psy_band->spread = band->active_lines * 2.0f / band_sizes[g];
825 psy_band->bits = PSY_3GPP_PE_TO_BITS(band->pe);
829 memcpy(pch->prev_band, pch->band, sizeof(pch->band));
832 static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
833 const float **coeffs, const FFPsyWindowInfo *wi)
836 FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel);
838 for (ch = 0; ch < group->num_ch; ch++)
839 psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]);
842 static av_cold void psy_3gpp_end(FFPsyContext *apc)
844 AacPsyContext *pctx = (AacPsyContext*) apc->model_priv_data;
846 av_freep(&apc->model_priv_data);
849 static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
851 int blocktype = ONLY_LONG_SEQUENCE;
853 if (ctx->next_window_seq == EIGHT_SHORT_SEQUENCE)
854 blocktype = LONG_STOP_SEQUENCE;
856 blocktype = EIGHT_SHORT_SEQUENCE;
857 if (ctx->next_window_seq == ONLY_LONG_SEQUENCE)
858 ctx->next_window_seq = LONG_START_SEQUENCE;
859 if (ctx->next_window_seq == LONG_STOP_SEQUENCE)
860 ctx->next_window_seq = EIGHT_SHORT_SEQUENCE;
863 wi->window_type[0] = ctx->next_window_seq;
864 ctx->next_window_seq = blocktype;
867 static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio,
868 const float *la, int channel, int prev_type)
870 AacPsyContext *pctx = (AacPsyContext*) ctx->model_priv_data;
871 AacPsyChannel *pch = &pctx->ch[channel];
873 int uselongblock = 1;
874 int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
875 float clippings[AAC_NUM_BLOCKS_SHORT];
877 FFPsyWindowInfo wi = { { 0 } };
880 float hpfsmpl[AAC_BLOCK_SIZE_LONG];
881 float const *pf = hpfsmpl;
882 float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
883 float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
884 float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
885 const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN);
888 /* LAME comment: apply high pass filter of fs/4 */
889 psy_hp_filter(firbuf, hpfsmpl, psy_fir_coeffs);
891 /* Calculate the energies of each sub-shortblock */
892 for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) {
893 energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)];
894 assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0);
895 attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)];
896 energy_short[0] += energy_subshort[i];
899 for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) {
900 float const *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS);
902 for (; pf < pfe; pf++)
903 p = FFMAX(p, fabsf(*pf));
904 pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
905 energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
906 /* NOTE: The indexes below are [i + 3 - 2] in the LAME source.
907 * Obviously the 3 and 2 have some significance, or this would be just [i + 1]
908 * (which is what we use here). What the 3 stands for is ambiguous, as it is both
909 * number of short blocks, and the number of sub-short blocks.
910 * It seems that LAME is comparing each sub-block to sub-block + 1 in the
913 if (p > energy_subshort[i + 1])
914 p = p / energy_subshort[i + 1];
915 else if (energy_subshort[i + 1] > p * 10.0f)
916 p = energy_subshort[i + 1] / (p * 10.0f);
919 attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
922 /* compare energy between sub-short blocks */
923 for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++)
924 if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS])
925 if (attack_intensity[i] > pch->attack_threshold)
926 attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1;
928 /* should have energy change between short blocks, in order to avoid periodic signals */
929 /* Good samples to show the effect are Trumpet test songs */
930 /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */
931 /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */
932 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) {
933 float const u = energy_short[i - 1];
934 float const v = energy_short[i];
935 float const m = FFMAX(u, v);
936 if (m < 40000) { /* (2) */
937 if (u < 1.7f * v && v < 1.7f * u) { /* (1) */
938 if (i == 1 && attacks[0] < attacks[i])
943 att_sum += attacks[i];
946 if (attacks[0] <= pch->prev_attack)
949 att_sum += attacks[0];
950 /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */
951 if (pch->prev_attack == 3 || att_sum) {
954 for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++)
955 if (attacks[i] && attacks[i-1])
959 /* We have no lookahead info, so just use same type as the previous sequence. */
960 uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE);
963 lame_apply_block_type(pch, &wi, uselongblock);
965 /* Calculate input sample maximums and evaluate clipping risk */
967 for (i = 0; i < AAC_NUM_BLOCKS_SHORT; i++) {
968 const float *wbuf = audio + i * AAC_BLOCK_SIZE_SHORT;
971 for (j = 0; j < AAC_BLOCK_SIZE_SHORT; j++)
972 max = FFMAX(max, fabsf(wbuf[j]));
976 for (i = 0; i < 8; i++)
980 wi.window_type[1] = prev_type;
981 if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
982 float clipping = 0.0f;
986 if (wi.window_type[0] == LONG_START_SEQUENCE)
991 for (i = 0; i < 8; i++)
992 clipping = FFMAX(clipping, clippings[i]);
993 wi.clipping[0] = clipping;
999 for (i = 0; i < 8; i++) {
1000 if (!((pch->next_grouping >> i) & 1))
1002 wi.grouping[lastgrp]++;
1005 for (i = 0; i < 8; i += wi.grouping[i]) {
1007 float clipping = 0.0f;
1008 for (w = 0; w < wi.grouping[i] && !clipping; w++)
1009 clipping = FFMAX(clipping, clippings[i+w]);
1010 wi.clipping[i] = clipping;
1014 /* Determine grouping, based on the location of the first attack, and save for
1016 * FIXME: Move this to analysis.
1017 * TODO: Tune groupings depending on attack location
1018 * TODO: Handle more than one attack in a group
1020 for (i = 0; i < 9; i++) {
1026 pch->next_grouping = window_grouping[grouping];
1028 pch->prev_attack = attacks[8];
1033 const FFPsyModel ff_aac_psy_model =
1035 .name = "3GPP TS 26.403-inspired model",
1036 .init = psy_3gpp_init,
1037 .window = psy_lame_window,
1038 .analyze = psy_3gpp_analyze,
1039 .end = psy_3gpp_end,