FFmpeg
 All Data Structures Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
aacpsy.c
Go to the documentation of this file.
1 /*
2  * AAC encoder psychoacoustic model
3  * Copyright (C) 2008 Konstantin Shishkov
4  *
5  * This file is part of FFmpeg.
6  *
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.
11  *
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.
16  *
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
20  */
21 
22 /**
23  * @file
24  * AAC encoder psychoacoustic model
25  */
26 
27 #include "libavutil/libm.h"
28 
29 #include "avcodec.h"
30 #include "aactab.h"
31 #include "psymodel.h"
32 
33 /***********************************
34  * TODOs:
35  * try other bitrate controlling mechanism (maybe use ratecontrol.c?)
36  * control quality for quality-based output
37  **********************************/
38 
39 /**
40  * constants for 3GPP AAC psychoacoustic model
41  * @{
42  */
43 #define PSY_3GPP_THR_SPREAD_HI 1.5f // spreading factor for low-to-hi threshold spreading (15 dB/Bark)
44 #define PSY_3GPP_THR_SPREAD_LOW 3.0f // spreading factor for hi-to-low threshold spreading (30 dB/Bark)
45 /* spreading factor for low-to-hi energy spreading, long block, > 22kbps/channel (20dB/Bark) */
46 #define PSY_3GPP_EN_SPREAD_HI_L1 2.0f
47 /* spreading factor for low-to-hi energy spreading, long block, <= 22kbps/channel (15dB/Bark) */
48 #define PSY_3GPP_EN_SPREAD_HI_L2 1.5f
49 /* spreading factor for low-to-hi energy spreading, short block (15 dB/Bark) */
50 #define PSY_3GPP_EN_SPREAD_HI_S 1.5f
51 /* spreading factor for hi-to-low energy spreading, long block (30dB/Bark) */
52 #define PSY_3GPP_EN_SPREAD_LOW_L 3.0f
53 /* spreading factor for hi-to-low energy spreading, short block (20dB/Bark) */
54 #define PSY_3GPP_EN_SPREAD_LOW_S 2.0f
55 
56 #define PSY_3GPP_RPEMIN 0.01f
57 #define PSY_3GPP_RPELEV 2.0f
58 
59 #define PSY_3GPP_C1 3.0f /* log2(8) */
60 #define PSY_3GPP_C2 1.3219281f /* log2(2.5) */
61 #define PSY_3GPP_C3 0.55935729f /* 1 - C2 / C1 */
62 
63 #define PSY_SNR_1DB 7.9432821e-1f /* -1dB */
64 #define PSY_SNR_25DB 3.1622776e-3f /* -25dB */
65 
66 #define PSY_3GPP_SAVE_SLOPE_L -0.46666667f
67 #define PSY_3GPP_SAVE_SLOPE_S -0.36363637f
68 #define PSY_3GPP_SAVE_ADD_L -0.84285712f
69 #define PSY_3GPP_SAVE_ADD_S -0.75f
70 #define PSY_3GPP_SPEND_SLOPE_L 0.66666669f
71 #define PSY_3GPP_SPEND_SLOPE_S 0.81818181f
72 #define PSY_3GPP_SPEND_ADD_L -0.35f
73 #define PSY_3GPP_SPEND_ADD_S -0.26111111f
74 #define PSY_3GPP_CLIP_LO_L 0.2f
75 #define PSY_3GPP_CLIP_LO_S 0.2f
76 #define PSY_3GPP_CLIP_HI_L 0.95f
77 #define PSY_3GPP_CLIP_HI_S 0.75f
78 
79 #define PSY_3GPP_AH_THR_LONG 0.5f
80 #define PSY_3GPP_AH_THR_SHORT 0.63f
81 
82 enum {
86 };
87 
88 #define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f)
89 
90 /* LAME psy model constants */
91 #define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order
92 #define AAC_BLOCK_SIZE_LONG 1024 ///< long block size
93 #define AAC_BLOCK_SIZE_SHORT 128 ///< short block size
94 #define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence
95 #define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block
96 
97 /**
98  * @}
99  */
100 
101 /**
102  * information for single band used by 3GPP TS26.403-inspired psychoacoustic model
103  */
104 typedef struct AacPsyBand{
105  float energy; ///< band energy
106  float thr; ///< energy threshold
107  float thr_quiet; ///< threshold in quiet
108  float nz_lines; ///< number of non-zero spectral lines
109  float active_lines; ///< number of active spectral lines
110  float pe; ///< perceptual entropy
111  float pe_const; ///< constant part of the PE calculation
112  float norm_fac; ///< normalization factor for linearization
113  int avoid_holes; ///< hole avoidance flag
114 }AacPsyBand;
115 
116 /**
117  * single/pair channel context for psychoacoustic model
118  */
119 typedef struct AacPsyChannel{
120  AacPsyBand band[128]; ///< bands information
121  AacPsyBand prev_band[128]; ///< bands information from the previous frame
122 
123  float win_energy; ///< sliding average of channel energy
124  float iir_state[2]; ///< hi-pass IIR filter state
125  uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence)
126  enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame
127  /* LAME psy model specific members */
128  float attack_threshold; ///< attack threshold for this channel
130  int prev_attack; ///< attack value for the last short block in the previous sequence
132 
133 /**
134  * psychoacoustic model frame type-dependent coefficients
135  */
136 typedef struct AacPsyCoeffs{
137  float ath; ///< absolute threshold of hearing per bands
138  float barks; ///< Bark value for each spectral band in long frame
139  float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame
140  float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame
141  float min_snr; ///< minimal SNR
142 }AacPsyCoeffs;
143 
144 /**
145  * 3GPP TS26.403-inspired psychoacoustic model specific data
146  */
147 typedef struct AacPsyContext{
148  int chan_bitrate; ///< bitrate per channel
149  int frame_bits; ///< average bits per frame
150  int fill_level; ///< bit reservoir fill level
151  struct {
152  float min; ///< minimum allowed PE for bit factor calculation
153  float max; ///< maximum allowed PE for bit factor calculation
154  float previous; ///< allowed PE of the previous frame
155  float correction; ///< PE correction factor
156  } pe;
160 
161 /**
162  * LAME psy model preset struct
163  */
164 typedef struct {
165  int quality; ///< Quality to map the rest of the vaules to.
166  /* This is overloaded to be both kbps per channel in ABR mode, and
167  * requested quality in constant quality mode.
168  */
169  float st_lrm; ///< short threshold for L, R, and M channels
170 } PsyLamePreset;
171 
172 /**
173  * LAME psy model preset table for ABR
174  */
175 static const PsyLamePreset psy_abr_map[] = {
176 /* TODO: Tuning. These were taken from LAME. */
177 /* kbps/ch st_lrm */
178  { 8, 6.60},
179  { 16, 6.60},
180  { 24, 6.60},
181  { 32, 6.60},
182  { 40, 6.60},
183  { 48, 6.60},
184  { 56, 6.60},
185  { 64, 6.40},
186  { 80, 6.00},
187  { 96, 5.60},
188  {112, 5.20},
189  {128, 5.20},
190  {160, 5.20}
191 };
192 
193 /**
194 * LAME psy model preset table for constant quality
195 */
196 static const PsyLamePreset psy_vbr_map[] = {
197 /* vbr_q st_lrm */
198  { 0, 4.20},
199  { 1, 4.20},
200  { 2, 4.20},
201  { 3, 4.20},
202  { 4, 4.20},
203  { 5, 4.20},
204  { 6, 4.20},
205  { 7, 4.20},
206  { 8, 4.20},
207  { 9, 4.20},
208  {10, 4.20}
209 };
210 
211 /**
212  * LAME psy model FIR coefficient table
213  */
214 static const float psy_fir_coeffs[] = {
215  -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
216  -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
217  -5.52212e-17 * 2, -0.313819 * 2
218 };
219 
220 /**
221  * Calculate the ABR attack threshold from the above LAME psymodel table.
222  */
223 static float lame_calc_attack_threshold(int bitrate)
224 {
225  /* Assume max bitrate to start with */
226  int lower_range = 12, upper_range = 12;
227  int lower_range_kbps = psy_abr_map[12].quality;
228  int upper_range_kbps = psy_abr_map[12].quality;
229  int i;
230 
231  /* Determine which bitrates the value specified falls between.
232  * If the loop ends without breaking our above assumption of 320kbps was correct.
233  */
234  for (i = 1; i < 13; i++) {
235  if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) {
236  upper_range = i;
237  upper_range_kbps = psy_abr_map[i ].quality;
238  lower_range = i - 1;
239  lower_range_kbps = psy_abr_map[i - 1].quality;
240  break; /* Upper range found */
241  }
242  }
243 
244  /* Determine which range the value specified is closer to */
245  if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps))
246  return psy_abr_map[lower_range].st_lrm;
247  return psy_abr_map[upper_range].st_lrm;
248 }
249 
250 /**
251  * LAME psy model specific initialization
252  */
253 static void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx) {
254  int i, j;
255 
256  for (i = 0; i < avctx->channels; i++) {
257  AacPsyChannel *pch = &ctx->ch[i];
258 
259  if (avctx->flags & CODEC_FLAG_QSCALE)
260  pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm;
261  else
262  pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000);
263 
264  for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++)
265  pch->prev_energy_subshort[j] = 10.0f;
266  }
267 }
268 
269 /**
270  * Calculate Bark value for given line.
271  */
272 static av_cold float calc_bark(float f)
273 {
274  return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f));
275 }
276 
277 #define ATH_ADD 4
278 /**
279  * Calculate ATH value for given frequency.
280  * Borrowed from Lame.
281  */
282 static av_cold float ath(float f, float add)
283 {
284  f /= 1000.0f;
285  return 3.64 * pow(f, -0.8)
286  - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4))
287  + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7))
288  + (0.6 + 0.04 * add) * 0.001 * f * f * f * f;
289 }
290 
292  AacPsyContext *pctx;
293  float bark;
294  int i, j, g, start;
295  float prev, minscale, minath, minsnr, pe_min;
296  const int chan_bitrate = ctx->avctx->bit_rate / ctx->avctx->channels;
297  const int bandwidth = ctx->avctx->cutoff ? ctx->avctx->cutoff : AAC_CUTOFF(ctx->avctx);
298  const float num_bark = calc_bark((float)bandwidth);
299 
300  ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
301  pctx = (AacPsyContext*) ctx->model_priv_data;
302 
303  pctx->chan_bitrate = chan_bitrate;
304  pctx->frame_bits = chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate;
305  pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
306  pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
307  ctx->bitres.size = 6144 - pctx->frame_bits;
308  ctx->bitres.size -= ctx->bitres.size % 8;
309  pctx->fill_level = ctx->bitres.size;
310  minath = ath(3410, ATH_ADD);
311  for (j = 0; j < 2; j++) {
312  AacPsyCoeffs *coeffs = pctx->psy_coef[j];
313  const uint8_t *band_sizes = ctx->bands[j];
314  float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
315  float avg_chan_bits = chan_bitrate / ctx->avctx->sample_rate * (j ? 128.0f : 1024.0f);
316  /* reference encoder uses 2.4% here instead of 60% like the spec says */
317  float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark;
318  float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L;
319  /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */
320  float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1;
321 
322  i = 0;
323  prev = 0.0;
324  for (g = 0; g < ctx->num_bands[j]; g++) {
325  i += band_sizes[g];
326  bark = calc_bark((i-1) * line_to_frequency);
327  coeffs[g].barks = (bark + prev) / 2.0;
328  prev = bark;
329  }
330  for (g = 0; g < ctx->num_bands[j] - 1; g++) {
331  AacPsyCoeffs *coeff = &coeffs[g];
332  float bark_width = coeffs[g+1].barks - coeffs->barks;
333  coeff->spread_low[0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_LOW);
334  coeff->spread_hi [0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_HI);
335  coeff->spread_low[1] = pow(10.0, -bark_width * en_spread_low);
336  coeff->spread_hi [1] = pow(10.0, -bark_width * en_spread_hi);
337  pe_min = bark_pe * bark_width;
338  minsnr = exp2(pe_min / band_sizes[g]) - 1.5f;
339  coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB);
340  }
341  start = 0;
342  for (g = 0; g < ctx->num_bands[j]; g++) {
343  minscale = ath(start * line_to_frequency, ATH_ADD);
344  for (i = 1; i < band_sizes[g]; i++)
345  minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD));
346  coeffs[g].ath = minscale - minath;
347  start += band_sizes[g];
348  }
349  }
350 
351  pctx->ch = av_mallocz(sizeof(AacPsyChannel) * ctx->avctx->channels);
352 
353  lame_window_init(pctx, ctx->avctx);
354 
355  return 0;
356 }
357 
358 /**
359  * IIR filter used in block switching decision
360  */
361 static float iir_filter(int in, float state[2])
362 {
363  float ret;
364 
365  ret = 0.7548f * (in - state[0]) + 0.5095f * state[1];
366  state[0] = in;
367  state[1] = ret;
368  return ret;
369 }
370 
371 /**
372  * window grouping information stored as bits (0 - new group, 1 - group continues)
373  */
374 static const uint8_t window_grouping[9] = {
375  0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
376 };
377 
378 /**
379  * Tell encoder which window types to use.
380  * @see 3GPP TS26.403 5.4.1 "Blockswitching"
381  */
383  const int16_t *audio,
384  const int16_t *la,
385  int channel, int prev_type)
386 {
387  int i, j;
388  int br = ctx->avctx->bit_rate / ctx->avctx->channels;
389  int attack_ratio = br <= 16000 ? 18 : 10;
391  AacPsyChannel *pch = &pctx->ch[channel];
392  uint8_t grouping = 0;
393  int next_type = pch->next_window_seq;
394  FFPsyWindowInfo wi = { { 0 } };
395 
396  if (la) {
397  float s[8], v;
398  int switch_to_eight = 0;
399  float sum = 0.0, sum2 = 0.0;
400  int attack_n = 0;
401  int stay_short = 0;
402  for (i = 0; i < 8; i++) {
403  for (j = 0; j < 128; j++) {
404  v = iir_filter(la[i*128+j], pch->iir_state);
405  sum += v*v;
406  }
407  s[i] = sum;
408  sum2 += sum;
409  }
410  for (i = 0; i < 8; i++) {
411  if (s[i] > pch->win_energy * attack_ratio) {
412  attack_n = i + 1;
413  switch_to_eight = 1;
414  break;
415  }
416  }
417  pch->win_energy = pch->win_energy*7/8 + sum2/64;
418 
419  wi.window_type[1] = prev_type;
420  switch (prev_type) {
421  case ONLY_LONG_SEQUENCE:
422  wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
423  next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
424  break;
425  case LONG_START_SEQUENCE:
426  wi.window_type[0] = EIGHT_SHORT_SEQUENCE;
427  grouping = pch->next_grouping;
428  next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
429  break;
430  case LONG_STOP_SEQUENCE:
431  wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
432  next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
433  break;
435  stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight;
436  wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
437  grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0;
438  next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
439  break;
440  }
441 
442  pch->next_grouping = window_grouping[attack_n];
443  pch->next_window_seq = next_type;
444  } else {
445  for (i = 0; i < 3; i++)
446  wi.window_type[i] = prev_type;
447  grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0;
448  }
449 
450  wi.window_shape = 1;
451  if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
452  wi.num_windows = 1;
453  wi.grouping[0] = 1;
454  } else {
455  int lastgrp = 0;
456  wi.num_windows = 8;
457  for (i = 0; i < 8; i++) {
458  if (!((grouping >> i) & 1))
459  lastgrp = i;
460  wi.grouping[lastgrp]++;
461  }
462  }
463 
464  return wi;
465 }
466 
467 /* 5.6.1.2 "Calculation of Bit Demand" */
468 static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size,
469  int short_window)
470 {
471  const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L;
472  const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L;
473  const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L;
474  const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L;
475  const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L;
476  const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L;
477  float clipped_pe, bit_save, bit_spend, bit_factor, fill_level;
478 
479  ctx->fill_level += ctx->frame_bits - bits;
480  ctx->fill_level = av_clip(ctx->fill_level, 0, size);
481  fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high);
482  clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max);
483  bit_save = (fill_level + bitsave_add) * bitsave_slope;
484  assert(bit_save <= 0.3f && bit_save >= -0.05000001f);
485  bit_spend = (fill_level + bitspend_add) * bitspend_slope;
486  assert(bit_spend <= 0.5f && bit_spend >= -0.1f);
487  /* The bit factor graph in the spec is obviously incorrect.
488  * bit_spend + ((bit_spend - bit_spend))...
489  * The reference encoder subtracts everything from 1, but also seems incorrect.
490  * 1 - bit_save + ((bit_spend + bit_save))...
491  * Hopefully below is correct.
492  */
493  bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min);
494  /* NOTE: The reference encoder attempts to center pe max/min around the current pe. */
495  ctx->pe.max = FFMAX(pe, ctx->pe.max);
496  ctx->pe.min = FFMIN(pe, ctx->pe.min);
497 
498  return FFMIN(ctx->frame_bits * bit_factor, ctx->frame_bits + size - bits);
499 }
500 
502 {
503  float pe, a;
504 
505  band->pe = 0.0f;
506  band->pe_const = 0.0f;
507  band->active_lines = 0.0f;
508  if (band->energy > band->thr) {
509  a = log2f(band->energy);
510  pe = a - log2f(band->thr);
511  band->active_lines = band->nz_lines;
512  if (pe < PSY_3GPP_C1) {
513  pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2;
514  a = a * PSY_3GPP_C3 + PSY_3GPP_C2;
515  band->active_lines *= PSY_3GPP_C3;
516  }
517  band->pe = pe * band->nz_lines;
518  band->pe_const = a * band->nz_lines;
519  }
520 
521  return band->pe;
522 }
523 
524 static float calc_reduction_3gpp(float a, float desired_pe, float pe,
525  float active_lines)
526 {
527  float thr_avg, reduction;
528 
529  if(active_lines == 0.0)
530  return 0;
531 
532  thr_avg = exp2f((a - pe) / (4.0f * active_lines));
533  reduction = exp2f((a - desired_pe) / (4.0f * active_lines)) - thr_avg;
534 
535  return FFMAX(reduction, 0.0f);
536 }
537 
538 static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr,
539  float reduction)
540 {
541  float thr = band->thr;
542 
543  if (band->energy > thr) {
544  thr = sqrtf(thr);
545  thr = sqrtf(thr) + reduction;
546  thr *= thr;
547  thr *= thr;
548 
549  /* This deviates from the 3GPP spec to match the reference encoder.
550  * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands
551  * that have hole avoidance on (active or inactive). It always reduces the
552  * threshold of bands with hole avoidance off.
553  */
554  if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) {
555  thr = FFMAX(band->thr, band->energy * min_snr);
557  }
558  }
559 
560  return thr;
561 }
562 
563 /**
564  * Calculate band thresholds as suggested in 3GPP TS26.403
565  */
566 static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel,
567  const float *coefs, const FFPsyWindowInfo *wi)
568 {
570  AacPsyChannel *pch = &pctx->ch[channel];
571  int start = 0;
572  int i, w, g;
573  float desired_bits, desired_pe, delta_pe, reduction= NAN, spread_en[128] = {0};
574  float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
575  float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f);
576  const int num_bands = ctx->num_bands[wi->num_windows == 8];
577  const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8];
578  AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
579  const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG;
580 
581  //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
582  for (w = 0; w < wi->num_windows*16; w += 16) {
583  for (g = 0; g < num_bands; g++) {
584  AacPsyBand *band = &pch->band[w+g];
585 
586  float form_factor = 0.0f;
587  float Temp;
588  band->energy = 0.0f;
589  for (i = 0; i < band_sizes[g]; i++) {
590  band->energy += coefs[start+i] * coefs[start+i];
591  form_factor += sqrtf(fabs(coefs[start+i]));
592  }
593  Temp = band->energy > 0 ? sqrtf((float)band_sizes[g] / band->energy) : 0;
594  band->thr = band->energy * 0.001258925f;
595  band->nz_lines = form_factor * sqrtf(Temp);
596 
597  start += band_sizes[g];
598  }
599  }
600  //modify thresholds and energies - spread, threshold in quiet, pre-echo control
601  for (w = 0; w < wi->num_windows*16; w += 16) {
602  AacPsyBand *bands = &pch->band[w];
603 
604  /* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */
605  spread_en[0] = bands[0].energy;
606  for (g = 1; g < num_bands; g++) {
607  bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]);
608  spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]);
609  }
610  for (g = num_bands - 2; g >= 0; g--) {
611  bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]);
612  spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]);
613  }
614  //5.4.2.4 "Threshold in quiet"
615  for (g = 0; g < num_bands; g++) {
616  AacPsyBand *band = &bands[g];
617 
618  band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath);
619  //5.4.2.5 "Pre-echo control"
620  if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (wi->window_type[1] == LONG_START_SEQUENCE && !w)))
621  band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr,
622  PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
623 
624  /* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */
625  pe += calc_pe_3gpp(band);
626  a += band->pe_const;
627  active_lines += band->active_lines;
628 
629  /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */
630  if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f)
632  else
634  }
635  }
636 
637  /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */
638  ctx->ch[channel].entropy = pe;
639  desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8);
640  desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits);
641  /* NOTE: PE correction is kept simple. During initial testing it had very
642  * little effect on the final bitrate. Probably a good idea to come
643  * back and do more testing later.
644  */
645  if (ctx->bitres.bits > 0)
646  desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits),
647  0.85f, 1.15f);
648  pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits);
649 
650  if (desired_pe < pe) {
651  /* 5.6.1.3.4 "First Estimation of the reduction value" */
652  for (w = 0; w < wi->num_windows*16; w += 16) {
653  reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines);
654  pe = 0.0f;
655  a = 0.0f;
656  active_lines = 0.0f;
657  for (g = 0; g < num_bands; g++) {
658  AacPsyBand *band = &pch->band[w+g];
659 
660  band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
661  /* recalculate PE */
662  pe += calc_pe_3gpp(band);
663  a += band->pe_const;
664  active_lines += band->active_lines;
665  }
666  }
667 
668  /* 5.6.1.3.5 "Second Estimation of the reduction value" */
669  for (i = 0; i < 2; i++) {
670  float pe_no_ah = 0.0f, desired_pe_no_ah;
671  active_lines = a = 0.0f;
672  for (w = 0; w < wi->num_windows*16; w += 16) {
673  for (g = 0; g < num_bands; g++) {
674  AacPsyBand *band = &pch->band[w+g];
675 
676  if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) {
677  pe_no_ah += band->pe;
678  a += band->pe_const;
679  active_lines += band->active_lines;
680  }
681  }
682  }
683  desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f);
684  if (active_lines > 0.0f)
685  reduction += calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines);
686 
687  pe = 0.0f;
688  for (w = 0; w < wi->num_windows*16; w += 16) {
689  for (g = 0; g < num_bands; g++) {
690  AacPsyBand *band = &pch->band[w+g];
691 
692  if (active_lines > 0.0f)
693  band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
694  pe += calc_pe_3gpp(band);
695  band->norm_fac = band->active_lines / band->thr;
696  norm_fac += band->norm_fac;
697  }
698  }
699  delta_pe = desired_pe - pe;
700  if (fabs(delta_pe) > 0.05f * desired_pe)
701  break;
702  }
703 
704  if (pe < 1.15f * desired_pe) {
705  /* 6.6.1.3.6 "Final threshold modification by linearization" */
706  norm_fac = 1.0f / norm_fac;
707  for (w = 0; w < wi->num_windows*16; w += 16) {
708  for (g = 0; g < num_bands; g++) {
709  AacPsyBand *band = &pch->band[w+g];
710 
711  if (band->active_lines > 0.5f) {
712  float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe;
713  float thr = band->thr;
714 
715  thr *= exp2f(delta_sfb_pe / band->active_lines);
716  if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE)
717  thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy);
718  band->thr = thr;
719  }
720  }
721  }
722  } else {
723  /* 5.6.1.3.7 "Further perceptual entropy reduction" */
724  g = num_bands;
725  while (pe > desired_pe && g--) {
726  for (w = 0; w < wi->num_windows*16; w+= 16) {
727  AacPsyBand *band = &pch->band[w+g];
728  if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) {
729  coeffs[g].min_snr = PSY_SNR_1DB;
730  band->thr = band->energy * PSY_SNR_1DB;
731  pe += band->active_lines * 1.5f - band->pe;
732  }
733  }
734  }
735  /* TODO: allow more holes (unused without mid/side) */
736  }
737  }
738 
739  for (w = 0; w < wi->num_windows*16; w += 16) {
740  for (g = 0; g < num_bands; g++) {
741  AacPsyBand *band = &pch->band[w+g];
742  FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g];
743 
744  psy_band->threshold = band->thr;
745  psy_band->energy = band->energy;
746  }
747  }
748 
749  memcpy(pch->prev_band, pch->band, sizeof(pch->band));
750 }
751 
752 static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
753  const float **coeffs, const FFPsyWindowInfo *wi)
754 {
755  int ch;
756  FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel);
757 
758  for (ch = 0; ch < group->num_ch; ch++)
759  psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]);
760 }
761 
763 {
765  av_freep(&pctx->ch);
766  av_freep(&apc->model_priv_data);
767 }
768 
769 static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
770 {
771  int blocktype = ONLY_LONG_SEQUENCE;
772  if (uselongblock) {
774  blocktype = LONG_STOP_SEQUENCE;
775  } else {
776  blocktype = EIGHT_SHORT_SEQUENCE;
781  }
782 
783  wi->window_type[0] = ctx->next_window_seq;
784  ctx->next_window_seq = blocktype;
785 }
786 
787 static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio,
788  const float *la, int channel, int prev_type)
789 {
791  AacPsyChannel *pch = &pctx->ch[channel];
792  int grouping = 0;
793  int uselongblock = 1;
794  int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
795  int i;
796  FFPsyWindowInfo wi = { { 0 } };
797 
798  if (la) {
799  float hpfsmpl[AAC_BLOCK_SIZE_LONG];
800  float const *pf = hpfsmpl;
801  float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
802  float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
803  float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
804  const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN);
805  int j, att_sum = 0;
806 
807  /* LAME comment: apply high pass filter of fs/4 */
808  for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
809  float sum1, sum2;
810  sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2];
811  sum2 = 0.0;
812  for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
813  sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]);
814  sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]);
815  }
816  /* NOTE: The LAME psymodel expects it's input in the range -32768 to 32768. Tuning this for normalized floats would be difficult. */
817  hpfsmpl[i] = (sum1 + sum2) * 32768.0f;
818  }
819 
820  /* Calculate the energies of each sub-shortblock */
821  for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) {
822  energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)];
823  assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0);
824  attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)];
825  energy_short[0] += energy_subshort[i];
826  }
827 
828  for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) {
829  float const *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS);
830  float p = 1.0f;
831  for (; pf < pfe; pf++)
832  p = FFMAX(p, fabsf(*pf));
833  pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
834  energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
835  /* NOTE: The indexes below are [i + 3 - 2] in the LAME source.
836  * Obviously the 3 and 2 have some significance, or this would be just [i + 1]
837  * (which is what we use here). What the 3 stands for is ambiguous, as it is both
838  * number of short blocks, and the number of sub-short blocks.
839  * It seems that LAME is comparing each sub-block to sub-block + 1 in the
840  * previous block.
841  */
842  if (p > energy_subshort[i + 1])
843  p = p / energy_subshort[i + 1];
844  else if (energy_subshort[i + 1] > p * 10.0f)
845  p = energy_subshort[i + 1] / (p * 10.0f);
846  else
847  p = 0.0;
848  attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
849  }
850 
851  /* compare energy between sub-short blocks */
852  for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++)
853  if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS])
854  if (attack_intensity[i] > pch->attack_threshold)
855  attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1;
856 
857  /* should have energy change between short blocks, in order to avoid periodic signals */
858  /* Good samples to show the effect are Trumpet test songs */
859  /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */
860  /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */
861  for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) {
862  float const u = energy_short[i - 1];
863  float const v = energy_short[i];
864  float const m = FFMAX(u, v);
865  if (m < 40000) { /* (2) */
866  if (u < 1.7f * v && v < 1.7f * u) { /* (1) */
867  if (i == 1 && attacks[0] < attacks[i])
868  attacks[0] = 0;
869  attacks[i] = 0;
870  }
871  }
872  att_sum += attacks[i];
873  }
874 
875  if (attacks[0] <= pch->prev_attack)
876  attacks[0] = 0;
877 
878  att_sum += attacks[0];
879  /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */
880  if (pch->prev_attack == 3 || att_sum) {
881  uselongblock = 0;
882 
883  for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++)
884  if (attacks[i] && attacks[i-1])
885  attacks[i] = 0;
886  }
887  } else {
888  /* We have no lookahead info, so just use same type as the previous sequence. */
889  uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE);
890  }
891 
892  lame_apply_block_type(pch, &wi, uselongblock);
893 
894  wi.window_type[1] = prev_type;
895  if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
896  wi.num_windows = 1;
897  wi.grouping[0] = 1;
898  if (wi.window_type[0] == LONG_START_SEQUENCE)
899  wi.window_shape = 0;
900  else
901  wi.window_shape = 1;
902  } else {
903  int lastgrp = 0;
904 
905  wi.num_windows = 8;
906  wi.window_shape = 0;
907  for (i = 0; i < 8; i++) {
908  if (!((pch->next_grouping >> i) & 1))
909  lastgrp = i;
910  wi.grouping[lastgrp]++;
911  }
912  }
913 
914  /* Determine grouping, based on the location of the first attack, and save for
915  * the next frame.
916  * FIXME: Move this to analysis.
917  * TODO: Tune groupings depending on attack location
918  * TODO: Handle more than one attack in a group
919  */
920  for (i = 0; i < 9; i++) {
921  if (attacks[i]) {
922  grouping = i;
923  break;
924  }
925  }
926  pch->next_grouping = window_grouping[grouping];
927 
928  pch->prev_attack = attacks[8];
929 
930  return wi;
931 }
932 
934 {
935  .name = "3GPP TS 26.403-inspired model",
936  .init = psy_3gpp_init,
937  .window = psy_lame_window,
938  .analyze = psy_3gpp_analyze,
939  .end = psy_3gpp_end,
940 };