FFmpeg
 All Data Structures Namespaces 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/attributes.h"
28 #include "libavutil/libm.h"
29 
30 #include "avcodec.h"
31 #include "aactab.h"
32 #include "psymodel.h"
33 
34 /***********************************
35  * TODOs:
36  * try other bitrate controlling mechanism (maybe use ratecontrol.c?)
37  * control quality for quality-based output
38  **********************************/
39 
40 /**
41  * constants for 3GPP AAC psychoacoustic model
42  * @{
43  */
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
56 
57 #define PSY_3GPP_RPEMIN 0.01f
58 #define PSY_3GPP_RPELEV 2.0f
59 
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 */
63 
64 #define PSY_SNR_1DB 7.9432821e-1f /* -1dB */
65 #define PSY_SNR_25DB 3.1622776e-3f /* -25dB */
66 
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
79 
80 #define PSY_3GPP_AH_THR_LONG 0.5f
81 #define PSY_3GPP_AH_THR_SHORT 0.63f
82 
83 enum {
87 };
88 
89 #define PSY_3GPP_BITS_TO_PE(bits) ((bits) * 1.18f)
90 
91 /* LAME psy model constants */
92 #define PSY_LAME_FIR_LEN 21 ///< LAME psy model FIR order
93 #define AAC_BLOCK_SIZE_LONG 1024 ///< long block size
94 #define AAC_BLOCK_SIZE_SHORT 128 ///< short block size
95 #define AAC_NUM_BLOCKS_SHORT 8 ///< number of blocks in a short sequence
96 #define PSY_LAME_NUM_SUBBLOCKS 3 ///< Number of sub-blocks in each short block
97 
98 /**
99  * @}
100  */
101 
102 /**
103  * information for single band used by 3GPP TS26.403-inspired psychoacoustic model
104  */
105 typedef struct AacPsyBand{
106  float energy; ///< band energy
107  float thr; ///< energy threshold
108  float thr_quiet; ///< threshold in quiet
109  float nz_lines; ///< number of non-zero spectral lines
110  float active_lines; ///< number of active spectral lines
111  float pe; ///< perceptual entropy
112  float pe_const; ///< constant part of the PE calculation
113  float norm_fac; ///< normalization factor for linearization
114  int avoid_holes; ///< hole avoidance flag
115 }AacPsyBand;
116 
117 /**
118  * single/pair channel context for psychoacoustic model
119  */
120 typedef struct AacPsyChannel{
121  AacPsyBand band[128]; ///< bands information
122  AacPsyBand prev_band[128]; ///< bands information from the previous frame
123 
124  float win_energy; ///< sliding average of channel energy
125  float iir_state[2]; ///< hi-pass IIR filter state
126  uint8_t next_grouping; ///< stored grouping scheme for the next frame (in case of 8 short window sequence)
127  enum WindowSequence next_window_seq; ///< window sequence to be used in the next frame
128  /* LAME psy model specific members */
129  float attack_threshold; ///< attack threshold for this channel
131  int prev_attack; ///< attack value for the last short block in the previous sequence
133 
134 /**
135  * psychoacoustic model frame type-dependent coefficients
136  */
137 typedef struct AacPsyCoeffs{
138  float ath; ///< absolute threshold of hearing per bands
139  float barks; ///< Bark value for each spectral band in long frame
140  float spread_low[2]; ///< spreading factor for low-to-high threshold spreading in long frame
141  float spread_hi [2]; ///< spreading factor for high-to-low threshold spreading in long frame
142  float min_snr; ///< minimal SNR
143 }AacPsyCoeffs;
144 
145 /**
146  * 3GPP TS26.403-inspired psychoacoustic model specific data
147  */
148 typedef struct AacPsyContext{
149  int chan_bitrate; ///< bitrate per channel
150  int frame_bits; ///< average bits per frame
151  int fill_level; ///< bit reservoir fill level
152  struct {
153  float min; ///< minimum allowed PE for bit factor calculation
154  float max; ///< maximum allowed PE for bit factor calculation
155  float previous; ///< allowed PE of the previous frame
156  float correction; ///< PE correction factor
157  } pe;
161 
162 /**
163  * LAME psy model preset struct
164  */
165 typedef struct PsyLamePreset {
166  int quality; ///< Quality to map the rest of the vaules to.
167  /* This is overloaded to be both kbps per channel in ABR mode, and
168  * requested quality in constant quality mode.
169  */
170  float st_lrm; ///< short threshold for L, R, and M channels
171 } PsyLamePreset;
172 
173 /**
174  * LAME psy model preset table for ABR
175  */
176 static const PsyLamePreset psy_abr_map[] = {
177 /* TODO: Tuning. These were taken from LAME. */
178 /* kbps/ch st_lrm */
179  { 8, 6.60},
180  { 16, 6.60},
181  { 24, 6.60},
182  { 32, 6.60},
183  { 40, 6.60},
184  { 48, 6.60},
185  { 56, 6.60},
186  { 64, 6.40},
187  { 80, 6.00},
188  { 96, 5.60},
189  {112, 5.20},
190  {128, 5.20},
191  {160, 5.20}
192 };
193 
194 /**
195 * LAME psy model preset table for constant quality
196 */
197 static const PsyLamePreset psy_vbr_map[] = {
198 /* vbr_q st_lrm */
199  { 0, 4.20},
200  { 1, 4.20},
201  { 2, 4.20},
202  { 3, 4.20},
203  { 4, 4.20},
204  { 5, 4.20},
205  { 6, 4.20},
206  { 7, 4.20},
207  { 8, 4.20},
208  { 9, 4.20},
209  {10, 4.20}
210 };
211 
212 /**
213  * LAME psy model FIR coefficient table
214  */
215 static const float psy_fir_coeffs[] = {
216  -8.65163e-18 * 2, -0.00851586 * 2, -6.74764e-18 * 2, 0.0209036 * 2,
217  -3.36639e-17 * 2, -0.0438162 * 2, -1.54175e-17 * 2, 0.0931738 * 2,
218  -5.52212e-17 * 2, -0.313819 * 2
219 };
220 
221 #if ARCH_MIPS
222 # include "mips/aacpsy_mips.h"
223 #endif /* ARCH_MIPS */
224 
225 /**
226  * Calculate the ABR attack threshold from the above LAME psymodel table.
227  */
228 static float lame_calc_attack_threshold(int bitrate)
229 {
230  /* Assume max bitrate to start with */
231  int lower_range = 12, upper_range = 12;
232  int lower_range_kbps = psy_abr_map[12].quality;
233  int upper_range_kbps = psy_abr_map[12].quality;
234  int i;
235 
236  /* Determine which bitrates the value specified falls between.
237  * If the loop ends without breaking our above assumption of 320kbps was correct.
238  */
239  for (i = 1; i < 13; i++) {
240  if (FFMAX(bitrate, psy_abr_map[i].quality) != bitrate) {
241  upper_range = i;
242  upper_range_kbps = psy_abr_map[i ].quality;
243  lower_range = i - 1;
244  lower_range_kbps = psy_abr_map[i - 1].quality;
245  break; /* Upper range found */
246  }
247  }
248 
249  /* Determine which range the value specified is closer to */
250  if ((upper_range_kbps - bitrate) > (bitrate - lower_range_kbps))
251  return psy_abr_map[lower_range].st_lrm;
252  return psy_abr_map[upper_range].st_lrm;
253 }
254 
255 /**
256  * LAME psy model specific initialization
257  */
259 {
260  int i, j;
261 
262  for (i = 0; i < avctx->channels; i++) {
263  AacPsyChannel *pch = &ctx->ch[i];
264 
265  if (avctx->flags & CODEC_FLAG_QSCALE)
266  pch->attack_threshold = psy_vbr_map[avctx->global_quality / FF_QP2LAMBDA].st_lrm;
267  else
268  pch->attack_threshold = lame_calc_attack_threshold(avctx->bit_rate / avctx->channels / 1000);
269 
270  for (j = 0; j < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; j++)
271  pch->prev_energy_subshort[j] = 10.0f;
272  }
273 }
274 
275 /**
276  * Calculate Bark value for given line.
277  */
278 static av_cold float calc_bark(float f)
279 {
280  return 13.3f * atanf(0.00076f * f) + 3.5f * atanf((f / 7500.0f) * (f / 7500.0f));
281 }
282 
283 #define ATH_ADD 4
284 /**
285  * Calculate ATH value for given frequency.
286  * Borrowed from Lame.
287  */
288 static av_cold float ath(float f, float add)
289 {
290  f /= 1000.0f;
291  return 3.64 * pow(f, -0.8)
292  - 6.8 * exp(-0.6 * (f - 3.4) * (f - 3.4))
293  + 6.0 * exp(-0.15 * (f - 8.7) * (f - 8.7))
294  + (0.6 + 0.04 * add) * 0.001 * f * f * f * f;
295 }
296 
298  AacPsyContext *pctx;
299  float bark;
300  int i, j, g, start;
301  float prev, minscale, minath, minsnr, pe_min;
302  const int chan_bitrate = ctx->avctx->bit_rate / ctx->avctx->channels;
303  const int bandwidth = ctx->avctx->cutoff ? ctx->avctx->cutoff : AAC_CUTOFF(ctx->avctx);
304  const float num_bark = calc_bark((float)bandwidth);
305 
306  ctx->model_priv_data = av_mallocz(sizeof(AacPsyContext));
307  if (!ctx->model_priv_data)
308  return AVERROR(ENOMEM);
309  pctx = (AacPsyContext*) ctx->model_priv_data;
310 
311  pctx->chan_bitrate = chan_bitrate;
312  pctx->frame_bits = chan_bitrate * AAC_BLOCK_SIZE_LONG / ctx->avctx->sample_rate;
313  pctx->pe.min = 8.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
314  pctx->pe.max = 12.0f * AAC_BLOCK_SIZE_LONG * bandwidth / (ctx->avctx->sample_rate * 2.0f);
315  ctx->bitres.size = 6144 - pctx->frame_bits;
316  ctx->bitres.size -= ctx->bitres.size % 8;
317  pctx->fill_level = ctx->bitres.size;
318  minath = ath(3410 - 0.733 * ATH_ADD, ATH_ADD);
319  for (j = 0; j < 2; j++) {
320  AacPsyCoeffs *coeffs = pctx->psy_coef[j];
321  const uint8_t *band_sizes = ctx->bands[j];
322  float line_to_frequency = ctx->avctx->sample_rate / (j ? 256.f : 2048.0f);
323  float avg_chan_bits = chan_bitrate * (j ? 128.0f : 1024.0f) / ctx->avctx->sample_rate;
324  /* reference encoder uses 2.4% here instead of 60% like the spec says */
325  float bark_pe = 0.024f * PSY_3GPP_BITS_TO_PE(avg_chan_bits) / num_bark;
326  float en_spread_low = j ? PSY_3GPP_EN_SPREAD_LOW_S : PSY_3GPP_EN_SPREAD_LOW_L;
327  /* High energy spreading for long blocks <= 22kbps/channel and short blocks are the same. */
328  float en_spread_hi = (j || (chan_bitrate <= 22.0f)) ? PSY_3GPP_EN_SPREAD_HI_S : PSY_3GPP_EN_SPREAD_HI_L1;
329 
330  i = 0;
331  prev = 0.0;
332  for (g = 0; g < ctx->num_bands[j]; g++) {
333  i += band_sizes[g];
334  bark = calc_bark((i-1) * line_to_frequency);
335  coeffs[g].barks = (bark + prev) / 2.0;
336  prev = bark;
337  }
338  for (g = 0; g < ctx->num_bands[j] - 1; g++) {
339  AacPsyCoeffs *coeff = &coeffs[g];
340  float bark_width = coeffs[g+1].barks - coeffs->barks;
341  coeff->spread_low[0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_LOW);
342  coeff->spread_hi [0] = pow(10.0, -bark_width * PSY_3GPP_THR_SPREAD_HI);
343  coeff->spread_low[1] = pow(10.0, -bark_width * en_spread_low);
344  coeff->spread_hi [1] = pow(10.0, -bark_width * en_spread_hi);
345  pe_min = bark_pe * bark_width;
346  minsnr = exp2(pe_min / band_sizes[g]) - 1.5f;
347  coeff->min_snr = av_clipf(1.0f / minsnr, PSY_SNR_25DB, PSY_SNR_1DB);
348  }
349  start = 0;
350  for (g = 0; g < ctx->num_bands[j]; g++) {
351  minscale = ath(start * line_to_frequency, ATH_ADD);
352  for (i = 1; i < band_sizes[g]; i++)
353  minscale = FFMIN(minscale, ath((start + i) * line_to_frequency, ATH_ADD));
354  coeffs[g].ath = minscale - minath;
355  start += band_sizes[g];
356  }
357  }
358 
359  pctx->ch = av_mallocz_array(ctx->avctx->channels, sizeof(AacPsyChannel));
360  if (!pctx->ch) {
361  av_freep(&ctx->model_priv_data);
362  return AVERROR(ENOMEM);
363  }
364 
365  lame_window_init(pctx, ctx->avctx);
366 
367  return 0;
368 }
369 
370 /**
371  * IIR filter used in block switching decision
372  */
373 static float iir_filter(int in, float state[2])
374 {
375  float ret;
376 
377  ret = 0.7548f * (in - state[0]) + 0.5095f * state[1];
378  state[0] = in;
379  state[1] = ret;
380  return ret;
381 }
382 
383 /**
384  * window grouping information stored as bits (0 - new group, 1 - group continues)
385  */
386 static const uint8_t window_grouping[9] = {
387  0xB6, 0x6C, 0xD8, 0xB2, 0x66, 0xC6, 0x96, 0x36, 0x36
388 };
389 
390 /**
391  * Tell encoder which window types to use.
392  * @see 3GPP TS26.403 5.4.1 "Blockswitching"
393  */
395  const int16_t *audio,
396  const int16_t *la,
397  int channel, int prev_type)
398 {
399  int i, j;
400  int br = ctx->avctx->bit_rate / ctx->avctx->channels;
401  int attack_ratio = br <= 16000 ? 18 : 10;
403  AacPsyChannel *pch = &pctx->ch[channel];
404  uint8_t grouping = 0;
405  int next_type = pch->next_window_seq;
406  FFPsyWindowInfo wi = { { 0 } };
407 
408  if (la) {
409  float s[8], v;
410  int switch_to_eight = 0;
411  float sum = 0.0, sum2 = 0.0;
412  int attack_n = 0;
413  int stay_short = 0;
414  for (i = 0; i < 8; i++) {
415  for (j = 0; j < 128; j++) {
416  v = iir_filter(la[i*128+j], pch->iir_state);
417  sum += v*v;
418  }
419  s[i] = sum;
420  sum2 += sum;
421  }
422  for (i = 0; i < 8; i++) {
423  if (s[i] > pch->win_energy * attack_ratio) {
424  attack_n = i + 1;
425  switch_to_eight = 1;
426  break;
427  }
428  }
429  pch->win_energy = pch->win_energy*7/8 + sum2/64;
430 
431  wi.window_type[1] = prev_type;
432  switch (prev_type) {
433  case ONLY_LONG_SEQUENCE:
434  wi.window_type[0] = switch_to_eight ? LONG_START_SEQUENCE : ONLY_LONG_SEQUENCE;
435  next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : ONLY_LONG_SEQUENCE;
436  break;
437  case LONG_START_SEQUENCE:
438  wi.window_type[0] = EIGHT_SHORT_SEQUENCE;
439  grouping = pch->next_grouping;
440  next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
441  break;
442  case LONG_STOP_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;
445  break;
447  stay_short = next_type == EIGHT_SHORT_SEQUENCE || switch_to_eight;
448  wi.window_type[0] = stay_short ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
449  grouping = next_type == EIGHT_SHORT_SEQUENCE ? pch->next_grouping : 0;
450  next_type = switch_to_eight ? EIGHT_SHORT_SEQUENCE : LONG_STOP_SEQUENCE;
451  break;
452  }
453 
454  pch->next_grouping = window_grouping[attack_n];
455  pch->next_window_seq = next_type;
456  } else {
457  for (i = 0; i < 3; i++)
458  wi.window_type[i] = prev_type;
459  grouping = (prev_type == EIGHT_SHORT_SEQUENCE) ? window_grouping[0] : 0;
460  }
461 
462  wi.window_shape = 1;
463  if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
464  wi.num_windows = 1;
465  wi.grouping[0] = 1;
466  } else {
467  int lastgrp = 0;
468  wi.num_windows = 8;
469  for (i = 0; i < 8; i++) {
470  if (!((grouping >> i) & 1))
471  lastgrp = i;
472  wi.grouping[lastgrp]++;
473  }
474  }
475 
476  return wi;
477 }
478 
479 /* 5.6.1.2 "Calculation of Bit Demand" */
480 static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size,
481  int short_window)
482 {
483  const float bitsave_slope = short_window ? PSY_3GPP_SAVE_SLOPE_S : PSY_3GPP_SAVE_SLOPE_L;
484  const float bitsave_add = short_window ? PSY_3GPP_SAVE_ADD_S : PSY_3GPP_SAVE_ADD_L;
485  const float bitspend_slope = short_window ? PSY_3GPP_SPEND_SLOPE_S : PSY_3GPP_SPEND_SLOPE_L;
486  const float bitspend_add = short_window ? PSY_3GPP_SPEND_ADD_S : PSY_3GPP_SPEND_ADD_L;
487  const float clip_low = short_window ? PSY_3GPP_CLIP_LO_S : PSY_3GPP_CLIP_LO_L;
488  const float clip_high = short_window ? PSY_3GPP_CLIP_HI_S : PSY_3GPP_CLIP_HI_L;
489  float clipped_pe, bit_save, bit_spend, bit_factor, fill_level;
490 
491  ctx->fill_level += ctx->frame_bits - bits;
492  ctx->fill_level = av_clip(ctx->fill_level, 0, size);
493  fill_level = av_clipf((float)ctx->fill_level / size, clip_low, clip_high);
494  clipped_pe = av_clipf(pe, ctx->pe.min, ctx->pe.max);
495  bit_save = (fill_level + bitsave_add) * bitsave_slope;
496  assert(bit_save <= 0.3f && bit_save >= -0.05000001f);
497  bit_spend = (fill_level + bitspend_add) * bitspend_slope;
498  assert(bit_spend <= 0.5f && bit_spend >= -0.1f);
499  /* The bit factor graph in the spec is obviously incorrect.
500  * bit_spend + ((bit_spend - bit_spend))...
501  * The reference encoder subtracts everything from 1, but also seems incorrect.
502  * 1 - bit_save + ((bit_spend + bit_save))...
503  * Hopefully below is correct.
504  */
505  bit_factor = 1.0f - bit_save + ((bit_spend - bit_save) / (ctx->pe.max - ctx->pe.min)) * (clipped_pe - ctx->pe.min);
506  /* NOTE: The reference encoder attempts to center pe max/min around the current pe. */
507  ctx->pe.max = FFMAX(pe, ctx->pe.max);
508  ctx->pe.min = FFMIN(pe, ctx->pe.min);
509 
510  return FFMIN(ctx->frame_bits * bit_factor, ctx->frame_bits + size - bits);
511 }
512 
514 {
515  float pe, a;
516 
517  band->pe = 0.0f;
518  band->pe_const = 0.0f;
519  band->active_lines = 0.0f;
520  if (band->energy > band->thr) {
521  a = log2f(band->energy);
522  pe = a - log2f(band->thr);
523  band->active_lines = band->nz_lines;
524  if (pe < PSY_3GPP_C1) {
525  pe = pe * PSY_3GPP_C3 + PSY_3GPP_C2;
526  a = a * PSY_3GPP_C3 + PSY_3GPP_C2;
527  band->active_lines *= PSY_3GPP_C3;
528  }
529  band->pe = pe * band->nz_lines;
530  band->pe_const = a * band->nz_lines;
531  }
532 
533  return band->pe;
534 }
535 
536 static float calc_reduction_3gpp(float a, float desired_pe, float pe,
537  float active_lines)
538 {
539  float thr_avg, reduction;
540 
541  if(active_lines == 0.0)
542  return 0;
543 
544  thr_avg = exp2f((a - pe) / (4.0f * active_lines));
545  reduction = exp2f((a - desired_pe) / (4.0f * active_lines)) - thr_avg;
546 
547  return FFMAX(reduction, 0.0f);
548 }
549 
550 static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr,
551  float reduction)
552 {
553  float thr = band->thr;
554 
555  if (band->energy > thr) {
556  thr = sqrtf(thr);
557  thr = sqrtf(thr) + reduction;
558  thr *= thr;
559  thr *= thr;
560 
561  /* This deviates from the 3GPP spec to match the reference encoder.
562  * It performs min(thr_reduced, max(thr, energy/min_snr)) only for bands
563  * that have hole avoidance on (active or inactive). It always reduces the
564  * threshold of bands with hole avoidance off.
565  */
566  if (thr > band->energy * min_snr && band->avoid_holes != PSY_3GPP_AH_NONE) {
567  thr = FFMAX(band->thr, band->energy * min_snr);
569  }
570  }
571 
572  return thr;
573 }
574 
575 #ifndef calc_thr_3gpp
576 static void calc_thr_3gpp(const FFPsyWindowInfo *wi, const int num_bands, AacPsyChannel *pch,
577  const uint8_t *band_sizes, const float *coefs)
578 {
579  int i, w, g;
580  int start = 0;
581  for (w = 0; w < wi->num_windows*16; w += 16) {
582  for (g = 0; g < num_bands; g++) {
583  AacPsyBand *band = &pch->band[w+g];
584 
585  float form_factor = 0.0f;
586  float Temp;
587  band->energy = 0.0f;
588  for (i = 0; i < band_sizes[g]; i++) {
589  band->energy += coefs[start+i] * coefs[start+i];
590  form_factor += sqrtf(fabs(coefs[start+i]));
591  }
592  Temp = band->energy > 0 ? sqrtf((float)band_sizes[g] / band->energy) : 0;
593  band->thr = band->energy * 0.001258925f;
594  band->nz_lines = form_factor * sqrtf(Temp);
595 
596  start += band_sizes[g];
597  }
598  }
599 }
600 #endif /* calc_thr_3gpp */
601 
602 #ifndef psy_hp_filter
603 static void psy_hp_filter(const float *firbuf, float *hpfsmpl, const float *psy_fir_coeffs)
604 {
605  int i, j;
606  for (i = 0; i < AAC_BLOCK_SIZE_LONG; i++) {
607  float sum1, sum2;
608  sum1 = firbuf[i + (PSY_LAME_FIR_LEN - 1) / 2];
609  sum2 = 0.0;
610  for (j = 0; j < ((PSY_LAME_FIR_LEN - 1) / 2) - 1; j += 2) {
611  sum1 += psy_fir_coeffs[j] * (firbuf[i + j] + firbuf[i + PSY_LAME_FIR_LEN - j]);
612  sum2 += psy_fir_coeffs[j + 1] * (firbuf[i + j + 1] + firbuf[i + PSY_LAME_FIR_LEN - j - 1]);
613  }
614  /* NOTE: The LAME psymodel expects it's input in the range -32768 to 32768.
615  * Tuning this for normalized floats would be difficult. */
616  hpfsmpl[i] = (sum1 + sum2) * 32768.0f;
617  }
618 }
619 #endif /* psy_hp_filter */
620 
621 /**
622  * Calculate band thresholds as suggested in 3GPP TS26.403
623  */
624 static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel,
625  const float *coefs, const FFPsyWindowInfo *wi)
626 {
628  AacPsyChannel *pch = &pctx->ch[channel];
629  int i, w, g;
630  float desired_bits, desired_pe, delta_pe, reduction= NAN, spread_en[128] = {0};
631  float a = 0.0f, active_lines = 0.0f, norm_fac = 0.0f;
632  float pe = pctx->chan_bitrate > 32000 ? 0.0f : FFMAX(50.0f, 100.0f - pctx->chan_bitrate * 100.0f / 32000.0f);
633  const int num_bands = ctx->num_bands[wi->num_windows == 8];
634  const uint8_t *band_sizes = ctx->bands[wi->num_windows == 8];
635  AacPsyCoeffs *coeffs = pctx->psy_coef[wi->num_windows == 8];
636  const float avoid_hole_thr = wi->num_windows == 8 ? PSY_3GPP_AH_THR_SHORT : PSY_3GPP_AH_THR_LONG;
637 
638  //calculate energies, initial thresholds and related values - 5.4.2 "Threshold Calculation"
639  calc_thr_3gpp(wi, num_bands, pch, band_sizes, coefs);
640 
641  //modify thresholds and energies - spread, threshold in quiet, pre-echo control
642  for (w = 0; w < wi->num_windows*16; w += 16) {
643  AacPsyBand *bands = &pch->band[w];
644 
645  /* 5.4.2.3 "Spreading" & 5.4.3 "Spread Energy Calculation" */
646  spread_en[0] = bands[0].energy;
647  for (g = 1; g < num_bands; g++) {
648  bands[g].thr = FFMAX(bands[g].thr, bands[g-1].thr * coeffs[g].spread_hi[0]);
649  spread_en[w+g] = FFMAX(bands[g].energy, spread_en[w+g-1] * coeffs[g].spread_hi[1]);
650  }
651  for (g = num_bands - 2; g >= 0; g--) {
652  bands[g].thr = FFMAX(bands[g].thr, bands[g+1].thr * coeffs[g].spread_low[0]);
653  spread_en[w+g] = FFMAX(spread_en[w+g], spread_en[w+g+1] * coeffs[g].spread_low[1]);
654  }
655  //5.4.2.4 "Threshold in quiet"
656  for (g = 0; g < num_bands; g++) {
657  AacPsyBand *band = &bands[g];
658 
659  band->thr_quiet = band->thr = FFMAX(band->thr, coeffs[g].ath);
660  //5.4.2.5 "Pre-echo control"
661  if (!(wi->window_type[0] == LONG_STOP_SEQUENCE || (wi->window_type[1] == LONG_START_SEQUENCE && !w)))
662  band->thr = FFMAX(PSY_3GPP_RPEMIN*band->thr, FFMIN(band->thr,
663  PSY_3GPP_RPELEV*pch->prev_band[w+g].thr_quiet));
664 
665  /* 5.6.1.3.1 "Preparatory steps of the perceptual entropy calculation" */
666  pe += calc_pe_3gpp(band);
667  a += band->pe_const;
668  active_lines += band->active_lines;
669 
670  /* 5.6.1.3.3 "Selection of the bands for avoidance of holes" */
671  if (spread_en[w+g] * avoid_hole_thr > band->energy || coeffs[g].min_snr > 1.0f)
673  else
675  }
676  }
677 
678  /* 5.6.1.3.2 "Calculation of the desired perceptual entropy" */
679  ctx->ch[channel].entropy = pe;
680  desired_bits = calc_bit_demand(pctx, pe, ctx->bitres.bits, ctx->bitres.size, wi->num_windows == 8);
681  desired_pe = PSY_3GPP_BITS_TO_PE(desired_bits);
682  /* NOTE: PE correction is kept simple. During initial testing it had very
683  * little effect on the final bitrate. Probably a good idea to come
684  * back and do more testing later.
685  */
686  if (ctx->bitres.bits > 0)
687  desired_pe *= av_clipf(pctx->pe.previous / PSY_3GPP_BITS_TO_PE(ctx->bitres.bits),
688  0.85f, 1.15f);
689  pctx->pe.previous = PSY_3GPP_BITS_TO_PE(desired_bits);
690 
691  if (desired_pe < pe) {
692  /* 5.6.1.3.4 "First Estimation of the reduction value" */
693  for (w = 0; w < wi->num_windows*16; w += 16) {
694  reduction = calc_reduction_3gpp(a, desired_pe, pe, active_lines);
695  pe = 0.0f;
696  a = 0.0f;
697  active_lines = 0.0f;
698  for (g = 0; g < num_bands; g++) {
699  AacPsyBand *band = &pch->band[w+g];
700 
701  band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
702  /* recalculate PE */
703  pe += calc_pe_3gpp(band);
704  a += band->pe_const;
705  active_lines += band->active_lines;
706  }
707  }
708 
709  /* 5.6.1.3.5 "Second Estimation of the reduction value" */
710  for (i = 0; i < 2; i++) {
711  float pe_no_ah = 0.0f, desired_pe_no_ah;
712  active_lines = a = 0.0f;
713  for (w = 0; w < wi->num_windows*16; w += 16) {
714  for (g = 0; g < num_bands; g++) {
715  AacPsyBand *band = &pch->band[w+g];
716 
717  if (band->avoid_holes != PSY_3GPP_AH_ACTIVE) {
718  pe_no_ah += band->pe;
719  a += band->pe_const;
720  active_lines += band->active_lines;
721  }
722  }
723  }
724  desired_pe_no_ah = FFMAX(desired_pe - (pe - pe_no_ah), 0.0f);
725  if (active_lines > 0.0f)
726  reduction = calc_reduction_3gpp(a, desired_pe_no_ah, pe_no_ah, active_lines);
727 
728  pe = 0.0f;
729  for (w = 0; w < wi->num_windows*16; w += 16) {
730  for (g = 0; g < num_bands; g++) {
731  AacPsyBand *band = &pch->band[w+g];
732 
733  if (active_lines > 0.0f)
734  band->thr = calc_reduced_thr_3gpp(band, coeffs[g].min_snr, reduction);
735  pe += calc_pe_3gpp(band);
736  if (band->thr > 0.0f)
737  band->norm_fac = band->active_lines / band->thr;
738  else
739  band->norm_fac = 0.0f;
740  norm_fac += band->norm_fac;
741  }
742  }
743  delta_pe = desired_pe - pe;
744  if (fabs(delta_pe) > 0.05f * desired_pe)
745  break;
746  }
747 
748  if (pe < 1.15f * desired_pe) {
749  /* 6.6.1.3.6 "Final threshold modification by linearization" */
750  norm_fac = 1.0f / norm_fac;
751  for (w = 0; w < wi->num_windows*16; w += 16) {
752  for (g = 0; g < num_bands; g++) {
753  AacPsyBand *band = &pch->band[w+g];
754 
755  if (band->active_lines > 0.5f) {
756  float delta_sfb_pe = band->norm_fac * norm_fac * delta_pe;
757  float thr = band->thr;
758 
759  thr *= exp2f(delta_sfb_pe / band->active_lines);
760  if (thr > coeffs[g].min_snr * band->energy && band->avoid_holes == PSY_3GPP_AH_INACTIVE)
761  thr = FFMAX(band->thr, coeffs[g].min_snr * band->energy);
762  band->thr = thr;
763  }
764  }
765  }
766  } else {
767  /* 5.6.1.3.7 "Further perceptual entropy reduction" */
768  g = num_bands;
769  while (pe > desired_pe && g--) {
770  for (w = 0; w < wi->num_windows*16; w+= 16) {
771  AacPsyBand *band = &pch->band[w+g];
772  if (band->avoid_holes != PSY_3GPP_AH_NONE && coeffs[g].min_snr < PSY_SNR_1DB) {
773  coeffs[g].min_snr = PSY_SNR_1DB;
774  band->thr = band->energy * PSY_SNR_1DB;
775  pe += band->active_lines * 1.5f - band->pe;
776  }
777  }
778  }
779  /* TODO: allow more holes (unused without mid/side) */
780  }
781  }
782 
783  for (w = 0; w < wi->num_windows*16; w += 16) {
784  for (g = 0; g < num_bands; g++) {
785  AacPsyBand *band = &pch->band[w+g];
786  FFPsyBand *psy_band = &ctx->ch[channel].psy_bands[w+g];
787 
788  psy_band->threshold = band->thr;
789  psy_band->energy = band->energy;
790  }
791  }
792 
793  memcpy(pch->prev_band, pch->band, sizeof(pch->band));
794 }
795 
796 static void psy_3gpp_analyze(FFPsyContext *ctx, int channel,
797  const float **coeffs, const FFPsyWindowInfo *wi)
798 {
799  int ch;
800  FFPsyChannelGroup *group = ff_psy_find_group(ctx, channel);
801 
802  for (ch = 0; ch < group->num_ch; ch++)
803  psy_3gpp_analyze_channel(ctx, channel + ch, coeffs[ch], &wi[ch]);
804 }
805 
807 {
809  av_freep(&pctx->ch);
810  av_freep(&apc->model_priv_data);
811 }
812 
813 static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
814 {
815  int blocktype = ONLY_LONG_SEQUENCE;
816  if (uselongblock) {
818  blocktype = LONG_STOP_SEQUENCE;
819  } else {
820  blocktype = EIGHT_SHORT_SEQUENCE;
825  }
826 
827  wi->window_type[0] = ctx->next_window_seq;
828  ctx->next_window_seq = blocktype;
829 }
830 
831 static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio,
832  const float *la, int channel, int prev_type)
833 {
835  AacPsyChannel *pch = &pctx->ch[channel];
836  int grouping = 0;
837  int uselongblock = 1;
838  int attacks[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
839  int i;
840  FFPsyWindowInfo wi = { { 0 } };
841 
842  if (la) {
843  float hpfsmpl[AAC_BLOCK_SIZE_LONG];
844  float const *pf = hpfsmpl;
845  float attack_intensity[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
846  float energy_subshort[(AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS];
847  float energy_short[AAC_NUM_BLOCKS_SHORT + 1] = { 0 };
848  const float *firbuf = la + (AAC_BLOCK_SIZE_SHORT/4 - PSY_LAME_FIR_LEN);
849  int att_sum = 0;
850 
851  /* LAME comment: apply high pass filter of fs/4 */
852  psy_hp_filter(firbuf, hpfsmpl, psy_fir_coeffs);
853 
854  /* Calculate the energies of each sub-shortblock */
855  for (i = 0; i < PSY_LAME_NUM_SUBBLOCKS; i++) {
856  energy_subshort[i] = pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 1) * PSY_LAME_NUM_SUBBLOCKS)];
857  assert(pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)] > 0);
858  attack_intensity[i] = energy_subshort[i] / pch->prev_energy_subshort[i + ((AAC_NUM_BLOCKS_SHORT - 2) * PSY_LAME_NUM_SUBBLOCKS + 1)];
859  energy_short[0] += energy_subshort[i];
860  }
861 
862  for (i = 0; i < AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS; i++) {
863  float const *const pfe = pf + AAC_BLOCK_SIZE_LONG / (AAC_NUM_BLOCKS_SHORT * PSY_LAME_NUM_SUBBLOCKS);
864  float p = 1.0f;
865  for (; pf < pfe; pf++)
866  p = FFMAX(p, fabsf(*pf));
867  pch->prev_energy_subshort[i] = energy_subshort[i + PSY_LAME_NUM_SUBBLOCKS] = p;
868  energy_short[1 + i / PSY_LAME_NUM_SUBBLOCKS] += p;
869  /* NOTE: The indexes below are [i + 3 - 2] in the LAME source.
870  * Obviously the 3 and 2 have some significance, or this would be just [i + 1]
871  * (which is what we use here). What the 3 stands for is ambiguous, as it is both
872  * number of short blocks, and the number of sub-short blocks.
873  * It seems that LAME is comparing each sub-block to sub-block + 1 in the
874  * previous block.
875  */
876  if (p > energy_subshort[i + 1])
877  p = p / energy_subshort[i + 1];
878  else if (energy_subshort[i + 1] > p * 10.0f)
879  p = energy_subshort[i + 1] / (p * 10.0f);
880  else
881  p = 0.0;
882  attack_intensity[i + PSY_LAME_NUM_SUBBLOCKS] = p;
883  }
884 
885  /* compare energy between sub-short blocks */
886  for (i = 0; i < (AAC_NUM_BLOCKS_SHORT + 1) * PSY_LAME_NUM_SUBBLOCKS; i++)
887  if (!attacks[i / PSY_LAME_NUM_SUBBLOCKS])
888  if (attack_intensity[i] > pch->attack_threshold)
889  attacks[i / PSY_LAME_NUM_SUBBLOCKS] = (i % PSY_LAME_NUM_SUBBLOCKS) + 1;
890 
891  /* should have energy change between short blocks, in order to avoid periodic signals */
892  /* Good samples to show the effect are Trumpet test songs */
893  /* GB: tuned (1) to avoid too many short blocks for test sample TRUMPET */
894  /* RH: tuned (2) to let enough short blocks through for test sample FSOL and SNAPS */
895  for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++) {
896  float const u = energy_short[i - 1];
897  float const v = energy_short[i];
898  float const m = FFMAX(u, v);
899  if (m < 40000) { /* (2) */
900  if (u < 1.7f * v && v < 1.7f * u) { /* (1) */
901  if (i == 1 && attacks[0] < attacks[i])
902  attacks[0] = 0;
903  attacks[i] = 0;
904  }
905  }
906  att_sum += attacks[i];
907  }
908 
909  if (attacks[0] <= pch->prev_attack)
910  attacks[0] = 0;
911 
912  att_sum += attacks[0];
913  /* 3 below indicates the previous attack happened in the last sub-block of the previous sequence */
914  if (pch->prev_attack == 3 || att_sum) {
915  uselongblock = 0;
916 
917  for (i = 1; i < AAC_NUM_BLOCKS_SHORT + 1; i++)
918  if (attacks[i] && attacks[i-1])
919  attacks[i] = 0;
920  }
921  } else {
922  /* We have no lookahead info, so just use same type as the previous sequence. */
923  uselongblock = !(prev_type == EIGHT_SHORT_SEQUENCE);
924  }
925 
926  lame_apply_block_type(pch, &wi, uselongblock);
927 
928  wi.window_type[1] = prev_type;
929  if (wi.window_type[0] != EIGHT_SHORT_SEQUENCE) {
930  wi.num_windows = 1;
931  wi.grouping[0] = 1;
932  if (wi.window_type[0] == LONG_START_SEQUENCE)
933  wi.window_shape = 0;
934  else
935  wi.window_shape = 1;
936  } else {
937  int lastgrp = 0;
938 
939  wi.num_windows = 8;
940  wi.window_shape = 0;
941  for (i = 0; i < 8; i++) {
942  if (!((pch->next_grouping >> i) & 1))
943  lastgrp = i;
944  wi.grouping[lastgrp]++;
945  }
946  }
947 
948  /* Determine grouping, based on the location of the first attack, and save for
949  * the next frame.
950  * FIXME: Move this to analysis.
951  * TODO: Tune groupings depending on attack location
952  * TODO: Handle more than one attack in a group
953  */
954  for (i = 0; i < 9; i++) {
955  if (attacks[i]) {
956  grouping = i;
957  break;
958  }
959  }
960  pch->next_grouping = window_grouping[grouping];
961 
962  pch->prev_attack = attacks[8];
963 
964  return wi;
965 }
966 
968 {
969  .name = "3GPP TS 26.403-inspired model",
970  .init = psy_3gpp_init,
971  .window = psy_lame_window,
972  .analyze = psy_3gpp_analyze,
973  .end = psy_3gpp_end,
974 };
int quality
Quality to map the rest of the vaules to.
Definition: aacpsy.c:166
float v
const char * s
Definition: avisynth_c.h:631
static const uint8_t window_grouping[9]
window grouping information stored as bits (0 - new group, 1 - group continues)
Definition: aacpsy.c:386
int grouping[8]
window grouping (for e.g. AAC)
Definition: psymodel.h:69
#define AAC_BLOCK_SIZE_SHORT
short block size
Definition: aacpsy.c:94
static int calc_bit_demand(AacPsyContext *ctx, float pe, int bits, int size, int short_window)
Definition: aacpsy.c:480
uint8_t ** bands
scalefactor band sizes for possible frame sizes
Definition: psymodel.h:84
#define PSY_3GPP_AH_THR_SHORT
Definition: aacpsy.c:81
float iir_state[2]
hi-pass IIR filter state
Definition: aacpsy.c:125
const char * g
Definition: vf_curves.c:108
static const PsyLamePreset psy_vbr_map[]
LAME psy model preset table for constant quality.
Definition: aacpsy.c:197
psychoacoustic information for an arbitrary group of channels
Definition: psymodel.h:56
static float calc_reduction_3gpp(float a, float desired_pe, float pe, float active_lines)
Definition: aacpsy.c:536
float ath
absolute threshold of hearing per bands
Definition: aacpsy.c:138
#define PSY_3GPP_EN_SPREAD_HI_L1
Definition: aacpsy.c:47
static av_cold float ath(float f, float add)
Calculate ATH value for given frequency.
Definition: aacpsy.c:288
float prev_energy_subshort[AAC_NUM_BLOCKS_SHORT *PSY_LAME_NUM_SUBBLOCKS]
Definition: aacpsy.c:130
enum WindowSequence next_window_seq
window sequence to be used in the next frame
Definition: aacpsy.c:127
#define PSY_SNR_25DB
Definition: aacpsy.c:65
#define AAC_BLOCK_SIZE_LONG
long block size
Definition: aacpsy.c:93
int * num_bands
number of scalefactor bands for possible frame sizes
Definition: psymodel.h:85
Macro definitions for various function/variable attributes.
LAME psy model preset struct.
Definition: aacpsy.c:165
float thr
energy threshold
Definition: aacpsy.c:107
float correction
PE correction factor.
Definition: aacpsy.c:156
static av_cold void psy_3gpp_end(FFPsyContext *apc)
Definition: aacpsy.c:806
float attack_threshold
attack threshold for this channel
Definition: aacpsy.c:129
#define PSY_3GPP_EN_SPREAD_LOW_L
Definition: aacpsy.c:53
float nz_lines
number of non-zero spectral lines
Definition: aacpsy.c:109
uint8_t bits
Definition: crc.c:295
uint8_t
psychoacoustic model frame type-dependent coefficients
Definition: aacpsy.c:137
#define av_cold
Definition: attributes.h:74
int size
size of the bitresevoir in bits
Definition: psymodel.h:89
static float calc_reduced_thr_3gpp(AacPsyBand *band, float min_snr, float reduction)
Definition: aacpsy.c:550
#define PSY_3GPP_C2
Definition: aacpsy.c:61
#define PSY_LAME_FIR_LEN
LAME psy model FIR order.
Definition: aacpsy.c:92
#define PSY_3GPP_CLIP_LO_L
Definition: aacpsy.c:75
#define PSY_3GPP_SPEND_SLOPE_S
Definition: aacpsy.c:72
#define PSY_3GPP_THR_SPREAD_LOW
Definition: aacpsy.c:45
context used by psychoacoustic model
Definition: psymodel.h:76
#define atanf(x)
Definition: libm.h:38
#define AAC_CUTOFF(s)
Definition: psymodel.h:32
single band psychoacoustic information
Definition: psymodel.h:37
ptrdiff_t size
Definition: opengl_enc.c:101
static float lame_calc_attack_threshold(int bitrate)
Calculate the ABR attack threshold from the above LAME psymodel table.
Definition: aacpsy.c:228
uint8_t next_grouping
stored grouping scheme for the next frame (in case of 8 short window sequence)
Definition: aacpsy.c:126
unsigned m
Definition: audioconvert.c:187
#define PSY_3GPP_SAVE_ADD_L
Definition: aacpsy.c:69
static void calc_thr_3gpp(const FFPsyWindowInfo *wi, const int num_bands, AacPsyChannel *pch, const uint8_t *band_sizes, const float *coefs)
Definition: aacpsy.c:576
static av_cold float calc_bark(float f)
Calculate Bark value for given line.
Definition: aacpsy.c:278
#define AVERROR(e)
Definition: error.h:43
#define PSY_3GPP_SPEND_ADD_S
Definition: aacpsy.c:74
#define PSY_SNR_1DB
Definition: aacpsy.c:64
AacPsyBand prev_band[128]
bands information from the previous frame
Definition: aacpsy.c:122
3GPP TS26.403-inspired psychoacoustic model specific data
Definition: aacpsy.c:148
struct FFPsyContext::@77 bitres
single/pair channel context for psychoacoustic model
Definition: aacpsy.c:120
static const float psy_fir_coeffs[]
LAME psy model FIR coefficient table.
Definition: aacpsy.c:215
float barks
Bark value for each spectral band in long frame.
Definition: aacpsy.c:139
int flags
CODEC_FLAG_*.
Definition: avcodec.h:1335
#define CODEC_FLAG_QSCALE
Use fixed qscale.
Definition: avcodec.h:712
float pe_const
constant part of the PE calculation
Definition: aacpsy.c:112
int num_windows
number of windows in a frame
Definition: psymodel.h:68
static FFPsyWindowInfo psy_lame_window(FFPsyContext *ctx, const float *audio, const float *la, int channel, int prev_type)
Definition: aacpsy.c:831
#define PSY_3GPP_SPEND_SLOPE_L
Definition: aacpsy.c:71
#define PSY_3GPP_THR_SPREAD_HI
constants for 3GPP AAC psychoacoustic model
Definition: aacpsy.c:44
float energy
Definition: psymodel.h:39
WindowSequence
Definition: aac.h:68
#define FFMAX(a, b)
Definition: common.h:64
Libavcodec external API header.
codec-specific psychoacoustic model implementation
Definition: psymodel.h:99
#define PSY_3GPP_RPELEV
Definition: aacpsy.c:58
struct AacPsyContext::@31 pe
float thr_quiet
threshold in quiet
Definition: aacpsy.c:108
static void psy_3gpp_analyze(FFPsyContext *ctx, int channel, const float **coeffs, const FFPsyWindowInfo *wi)
Definition: aacpsy.c:796
int bit_rate
the average bitrate
Definition: avcodec.h:1305
#define FFMIN(a, b)
Definition: common.h:66
int prev_attack
attack value for the last short block in the previous sequence
Definition: aacpsy.c:131
#define PSY_3GPP_SAVE_SLOPE_S
Definition: aacpsy.c:68
ret
Definition: avfilter.c:974
#define PSY_3GPP_C3
Definition: aacpsy.c:62
uint8_t num_ch
number of channels in this group
Definition: psymodel.h:58
int frame_bits
average bits per frame
Definition: aacpsy.c:150
int fill_level
bit reservoir fill level
Definition: aacpsy.c:151
static void lame_apply_block_type(AacPsyChannel *ctx, FFPsyWindowInfo *wi, int uselongblock)
Definition: aacpsy.c:813
#define PSY_3GPP_SAVE_SLOPE_L
Definition: aacpsy.c:67
Reference: libavcodec/aacpsy.c.
float u
#define PSY_LAME_NUM_SUBBLOCKS
Number of sub-blocks in each short block.
Definition: aacpsy.c:96
#define ATH_ADD
Definition: aacpsy.c:283
float energy
band energy
Definition: aacpsy.c:106
const FFPsyModel ff_aac_psy_model
Definition: aacpsy.c:967
static void psy_3gpp_analyze_channel(FFPsyContext *ctx, int channel, const float *coefs, const FFPsyWindowInfo *wi)
Calculate band thresholds as suggested in 3GPP TS26.403.
Definition: aacpsy.c:624
float st_lrm
short threshold for L, R, and M channels
Definition: aacpsy.c:170
#define PSY_3GPP_EN_SPREAD_LOW_S
Definition: aacpsy.c:55
#define exp2f(x)
Definition: libm.h:82
int sample_rate
samples per second
Definition: avcodec.h:1985
FFPsyChannelGroup * ff_psy_find_group(FFPsyContext *ctx, int channel)
Determine what group a channel belongs to.
Definition: psymodel.c:72
main external API structure.
Definition: avcodec.h:1241
float win_energy
sliding average of channel energy
Definition: aacpsy.c:124
void * model_priv_data
psychoacoustic model implementation private data
Definition: psymodel.h:93
float active_lines
number of active spectral lines
Definition: aacpsy.c:110
uint8_t pi<< 24) CONV_FUNC_GROUP(AV_SAMPLE_FMT_FLT, float, AV_SAMPLE_FMT_U8, uint8_t,(*(constuint8_t *) pi-0x80)*(1.0f/(1<< 7))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_DBL, double, AV_SAMPLE_FMT_U8, uint8_t,(*(constuint8_t *) pi-0x80)*(1.0/(1<< 7))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_U8, uint8_t, AV_SAMPLE_FMT_S16, int16_t,(*(constint16_t *) pi >>8)+0x80) CONV_FUNC_GROUP(AV_SAMPLE_FMT_FLT, float, AV_SAMPLE_FMT_S16, int16_t,*(constint16_t *) pi *(1.0f/(1<< 15))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_DBL, double, AV_SAMPLE_FMT_S16, int16_t,*(constint16_t *) pi *(1.0/(1<< 15))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_U8, uint8_t, AV_SAMPLE_FMT_S32, int32_t,(*(constint32_t *) pi >>24)+0x80) CONV_FUNC_GROUP(AV_SAMPLE_FMT_FLT, float, AV_SAMPLE_FMT_S32, int32_t,*(constint32_t *) pi *(1.0f/(1U<< 31))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_DBL, double, AV_SAMPLE_FMT_S32, int32_t,*(constint32_t *) pi *(1.0/(1U<< 31))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_U8, uint8_t, AV_SAMPLE_FMT_FLT, float, av_clip_uint8(lrintf(*(constfloat *) pi *(1<< 7))+0x80)) CONV_FUNC_GROUP(AV_SAMPLE_FMT_S16, int16_t, AV_SAMPLE_FMT_FLT, float, av_clip_int16(lrintf(*(constfloat *) pi *(1<< 15)))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_S32, int32_t, AV_SAMPLE_FMT_FLT, float, av_clipl_int32(llrintf(*(constfloat *) pi *(1U<< 31)))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_U8, uint8_t, AV_SAMPLE_FMT_DBL, double, av_clip_uint8(lrint(*(constdouble *) pi *(1<< 7))+0x80)) CONV_FUNC_GROUP(AV_SAMPLE_FMT_S16, int16_t, AV_SAMPLE_FMT_DBL, double, av_clip_int16(lrint(*(constdouble *) pi *(1<< 15)))) CONV_FUNC_GROUP(AV_SAMPLE_FMT_S32, int32_t, AV_SAMPLE_FMT_DBL, double, av_clipl_int32(llrint(*(constdouble *) pi *(1U<< 31))))#defineSET_CONV_FUNC_GROUP(ofmt, ifmt) staticvoidset_generic_function(AudioConvert *ac){}voidff_audio_convert_free(AudioConvert **ac){if(!*ac) return;ff_dither_free(&(*ac) ->dc);av_freep(ac);}AudioConvert *ff_audio_convert_alloc(AVAudioResampleContext *avr, enumAVSampleFormatout_fmt, enumAVSampleFormatin_fmt, intchannels, intsample_rate, intapply_map){AudioConvert *ac;intin_planar, out_planar;ac=av_mallocz(sizeof(*ac));if(!ac) returnNULL;ac->avr=avr;ac->out_fmt=out_fmt;ac->in_fmt=in_fmt;ac->channels=channels;ac->apply_map=apply_map;if(avr->dither_method!=AV_RESAMPLE_DITHER_NONE &&av_get_packed_sample_fmt(out_fmt)==AV_SAMPLE_FMT_S16 &&av_get_bytes_per_sample(in_fmt)>2){ac->dc=ff_dither_alloc(avr, out_fmt, in_fmt, channels, sample_rate, apply_map);if(!ac->dc){av_free(ac);returnNULL;}returnac;}in_planar=ff_sample_fmt_is_planar(in_fmt, channels);out_planar=ff_sample_fmt_is_planar(out_fmt, channels);if(in_planar==out_planar){ac->func_type=CONV_FUNC_TYPE_FLAT;ac->planes=in_planar?ac->channels:1;}elseif(in_planar) ac->func_type=CONV_FUNC_TYPE_INTERLEAVE;elseac->func_type=CONV_FUNC_TYPE_DEINTERLEAVE;set_generic_function(ac);if(ARCH_AARCH64) ff_audio_convert_init_aarch64(ac);if(ARCH_ARM) ff_audio_convert_init_arm(ac);if(ARCH_X86) ff_audio_convert_init_x86(ac);returnac;}intff_audio_convert(AudioConvert *ac, AudioData *out, AudioData *in){intuse_generic=1;intlen=in->nb_samples;intp;if(ac->dc){av_log(ac->avr, AV_LOG_TRACE,"%dsamples-audio_convert:%sto%s(dithered)\n", len, av_get_sample_fmt_name(ac->in_fmt), av_get_sample_fmt_name(ac->out_fmt));returnff_convert_dither(ac-> in
static float iir_filter(int in, float state[2])
IIR filter used in block switching decision.
Definition: aacpsy.c:373
int avoid_holes
hole avoidance flag
Definition: aacpsy.c:114
Replacements for frequently missing libm functions.
AacPsyBand band[128]
bands information
Definition: aacpsy.c:121
#define PSY_3GPP_CLIP_HI_S
Definition: aacpsy.c:78
#define PSY_3GPP_RPEMIN
Definition: aacpsy.c:57
static const PsyLamePreset psy_abr_map[]
LAME psy model preset table for ABR.
Definition: aacpsy.c:176
int window_shape
window shape (sine/KBD/whatever)
Definition: psymodel.h:67
float min_snr
minimal SNR
Definition: aacpsy.c:142
float max
maximum allowed PE for bit factor calculation
Definition: aacpsy.c:154
float previous
allowed PE of the previous frame
Definition: aacpsy.c:155
AacPsyCoeffs psy_coef[2][64]
Definition: aacpsy.c:158
float min
minimum allowed PE for bit factor calculation
Definition: aacpsy.c:153
int global_quality
Global quality for codecs which cannot change it per frame.
Definition: avcodec.h:1321
static av_cold int psy_3gpp_init(FFPsyContext *ctx)
Definition: aacpsy.c:297
static uint32_t state
Definition: trasher.c:27
static void psy_hp_filter(const float *firbuf, float *hpfsmpl, const float *psy_fir_coeffs)
Definition: aacpsy.c:603
float spread_hi[2]
spreading factor for high-to-low threshold spreading in long frame
Definition: aacpsy.c:141
const char * name
Definition: psymodel.h:100
static av_unused FFPsyWindowInfo psy_3gpp_window(FFPsyContext *ctx, const int16_t *audio, const int16_t *la, int channel, int prev_type)
Tell encoder which window types to use.
Definition: aacpsy.c:394
static float calc_pe_3gpp(AacPsyBand *band)
Definition: aacpsy.c:513
#define exp2(x)
Definition: libm.h:77
windowing related information
Definition: psymodel.h:65
#define log2f(x)
Definition: libm.h:127
#define PSY_3GPP_BITS_TO_PE(bits)
Definition: aacpsy.c:89
#define PSY_3GPP_C1
Definition: aacpsy.c:60
float norm_fac
normalization factor for linearization
Definition: aacpsy.c:113
int chan_bitrate
bitrate per channel
Definition: aacpsy.c:149
int cutoff
Audio cutoff bandwidth (0 means "automatic")
Definition: avcodec.h:2029
#define PSY_3GPP_CLIP_LO_S
Definition: aacpsy.c:76
#define PSY_3GPP_AH_THR_LONG
Definition: aacpsy.c:80
#define NAN
Definition: math.h:28
int channels
number of audio channels
Definition: avcodec.h:1986
float pe
perceptual entropy
Definition: aacpsy.c:111
#define PSY_3GPP_EN_SPREAD_HI_S
Definition: aacpsy.c:51
static const double coeff[2][5]
Definition: vf_owdenoise.c:71
#define FF_QP2LAMBDA
factor to convert from H.263 QP to lambda
Definition: avutil.h:220
AacPsyChannel * ch
Definition: aacpsy.c:159
static void * av_mallocz_array(size_t nmemb, size_t size)
Definition: mem.h:228
#define PSY_3GPP_SAVE_ADD_S
Definition: aacpsy.c:70
#define av_freep(p)
void INT64 start
Definition: avisynth_c.h:553
information for single band used by 3GPP TS26.403-inspired psychoacoustic model
Definition: aacpsy.c:105
AVCodecContext * avctx
encoder context
Definition: psymodel.h:77
float threshold
Definition: psymodel.h:40
AAC data declarations.
float spread_low[2]
spreading factor for low-to-high threshold spreading in long frame
Definition: aacpsy.c:140
#define PSY_3GPP_CLIP_HI_L
Definition: aacpsy.c:77
int window_type[3]
window type (short/long/transitional, etc.) - current, previous and next
Definition: psymodel.h:66
void * av_mallocz(size_t size)
Allocate a block of size bytes with alignment suitable for all memory accesses (including vectors if ...
Definition: mem.c:250
#define AAC_NUM_BLOCKS_SHORT
number of blocks in a short sequence
Definition: aacpsy.c:95
#define av_unused
Definition: attributes.h:118
#define PSY_3GPP_SPEND_ADD_L
Definition: aacpsy.c:73
static av_cold void lame_window_init(AacPsyContext *ctx, AVCodecContext *avctx)
LAME psy model specific initialization.
Definition: aacpsy.c:258