FFmpeg
aaccoder_twoloop.h
Go to the documentation of this file.
1 /*
2  * AAC encoder twoloop coder
3  * Copyright (C) 2008-2009 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 twoloop coder
25  * @author Konstantin Shishkov, Claudio Freire
26  */
27 
28 /**
29  * This file contains a template for the twoloop coder function.
30  * It needs to be provided, externally, as an already included declaration,
31  * the following functions from aacenc_quantization/util.h. They're not included
32  * explicitly here to make it possible to provide alternative implementations:
33  * - quantize_band_cost
34  * - abs_pow34_v
35  * - find_max_val
36  * - find_min_book
37  * - find_form_factor
38  */
39 
40 #ifndef AVCODEC_AACCODER_TWOLOOP_H
41 #define AVCODEC_AACCODER_TWOLOOP_H
42 
43 #include <float.h>
44 #include "libavutil/mathematics.h"
45 #include "mathops.h"
46 #include "avcodec.h"
47 #include "put_bits.h"
48 #include "aac.h"
49 #include "aacenc.h"
50 #include "aactab.h"
51 #include "aacenctab.h"
52 
53 /** Frequency in Hz for lower limit of noise substitution **/
54 #define NOISE_LOW_LIMIT 4000
55 
56 #define sclip(x) av_clip(x,60,218)
57 
58 /* Reflects the cost to change codebooks */
59 static inline int ff_pns_bits(SingleChannelElement *sce, int w, int g)
60 {
61  return (!g || !sce->zeroes[w*16+g-1] || !sce->can_pns[w*16+g-1]) ? 9 : 5;
62 }
63 
64 /**
65  * two-loop quantizers search taken from ISO 13818-7 Appendix C
66  */
70  const float lambda)
71 {
72  int start = 0, i, w, w2, g, recomprd;
73  int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
74  / ((avctx->flags & AV_CODEC_FLAG_QSCALE) ? 2.0f : avctx->ch_layout.nb_channels)
75  * (lambda / 120.f);
76  int refbits = destbits;
77  int toomanybits, toofewbits;
78  char nzs[128];
79  uint8_t nextband[128];
80  int maxsf[128], minsf[128];
81  float dists[128] = { 0 }, qenergies[128] = { 0 }, uplims[128], euplims[128], energies[128];
82  float maxvals[128], spread_thr_r[128];
83  float min_spread_thr_r, max_spread_thr_r;
84 
85  /**
86  * rdlambda controls the maximum tolerated distortion. Twoloop
87  * will keep iterating until it fails to lower it or it reaches
88  * ulimit * rdlambda. Keeping it low increases quality on difficult
89  * signals, but lower it too much, and bits will be taken from weak
90  * signals, creating "holes". A balance is necessary.
91  * rdmax and rdmin specify the relative deviation from rdlambda
92  * allowed for tonality compensation
93  */
94  float rdlambda = av_clipf(2.0f * 120.f / lambda, 0.0625f, 16.0f);
95  const float nzslope = 1.5f;
96  float rdmin = 0.03125f;
97  float rdmax = 1.0f;
98 
99  /**
100  * sfoffs controls an offset of optmium allocation that will be
101  * applied based on lambda. Keep it real and modest, the loop
102  * will take care of the rest, this just accelerates convergence
103  */
104  float sfoffs = av_clipf(log2f(120.0f / lambda) * 4.0f, -5, 10);
105 
106  int fflag, minscaler, maxscaler, nminscaler;
107  int its = 0;
108  int maxits = 30;
109  int allz = 0;
110  int tbits;
111  int cutoff = 1024;
112  int pns_start_pos;
113  int prev;
114 
115  /**
116  * zeroscale controls a multiplier of the threshold, if band energy
117  * is below this, a zero is forced. Keep it lower than 1, unless
118  * low lambda is used, because energy < threshold doesn't mean there's
119  * no audible signal outright, it's just energy. Also make it rise
120  * slower than rdlambda, as rdscale has due compensation with
121  * noisy band depriorization below, whereas zeroing logic is rather dumb
122  */
123  float zeroscale;
124  if (lambda > 120.f) {
125  zeroscale = av_clipf(powf(120.f / lambda, 0.25f), 0.0625f, 1.0f);
126  } else {
127  zeroscale = 1.f;
128  }
129 
130  if (s->psy.bitres.alloc >= 0) {
131  /**
132  * Psy granted us extra bits to use, from the reservoire
133  * adjust for lambda except what psy already did
134  */
135  destbits = s->psy.bitres.alloc
136  * (lambda / (avctx->global_quality ? avctx->global_quality : 120));
137  }
138 
139  if (avctx->flags & AV_CODEC_FLAG_QSCALE) {
140  /**
141  * Constant Q-scale doesn't compensate MS coding on its own
142  * No need to be overly precise, this only controls RD
143  * adjustment CB limits when going overboard
144  */
145  if (s->options.mid_side && s->cur_type == TYPE_CPE)
146  destbits *= 2;
147 
148  /**
149  * When using a constant Q-scale, don't adjust bits, just use RD
150  * Don't let it go overboard, though... 8x psy target is enough
151  */
152  toomanybits = 5800;
153  toofewbits = destbits / 16;
154 
155  /** Don't offset scalers, just RD */
156  sfoffs = sce->ics.num_windows - 1;
157  rdlambda = sqrtf(rdlambda);
158 
159  /** search further */
160  maxits *= 2;
161  } else {
162  /* When using ABR, be strict, but a reasonable leeway is
163  * critical to allow RC to smoothly track desired bitrate
164  * without sudden quality drops that cause audible artifacts.
165  * Symmetry is also desirable, to avoid systematic bias.
166  */
167  toomanybits = destbits + destbits/8;
168  toofewbits = destbits - destbits/8;
169 
170  sfoffs = 0;
171  rdlambda = sqrtf(rdlambda);
172  }
173 
174  /** and zero out above cutoff frequency */
175  {
176  int wlen = 1024 / sce->ics.num_windows;
177  int bandwidth;
178 
179  /**
180  * Scale, psy gives us constant quality, this LP only scales
181  * bitrate by lambda, so we save bits on subjectively unimportant HF
182  * rather than increase quantization noise. Adjust nominal bitrate
183  * to effective bitrate according to encoding parameters,
184  * AAC_CUTOFF_FROM_BITRATE is calibrated for effective bitrate.
185  */
186  float rate_bandwidth_multiplier = 1.5f;
187  int frame_bit_rate = (avctx->flags & AV_CODEC_FLAG_QSCALE)
188  ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
189  : (avctx->bit_rate / avctx->ch_layout.nb_channels);
190 
191  /** Compensate for extensions that increase efficiency */
192  if (s->options.pns || s->options.intensity_stereo)
193  frame_bit_rate *= 1.15f;
194 
195  if (avctx->cutoff > 0) {
196  bandwidth = avctx->cutoff;
197  } else {
198  bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
199  s->psy.cutoff = bandwidth;
200  }
201 
202  cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
203  pns_start_pos = NOISE_LOW_LIMIT * 2 * wlen / avctx->sample_rate;
204  }
205 
206  /**
207  * for values above this the decoder might end up in an endless loop
208  * due to always having more bits than what can be encoded.
209  */
210  destbits = FFMIN(destbits, 5800);
211  toomanybits = FFMIN(toomanybits, 5800);
212  toofewbits = FFMIN(toofewbits, 5800);
213  /**
214  * XXX: some heuristic to determine initial quantizers will reduce search time
215  * determine zero bands and upper distortion limits
216  */
217  min_spread_thr_r = -1;
218  max_spread_thr_r = -1;
219  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
220  for (g = start = 0; g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
221  int nz = 0;
222  float uplim = 0.0f, energy = 0.0f, spread = 0.0f;
223  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
224  FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
225  if (start >= cutoff || band->energy <= (band->threshold * zeroscale) || band->threshold == 0.0f) {
226  sce->zeroes[(w+w2)*16+g] = 1;
227  continue;
228  }
229  nz = 1;
230  }
231  if (!nz) {
232  uplim = 0.0f;
233  } else {
234  nz = 0;
235  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
236  FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
237  if (band->energy <= (band->threshold * zeroscale) || band->threshold == 0.0f)
238  continue;
239  uplim += band->threshold;
240  energy += band->energy;
241  spread += band->spread;
242  nz++;
243  }
244  }
245  uplims[w*16+g] = uplim;
246  energies[w*16+g] = energy;
247  nzs[w*16+g] = nz;
248  sce->zeroes[w*16+g] = !nz;
249  allz |= nz;
250  if (nz && sce->can_pns[w*16+g]) {
251  spread_thr_r[w*16+g] = energy * nz / (uplim * spread);
252  if (min_spread_thr_r < 0) {
253  min_spread_thr_r = max_spread_thr_r = spread_thr_r[w*16+g];
254  } else {
255  min_spread_thr_r = FFMIN(min_spread_thr_r, spread_thr_r[w*16+g]);
256  max_spread_thr_r = FFMAX(max_spread_thr_r, spread_thr_r[w*16+g]);
257  }
258  }
259  }
260  }
261 
262  /** Compute initial scalers */
263  minscaler = 65535;
264  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
265  for (g = 0; g < sce->ics.num_swb; g++) {
266  if (sce->zeroes[w*16+g]) {
267  sce->sf_idx[w*16+g] = SCALE_ONE_POS;
268  continue;
269  }
270  /**
271  * log2f-to-distortion ratio is, technically, 2 (1.5db = 4, but it's power vs level so it's 2).
272  * But, as offsets are applied, low-frequency signals are too sensitive to the induced distortion,
273  * so we make scaling more conservative by choosing a lower log2f-to-distortion ratio, and thus
274  * more robust.
275  */
276  sce->sf_idx[w*16+g] = av_clip(
278  + 1.75*log2f(FFMAX(0.00125f,uplims[w*16+g]) / sce->ics.swb_sizes[g])
279  + sfoffs,
280  60, SCALE_MAX_POS);
281  minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
282  }
283  }
284 
285  /** Clip */
286  minscaler = av_clip(minscaler, SCALE_ONE_POS - SCALE_DIV_512, SCALE_MAX_POS - SCALE_DIV_512);
287  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
288  for (g = 0; g < sce->ics.num_swb; g++)
289  if (!sce->zeroes[w*16+g])
290  sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF - 1);
291 
292  if (!allz)
293  return;
294  s->aacdsp.abs_pow34(s->scoefs, sce->coeffs, 1024);
296 
297  for (i = 0; i < sizeof(minsf) / sizeof(minsf[0]); ++i)
298  minsf[i] = 0;
299  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
300  start = w*128;
301  for (g = 0; g < sce->ics.num_swb; g++) {
302  const float *scaled = s->scoefs + start;
303  int minsfidx;
304  maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
305  if (maxvals[w*16+g] > 0) {
306  minsfidx = coef2minsf(maxvals[w*16+g]);
307  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
308  minsf[(w+w2)*16+g] = minsfidx;
309  }
310  start += sce->ics.swb_sizes[g];
311  }
312  }
313 
314  /**
315  * Scale uplims to match rate distortion to quality
316  * bu applying noisy band depriorization and tonal band priorization.
317  * Maxval-energy ratio gives us an idea of how noisy/tonal the band is.
318  * If maxval^2 ~ energy, then that band is mostly noise, and we can relax
319  * rate distortion requirements.
320  */
321  memcpy(euplims, uplims, sizeof(euplims));
322  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
323  /** psy already priorizes transients to some extent */
324  float de_psy_factor = (sce->ics.num_windows > 1) ? 8.0f / sce->ics.group_len[w] : 1.0f;
325  start = w*128;
326  for (g = 0; g < sce->ics.num_swb; g++) {
327  if (nzs[g] > 0) {
328  float cleanup_factor = ff_sqrf(av_clipf(start / (cutoff * 0.75f), 1.0f, 2.0f));
329  float energy2uplim = find_form_factor(
330  sce->ics.group_len[w], sce->ics.swb_sizes[g],
331  uplims[w*16+g] / (nzs[g] * sce->ics.swb_sizes[w]),
332  sce->coeffs + start,
333  nzslope * cleanup_factor);
334  energy2uplim *= de_psy_factor;
335  if (!(avctx->flags & AV_CODEC_FLAG_QSCALE)) {
336  /** In ABR, we need to priorize less and let rate control do its thing */
337  energy2uplim = sqrtf(energy2uplim);
338  }
339  energy2uplim = FFMAX(0.015625f, FFMIN(1.0f, energy2uplim));
340  uplims[w*16+g] *= av_clipf(rdlambda * energy2uplim, rdmin, rdmax)
341  * sce->ics.group_len[w];
342 
343  energy2uplim = find_form_factor(
344  sce->ics.group_len[w], sce->ics.swb_sizes[g],
345  uplims[w*16+g] / (nzs[g] * sce->ics.swb_sizes[w]),
346  sce->coeffs + start,
347  2.0f);
348  energy2uplim *= de_psy_factor;
349  if (!(avctx->flags & AV_CODEC_FLAG_QSCALE)) {
350  /** In ABR, we need to priorize less and let rate control do its thing */
351  energy2uplim = sqrtf(energy2uplim);
352  }
353  energy2uplim = FFMAX(0.015625f, FFMIN(1.0f, energy2uplim));
354  euplims[w*16+g] *= av_clipf(rdlambda * energy2uplim * sce->ics.group_len[w],
355  0.5f, 1.0f);
356  }
357  start += sce->ics.swb_sizes[g];
358  }
359  }
360 
361  for (i = 0; i < sizeof(maxsf) / sizeof(maxsf[0]); ++i)
362  maxsf[i] = SCALE_MAX_POS;
363 
364  //perform two-loop search
365  //outer loop - improve quality
366  do {
367  //inner loop - quantize spectrum to fit into given number of bits
368  int overdist;
369  int qstep = its ? 1 : 32;
370  do {
371  int changed = 0;
372  prev = -1;
373  recomprd = 0;
374  tbits = 0;
375  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
376  start = w*128;
377  for (g = 0; g < sce->ics.num_swb; g++) {
378  const float *coefs = &sce->coeffs[start];
379  const float *scaled = &s->scoefs[start];
380  int bits = 0;
381  int cb;
382  float dist = 0.0f;
383  float qenergy = 0.0f;
384 
385  if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
386  start += sce->ics.swb_sizes[g];
387  if (sce->can_pns[w*16+g]) {
388  /** PNS isn't free */
389  tbits += ff_pns_bits(sce, w, g);
390  }
391  continue;
392  }
393  cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
394  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
395  int b;
396  float sqenergy;
397  dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
398  scaled + w2*128,
399  sce->ics.swb_sizes[g],
400  sce->sf_idx[w*16+g],
401  cb,
402  1.0f,
403  INFINITY,
404  &b, &sqenergy,
405  0);
406  bits += b;
407  qenergy += sqenergy;
408  }
409  dists[w*16+g] = dist - bits;
410  qenergies[w*16+g] = qenergy;
411  if (prev != -1) {
412  int sfdiff = av_clip(sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO, 0, 2*SCALE_MAX_DIFF);
413  bits += ff_aac_scalefactor_bits[sfdiff];
414  }
415  tbits += bits;
416  start += sce->ics.swb_sizes[g];
417  prev = sce->sf_idx[w*16+g];
418  }
419  }
420  if (tbits > toomanybits) {
421  recomprd = 1;
422  for (i = 0; i < 128; i++) {
423  if (sce->sf_idx[i] < (SCALE_MAX_POS - SCALE_DIV_512)) {
424  int maxsf_i = (tbits > 5800) ? SCALE_MAX_POS : maxsf[i];
425  int new_sf = FFMIN(maxsf_i, sce->sf_idx[i] + qstep);
426  if (new_sf != sce->sf_idx[i]) {
427  sce->sf_idx[i] = new_sf;
428  changed = 1;
429  }
430  }
431  }
432  } else if (tbits < toofewbits) {
433  recomprd = 1;
434  for (i = 0; i < 128; i++) {
435  if (sce->sf_idx[i] > SCALE_ONE_POS) {
436  int new_sf = FFMAX3(minsf[i], SCALE_ONE_POS, sce->sf_idx[i] - qstep);
437  if (new_sf != sce->sf_idx[i]) {
438  sce->sf_idx[i] = new_sf;
439  changed = 1;
440  }
441  }
442  }
443  }
444  qstep >>= 1;
445  if (!qstep && tbits > toomanybits && sce->sf_idx[0] < 217 && changed)
446  qstep = 1;
447  } while (qstep);
448 
449  overdist = 1;
450  fflag = tbits < toofewbits;
451  for (i = 0; i < 2 && (overdist || recomprd); ++i) {
452  if (recomprd) {
453  /** Must recompute distortion */
454  prev = -1;
455  tbits = 0;
456  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
457  start = w*128;
458  for (g = 0; g < sce->ics.num_swb; g++) {
459  const float *coefs = sce->coeffs + start;
460  const float *scaled = s->scoefs + start;
461  int bits = 0;
462  int cb;
463  float dist = 0.0f;
464  float qenergy = 0.0f;
465 
466  if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
467  start += sce->ics.swb_sizes[g];
468  if (sce->can_pns[w*16+g]) {
469  /** PNS isn't free */
470  tbits += ff_pns_bits(sce, w, g);
471  }
472  continue;
473  }
474  cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
475  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
476  int b;
477  float sqenergy;
478  dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
479  scaled + w2*128,
480  sce->ics.swb_sizes[g],
481  sce->sf_idx[w*16+g],
482  cb,
483  1.0f,
484  INFINITY,
485  &b, &sqenergy,
486  0);
487  bits += b;
488  qenergy += sqenergy;
489  }
490  dists[w*16+g] = dist - bits;
491  qenergies[w*16+g] = qenergy;
492  if (prev != -1) {
493  int sfdiff = av_clip(sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO, 0, 2*SCALE_MAX_DIFF);
494  bits += ff_aac_scalefactor_bits[sfdiff];
495  }
496  tbits += bits;
497  start += sce->ics.swb_sizes[g];
498  prev = sce->sf_idx[w*16+g];
499  }
500  }
501  }
502  if (!i && s->options.pns && its > maxits/2 && tbits > toofewbits) {
503  float maxoverdist = 0.0f;
504  float ovrfactor = 1.f+(maxits-its)*16.f/maxits;
505  overdist = recomprd = 0;
506  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
507  for (g = start = 0; g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
508  if (!sce->zeroes[w*16+g] && sce->sf_idx[w*16+g] > SCALE_ONE_POS && dists[w*16+g] > uplims[w*16+g]*ovrfactor) {
509  float ovrdist = dists[w*16+g] / FFMAX(uplims[w*16+g],euplims[w*16+g]);
510  maxoverdist = FFMAX(maxoverdist, ovrdist);
511  overdist++;
512  }
513  }
514  }
515  if (overdist) {
516  /* We have overdistorted bands, trade for zeroes (that can be noise)
517  * Zero the bands in the lowest 1.25% spread-energy-threshold ranking
518  */
519  float minspread = max_spread_thr_r;
520  float maxspread = min_spread_thr_r;
521  float zspread;
522  int zeroable = 0;
523  int zeroed = 0;
524  int maxzeroed, zloop;
525  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
526  for (g = start = 0; g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
527  if (start >= pns_start_pos && !sce->zeroes[w*16+g] && sce->can_pns[w*16+g]) {
528  minspread = FFMIN(minspread, spread_thr_r[w*16+g]);
529  maxspread = FFMAX(maxspread, spread_thr_r[w*16+g]);
530  zeroable++;
531  }
532  }
533  }
534  zspread = (maxspread-minspread) * 0.0125f + minspread;
535  /* Don't PNS everything even if allowed. It suppresses bit starvation signals from RC,
536  * and forced the hand of the later search_for_pns step.
537  * Instead, PNS a fraction of the spread_thr_r range depending on how starved for bits we are,
538  * and leave further PNSing to search_for_pns if worthwhile.
539  */
540  zspread = FFMIN3(min_spread_thr_r * 8.f, zspread,
541  ((toomanybits - tbits) * min_spread_thr_r + (tbits - toofewbits) * max_spread_thr_r) / (toomanybits - toofewbits + 1));
542  maxzeroed = FFMIN(zeroable, FFMAX(1, (zeroable * its + maxits - 1) / (2 * maxits)));
543  for (zloop = 0; zloop < 2; zloop++) {
544  /* Two passes: first distorted stuff - two birds in one shot and all that,
545  * then anything viable. Viable means not zero, but either CB=zero-able
546  * (too high SF), not SF <= 1 (that means we'd be operating at very high
547  * quality, we don't want PNS when doing VHQ), PNS allowed, and within
548  * the lowest ranking percentile.
549  */
550  float loopovrfactor = (zloop) ? 1.0f : ovrfactor;
551  int loopminsf = (zloop) ? (SCALE_ONE_POS - SCALE_DIV_512) : SCALE_ONE_POS;
552  int mcb;
553  for (g = sce->ics.num_swb-1; g > 0 && zeroed < maxzeroed; g--) {
554  if (sce->ics.swb_offset[g] < pns_start_pos)
555  continue;
556  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
557  if (!sce->zeroes[w*16+g] && sce->can_pns[w*16+g] && spread_thr_r[w*16+g] <= zspread
558  && sce->sf_idx[w*16+g] > loopminsf
559  && (dists[w*16+g] > loopovrfactor*uplims[w*16+g] || !(mcb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]))
560  || (mcb <= 1 && dists[w*16+g] > FFMIN(uplims[w*16+g], euplims[w*16+g]))) ) {
561  sce->zeroes[w*16+g] = 1;
562  sce->band_type[w*16+g] = 0;
563  zeroed++;
564  }
565  }
566  }
567  }
568  if (zeroed)
569  recomprd = fflag = 1;
570  } else {
571  overdist = 0;
572  }
573  }
574  }
575 
576  minscaler = SCALE_MAX_POS;
577  maxscaler = 0;
578  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
579  for (g = 0; g < sce->ics.num_swb; g++) {
580  if (!sce->zeroes[w*16+g]) {
581  minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
582  maxscaler = FFMAX(maxscaler, sce->sf_idx[w*16+g]);
583  }
584  }
585  }
586 
587  minscaler = nminscaler = av_clip(minscaler, SCALE_ONE_POS - SCALE_DIV_512, SCALE_MAX_POS - SCALE_DIV_512);
588  prev = -1;
589  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
590  /** Start with big steps, end up fine-tunning */
591  int depth = (its > maxits/2) ? ((its > maxits*2/3) ? 1 : 3) : 10;
592  int edepth = depth+2;
593  float uplmax = its / (maxits*0.25f) + 1.0f;
594  uplmax *= (tbits > destbits) ? FFMIN(2.0f, tbits / (float)FFMAX(1,destbits)) : 1.0f;
595  start = w * 128;
596  for (g = 0; g < sce->ics.num_swb; g++) {
597  int prevsc = sce->sf_idx[w*16+g];
598  if (prev < 0 && !sce->zeroes[w*16+g])
599  prev = sce->sf_idx[0];
600  if (!sce->zeroes[w*16+g]) {
601  const float *coefs = sce->coeffs + start;
602  const float *scaled = s->scoefs + start;
603  int cmb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
604  int mindeltasf = FFMAX(0, prev - SCALE_MAX_DIFF);
605  int maxdeltasf = FFMIN(SCALE_MAX_POS - SCALE_DIV_512, prev + SCALE_MAX_DIFF);
606  if ((!cmb || dists[w*16+g] > uplims[w*16+g]) && sce->sf_idx[w*16+g] > FFMAX(mindeltasf, minsf[w*16+g])) {
607  /* Try to make sure there is some energy in every nonzero band
608  * NOTE: This algorithm must be forcibly imbalanced, pushing harder
609  * on holes or more distorted bands at first, otherwise there's
610  * no net gain (since the next iteration will offset all bands
611  * on the opposite direction to compensate for extra bits)
612  */
613  for (i = 0; i < edepth && sce->sf_idx[w*16+g] > mindeltasf; ++i) {
614  int cb, bits;
615  float dist, qenergy;
616  int mb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1);
617  cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
618  dist = qenergy = 0.f;
619  bits = 0;
620  if (!cb) {
621  maxsf[w*16+g] = FFMIN(sce->sf_idx[w*16+g]-1, maxsf[w*16+g]);
622  } else if (i >= depth && dists[w*16+g] < euplims[w*16+g]) {
623  break;
624  }
625  /* !g is the DC band, it's important, since quantization error here
626  * applies to less than a cycle, it creates horrible intermodulation
627  * distortion if it doesn't stick to what psy requests
628  */
629  if (!g && sce->ics.num_windows > 1 && dists[w*16+g] >= euplims[w*16+g])
630  maxsf[w*16+g] = FFMIN(sce->sf_idx[w*16+g], maxsf[w*16+g]);
631  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
632  int b;
633  float sqenergy;
634  dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
635  scaled + w2*128,
636  sce->ics.swb_sizes[g],
637  sce->sf_idx[w*16+g]-1,
638  cb,
639  1.0f,
640  INFINITY,
641  &b, &sqenergy,
642  0);
643  bits += b;
644  qenergy += sqenergy;
645  }
646  sce->sf_idx[w*16+g]--;
647  dists[w*16+g] = dist - bits;
648  qenergies[w*16+g] = qenergy;
649  if (mb && (sce->sf_idx[w*16+g] < mindeltasf || (
650  (dists[w*16+g] < FFMIN(uplmax*uplims[w*16+g], euplims[w*16+g]))
651  && (fabsf(qenergies[w*16+g]-energies[w*16+g]) < euplims[w*16+g])
652  ) )) {
653  break;
654  }
655  }
656  } else if (tbits > toofewbits && sce->sf_idx[w*16+g] < FFMIN(maxdeltasf, maxsf[w*16+g])
657  && (dists[w*16+g] < FFMIN(euplims[w*16+g], uplims[w*16+g]))
658  && (fabsf(qenergies[w*16+g]-energies[w*16+g]) < euplims[w*16+g])
659  ) {
660  /** Um... over target. Save bits for more important stuff. */
661  for (i = 0; i < depth && sce->sf_idx[w*16+g] < maxdeltasf; ++i) {
662  int cb, bits;
663  float dist, qenergy;
664  cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]+1);
665  if (cb > 0) {
666  dist = qenergy = 0.f;
667  bits = 0;
668  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
669  int b;
670  float sqenergy;
671  dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
672  scaled + w2*128,
673  sce->ics.swb_sizes[g],
674  sce->sf_idx[w*16+g]+1,
675  cb,
676  1.0f,
677  INFINITY,
678  &b, &sqenergy,
679  0);
680  bits += b;
681  qenergy += sqenergy;
682  }
683  dist -= bits;
684  if (dist < FFMIN(euplims[w*16+g], uplims[w*16+g])) {
685  sce->sf_idx[w*16+g]++;
686  dists[w*16+g] = dist;
687  qenergies[w*16+g] = qenergy;
688  } else {
689  break;
690  }
691  } else {
692  maxsf[w*16+g] = FFMIN(sce->sf_idx[w*16+g], maxsf[w*16+g]);
693  break;
694  }
695  }
696  }
697  prev = sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], mindeltasf, maxdeltasf);
698  if (sce->sf_idx[w*16+g] != prevsc)
699  fflag = 1;
700  nminscaler = FFMIN(nminscaler, sce->sf_idx[w*16+g]);
701  sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
702  }
703  start += sce->ics.swb_sizes[g];
704  }
705  }
706 
707  /** SF difference limit violation risk. Must re-clamp. */
708  prev = -1;
709  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
710  for (g = 0; g < sce->ics.num_swb; g++) {
711  if (!sce->zeroes[w*16+g]) {
712  int prevsf = sce->sf_idx[w*16+g];
713  if (prev < 0)
714  prev = prevsf;
715  sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], prev - SCALE_MAX_DIFF, prev + SCALE_MAX_DIFF);
716  sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
717  prev = sce->sf_idx[w*16+g];
718  if (!fflag && prevsf != sce->sf_idx[w*16+g])
719  fflag = 1;
720  }
721  }
722  }
723 
724  its++;
725  } while (fflag && its < maxits);
726 
727  /** Scout out next nonzero bands */
728  ff_init_nextband_map(sce, nextband);
729 
730  prev = -1;
731  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
732  /** Make sure proper codebooks are set */
733  for (g = 0; g < sce->ics.num_swb; g++) {
734  if (!sce->zeroes[w*16+g]) {
735  sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
736  if (sce->band_type[w*16+g] <= 0) {
737  if (!ff_sfdelta_can_remove_band(sce, nextband, prev, w*16+g)) {
738  /** Cannot zero out, make sure it's not attempted */
739  sce->band_type[w*16+g] = 1;
740  } else {
741  sce->zeroes[w*16+g] = 1;
742  sce->band_type[w*16+g] = 0;
743  }
744  }
745  } else {
746  sce->band_type[w*16+g] = 0;
747  }
748  /** Check that there's no SF delta range violations */
749  if (!sce->zeroes[w*16+g]) {
750  if (prev != -1) {
751  av_unused int sfdiff = sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO;
752  av_assert1(sfdiff >= 0 && sfdiff <= 2*SCALE_MAX_DIFF);
753  } else if (sce->zeroes[0]) {
754  /** Set global gain to something useful */
755  sce->sf_idx[0] = sce->sf_idx[w*16+g];
756  }
757  prev = sce->sf_idx[w*16+g];
758  }
759  }
760  }
761 }
762 
763 #endif /* AVCODEC_AACCODER_TWOLOOP_H */
INFINITY
#define INFINITY
Definition: mathematics.h:118
av_clip
#define av_clip
Definition: common.h:98
SingleChannelElement::can_pns
uint8_t can_pns[128]
band is allowed to PNS (informative)
Definition: aacenc.h:121
AVCodecContext::sample_rate
int sample_rate
samples per second
Definition: avcodec.h:1050
cb
static double cb(void *priv, double x, double y)
Definition: vf_geq.c:241
aacenctab.h
log2f
#define log2f(x)
Definition: libm.h:409
AV_CODEC_FLAG_QSCALE
#define AV_CODEC_FLAG_QSCALE
Use fixed qscale.
Definition: avcodec.h:224
SingleChannelElement::zeroes
uint8_t zeroes[128]
band is not coded
Definition: aacenc.h:120
av_unused
#define av_unused
Definition: attributes.h:131
w
uint8_t w
Definition: llviddspenc.c:38
b
#define b
Definition: input.c:41
float.h
mathematics.h
ff_sfdelta_can_remove_band
static int ff_sfdelta_can_remove_band(const SingleChannelElement *sce, const uint8_t *nextband, int prev_sf, int band)
Definition: aacenc_utils.h:208
FFMAX
#define FFMAX(a, b)
Definition: macros.h:47
AVChannelLayout::nb_channels
int nb_channels
Number of channels in this layout.
Definition: channel_layout.h:313
SCALE_MAX_POS
#define SCALE_MAX_POS
scalefactor index maximum value
Definition: aac.h:106
IndividualChannelStream::num_swb
int num_swb
number of scalefactor window bands
Definition: aacdec.h:92
AVCodecContext::ch_layout
AVChannelLayout ch_layout
Audio channel layout.
Definition: avcodec.h:1065
SCALE_DIV_512
#define SCALE_DIV_512
scalefactor difference that corresponds to scale difference in 512 times
Definition: aac.h:104
TYPE_CPE
@ TYPE_CPE
Definition: aac.h:44
find_form_factor
static float find_form_factor(int group_len, int swb_size, float thresh, const float *scaled, float nzslope)
Definition: aacenc_utils.h:80
AVCodecContext::flags
int flags
AV_CODEC_FLAG_*.
Definition: avcodec.h:502
fabsf
static __device__ float fabsf(float a)
Definition: cuda_runtime.h:181
SingleChannelElement::ics
IndividualChannelStream ics
Definition: aacdec.h:132
s
#define s(width, name)
Definition: cbs_vp9.c:198
SingleChannelElement::coeffs
INTFLOAT coeffs[1024]
coefficients for IMDCT, maybe processed
Definition: aacdec.h:137
AVCodecContext::global_quality
int global_quality
Global quality for codecs which cannot change it per frame.
Definition: avcodec.h:1239
IndividualChannelStream::swb_sizes
const uint8_t * swb_sizes
table of scalefactor band sizes for a particular window
Definition: aacenc.h:84
g
const char * g
Definition: vf_curves.c:127
bits
uint8_t bits
Definition: vp3data.h:128
SCALE_DIFF_ZERO
#define SCALE_DIFF_ZERO
codebook index corresponding to zero scalefactor indices difference
Definition: aac.h:108
quantize_band_cost_cached
static float quantize_band_cost_cached(struct AACEncContext *s, int w, int g, const float *in, const float *scaled, int size, int scale_idx, int cb, const float lambda, const float uplim, int *bits, float *energy, int rtz)
Definition: aacenc_quantization_misc.h:31
AVCodecContext::bit_rate
int64_t bit_rate
the average bitrate
Definition: avcodec.h:495
mathops.h
FFPsyBand
single band psychoacoustic information
Definition: psymodel.h:50
aac.h
aactab.h
sqrtf
static __device__ float sqrtf(float a)
Definition: cuda_runtime.h:184
ff_init_nextband_map
static void ff_init_nextband_map(const SingleChannelElement *sce, uint8_t *nextband)
Definition: aacenc_utils.h:175
av_clipf
av_clipf
Definition: af_crystalizer.c:121
SingleChannelElement::sf_idx
int sf_idx[128]
scalefactor indices
Definition: aacenc.h:119
ff_aac_scalefactor_bits
const uint8_t ff_aac_scalefactor_bits[121]
Definition: aactab.c:196
coef2minsf
static uint8_t coef2minsf(float coef)
Return the minimum scalefactor where the quantized coef does not clip.
Definition: aacenc_utils.h:133
f
f
Definition: af_crystalizer.c:121
powf
#define powf(x, y)
Definition: libm.h:50
for
for(k=2;k<=8;++k)
Definition: h264pred_template.c:425
search_for_quantizers_twoloop
static void search_for_quantizers_twoloop(AVCodecContext *avctx, AACEncContext *s, SingleChannelElement *sce, const float lambda)
two-loop quantizers search taken from ISO 13818-7 Appendix C
Definition: aaccoder_twoloop.h:67
SingleChannelElement::band_type
enum BandType band_type[128]
band types
Definition: aacdec.h:134
mb
#define mb
Definition: vf_colormatrix.c:99
SCALE_MAX_DIFF
#define SCALE_MAX_DIFF
maximum scalefactor difference allowed by standard
Definition: aac.h:107
AAC_CUTOFF_FROM_BITRATE
#define AAC_CUTOFF_FROM_BITRATE(bit_rate, channels, sample_rate)
Definition: psymodel.h:35
SingleChannelElement
Single Channel Element - used for both SCE and LFE elements.
Definition: aacdec.h:131
i
#define i(width, name, range_min, range_max)
Definition: cbs_h2645.c:255
IndividualChannelStream::num_windows
int num_windows
Definition: aacdec.h:93
SCALE_ONE_POS
#define SCALE_ONE_POS
scalefactor index that corresponds to scale=1.0
Definition: aac.h:105
find_min_book
static int find_min_book(float maxval, int sf)
Definition: aacenc_utils.h:68
FFMIN3
#define FFMIN3(a, b, c)
Definition: macros.h:50
FFPsyBand::threshold
float threshold
Definition: psymodel.h:53
IndividualChannelStream::swb_offset
const uint16_t * swb_offset
table of offsets to the lowest spectral coefficient of a scalefactor band, sfb, for a particular wind...
Definition: aacdec.h:91
AVCodecContext::cutoff
int cutoff
Audio cutoff bandwidth (0 means "automatic")
Definition: avcodec.h:1090
av_assert1
#define av_assert1(cond)
assert() equivalent, that does not lie in speed critical code.
Definition: avassert.h:56
FFMIN
#define FFMIN(a, b)
Definition: macros.h:49
NOISE_LOW_LIMIT
#define NOISE_LOW_LIMIT
This file contains a template for the twoloop coder function.
Definition: aaccoder_twoloop.h:54
ff_sqrf
static av_const float ff_sqrf(float a)
Definition: mathops.h:240
avcodec.h
AACEncContext
AAC encoder context.
Definition: aacenc.h:198
FFPsyBand::energy
float energy
Definition: psymodel.h:52
AVCodecContext
main external API structure.
Definition: avcodec.h:445
find_max_val
static float find_max_val(int group_len, int swb_size, const float *scaled)
Definition: aacenc_utils.h:56
FFMAX3
#define FFMAX3(a, b, c)
Definition: macros.h:48
FFPsyBand::spread
float spread
Definition: psymodel.h:54
put_bits.h
IndividualChannelStream::group_len
uint8_t group_len[8]
Definition: aacdec.h:89
ff_pns_bits
static int ff_pns_bits(SingleChannelElement *sce, int w, int g)
Definition: aaccoder_twoloop.h:59
ff_quantize_band_cost_cache_init
void ff_quantize_band_cost_cache_init(struct AACEncContext *s)
Definition: aacenc.c:400
aacenc.h