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 /* Reflects the cost to change codebooks */
57 static inline int ff_pns_bits(SingleChannelElement *sce, int w, int g)
58 {
59  return (!g || !sce->zeroes[w*16+g-1] || !sce->can_pns[w*16+g-1]) ? 9 : 5;
60 }
61 
62 /**
63  * two-loop quantizers search taken from ISO 13818-7 Appendix C
64  */
65 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
66  AACEncContext *s,
67  SingleChannelElement *sce,
68  const float lambda)
69 {
70  int start = 0, i, w, w2, g, recomprd;
71  int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
72  / ((avctx->flags & AV_CODEC_FLAG_QSCALE) ? 2.0f : avctx->ch_layout.nb_channels)
73  * (lambda / 120.f);
74  int refbits = destbits;
75  int toomanybits, toofewbits;
76  char nzs[128];
77  uint8_t nextband[128];
78  int maxsf[128], minsf[128];
79  float dists[128] = { 0 }, qenergies[128] = { 0 }, uplims[128], euplims[128], energies[128];
80  float maxvals[128], spread_thr_r[128];
81  float min_spread_thr_r, max_spread_thr_r;
82 
83  /**
84  * rdlambda controls the maximum tolerated distortion. Twoloop
85  * will keep iterating until it fails to lower it or it reaches
86  * ulimit * rdlambda. Keeping it low increases quality on difficult
87  * signals, but lower it too much, and bits will be taken from weak
88  * signals, creating "holes". A balance is necessary.
89  * rdmax and rdmin specify the relative deviation from rdlambda
90  * allowed for tonality compensation
91  */
92  float rdlambda = av_clipf(2.0f * 120.f / lambda, 0.0625f, 16.0f);
93  const float nzslope = 1.5f;
94  float rdmin = 0.03125f;
95  float rdmax = 1.0f;
96 
97  /**
98  * sfoffs controls an offset of optmium allocation that will be
99  * applied based on lambda. Keep it real and modest, the loop
100  * will take care of the rest, this just accelerates convergence
101  */
102  float sfoffs = av_clipf(log2f(120.0f / lambda) * 4.0f, -5, 10);
103 
104  int fflag, minscaler, nminscaler;
105  int its = 0;
106  int maxits = 30;
107  int allz = 0;
108  int tbits;
109  int cutoff = 1024;
110  int pns_start_pos;
111  int prev;
112 
113  /**
114  * zeroscale controls a multiplier of the threshold, if band energy
115  * is below this, a zero is forced. Keep it lower than 1, unless
116  * low lambda is used, because energy < threshold doesn't mean there's
117  * no audible signal outright, it's just energy. Also make it rise
118  * slower than rdlambda, as rdscale has due compensation with
119  * noisy band depriorization below, whereas zeroing logic is rather dumb
120  */
121  float zeroscale;
122  if (lambda > 120.f) {
123  zeroscale = av_clipf(powf(120.f / lambda, 0.25f), 0.0625f, 1.0f);
124  } else {
125  zeroscale = 1.f;
126  }
127 
128  if (s->psy.bitres.alloc >= 0) {
129  /**
130  * Psy granted us extra bits to use, from the reservoire
131  * adjust for lambda except what psy already did
132  */
133  destbits = s->psy.bitres.alloc
134  * (lambda / (avctx->global_quality ? avctx->global_quality : 120));
135  }
136 
137  if (avctx->flags & AV_CODEC_FLAG_QSCALE) {
138  /**
139  * Constant Q-scale doesn't compensate MS coding on its own
140  * No need to be overly precise, this only controls RD
141  * adjustment CB limits when going overboard
142  */
143  if (s->options.mid_side && s->cur_type == TYPE_CPE)
144  destbits *= 2;
145 
146  /**
147  * When using a constant Q-scale, don't adjust bits, just use RD
148  * Don't let it go overboard, though... 8x psy target is enough
149  */
150  toomanybits = 5800;
151  toofewbits = destbits / 16;
152 
153  /** Don't offset scalers, just RD */
154  sfoffs = sce->ics.num_windows - 1;
155  rdlambda = sqrtf(rdlambda);
156 
157  /** search further */
158  maxits *= 2;
159  } else {
160  /* When using ABR, be strict, but a reasonable leeway is
161  * critical to allow RC to smoothly track desired bitrate
162  * without sudden quality drops that cause audible artifacts.
163  * Symmetry is also desirable, to avoid systematic bias.
164  */
165  toomanybits = destbits + destbits/8;
166  toofewbits = destbits - destbits/8;
167 
168  sfoffs = 0;
169  rdlambda = sqrtf(rdlambda);
170  }
171 
172  /** and zero out above cutoff frequency */
173  {
174  int wlen = 1024 / sce->ics.num_windows;
175  int bandwidth;
176 
177  /**
178  * Scale, psy gives us constant quality, this LP only scales
179  * bitrate by lambda, so we save bits on subjectively unimportant HF
180  * rather than increase quantization noise. Adjust nominal bitrate
181  * to effective bitrate according to encoding parameters,
182  * AAC_CUTOFF_FROM_BITRATE is calibrated for effective bitrate.
183  */
184  float rate_bandwidth_multiplier = 1.5f;
185  int frame_bit_rate = (avctx->flags & AV_CODEC_FLAG_QSCALE)
186  ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
187  : (avctx->bit_rate / avctx->ch_layout.nb_channels);
188 
189  /** Compensate for extensions that increase efficiency */
190  if (s->options.pns || s->options.intensity_stereo)
191  frame_bit_rate *= 1.15f;
192 
193  if (avctx->cutoff > 0) {
194  bandwidth = avctx->cutoff;
195  } else {
196  bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
197  s->psy.cutoff = bandwidth;
198  }
199 
200  cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
201  pns_start_pos = NOISE_LOW_LIMIT * 2 * wlen / avctx->sample_rate;
202  }
203 
204  /**
205  * for values above this the decoder might end up in an endless loop
206  * due to always having more bits than what can be encoded.
207  */
208  destbits = FFMIN(destbits, 5800);
209  toomanybits = FFMIN(toomanybits, 5800);
210  toofewbits = FFMIN(toofewbits, 5800);
211  /**
212  * XXX: some heuristic to determine initial quantizers will reduce search time
213  * determine zero bands and upper distortion limits
214  */
215  min_spread_thr_r = -1;
216  max_spread_thr_r = -1;
217  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
218  for (g = start = 0; g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
219  int nz = 0;
220  float uplim = 0.0f, energy = 0.0f, spread = 0.0f;
221  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
222  FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
223  if (start >= cutoff || band->energy <= (band->threshold * zeroscale) || band->threshold == 0.0f) {
224  sce->zeroes[(w+w2)*16+g] = 1;
225  continue;
226  }
227  nz = 1;
228  }
229  if (!nz) {
230  uplim = 0.0f;
231  } else {
232  nz = 0;
233  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
234  FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
235  if (band->energy <= (band->threshold * zeroscale) || band->threshold == 0.0f)
236  continue;
237  uplim += band->threshold;
238  energy += band->energy;
239  spread += band->spread;
240  nz++;
241  }
242  }
243  uplims[w*16+g] = uplim;
244  energies[w*16+g] = energy;
245  nzs[w*16+g] = nz;
246  sce->zeroes[w*16+g] = !nz;
247  allz |= nz;
248  if (nz && sce->can_pns[w*16+g]) {
249  spread_thr_r[w*16+g] = energy * nz / (uplim * spread);
250  if (min_spread_thr_r < 0) {
251  min_spread_thr_r = max_spread_thr_r = spread_thr_r[w*16+g];
252  } else {
253  min_spread_thr_r = FFMIN(min_spread_thr_r, spread_thr_r[w*16+g]);
254  max_spread_thr_r = FFMAX(max_spread_thr_r, spread_thr_r[w*16+g]);
255  }
256  }
257  }
258  }
259 
260  /** Compute initial scalers */
261  minscaler = 65535;
262  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
263  for (g = 0; g < sce->ics.num_swb; g++) {
264  if (sce->zeroes[w*16+g]) {
265  sce->sf_idx[w*16+g] = SCALE_ONE_POS;
266  continue;
267  }
268  /**
269  * log2f-to-distortion ratio is, technically, 2 (1.5db = 4, but it's power vs level so it's 2).
270  * But, as offsets are applied, low-frequency signals are too sensitive to the induced distortion,
271  * so we make scaling more conservative by choosing a lower log2f-to-distortion ratio, and thus
272  * more robust.
273  */
274  sce->sf_idx[w*16+g] = av_clip(
275  SCALE_ONE_POS
276  + 1.75*log2f(FFMAX(0.00125f,uplims[w*16+g]) / sce->ics.swb_sizes[g])
277  + sfoffs,
278  60, SCALE_MAX_POS);
279  minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
280  }
281  }
282 
283  /** Clip */
284  minscaler = av_clip(minscaler, SCALE_ONE_POS - SCALE_DIV_512, SCALE_MAX_POS - SCALE_DIV_512);
285  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
286  for (g = 0; g < sce->ics.num_swb; g++)
287  if (!sce->zeroes[w*16+g])
288  sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF - 1);
289 
290  if (!allz)
291  return;
292  s->aacdsp.abs_pow34(s->scoefs, sce->coeffs, 1024);
293  ff_quantize_band_cost_cache_init(s);
294 
295  for (i = 0; i < sizeof(minsf) / sizeof(minsf[0]); ++i)
296  minsf[i] = 0;
297  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
298  start = w*128;
299  for (g = 0; g < sce->ics.num_swb; g++) {
300  const float *scaled = s->scoefs + start;
301  int minsfidx;
302  maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
303  if (maxvals[w*16+g] > 0) {
304  minsfidx = coef2minsf(maxvals[w*16+g]);
305  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++)
306  minsf[(w+w2)*16+g] = minsfidx;
307  }
308  start += sce->ics.swb_sizes[g];
309  }
310  }
311 
312  /**
313  * Scale uplims to match rate distortion to quality
314  * bu applying noisy band depriorization and tonal band priorization.
315  * Maxval-energy ratio gives us an idea of how noisy/tonal the band is.
316  * If maxval^2 ~ energy, then that band is mostly noise, and we can relax
317  * rate distortion requirements.
318  */
319  memcpy(euplims, uplims, sizeof(euplims));
320  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
321  /** psy already priorizes transients to some extent */
322  float de_psy_factor = (sce->ics.num_windows > 1) ? 8.0f / sce->ics.group_len[w] : 1.0f;
323  start = w*128;
324  for (g = 0; g < sce->ics.num_swb; g++) {
325  if (nzs[g] > 0) {
326  float cleanup_factor = ff_sqrf(av_clipf(start / (cutoff * 0.75f), 1.0f, 2.0f));
327  float energy2uplim = find_form_factor(
328  sce->ics.group_len[w], sce->ics.swb_sizes[g],
329  uplims[w*16+g] / (nzs[g] * sce->ics.swb_sizes[w]),
330  sce->coeffs + start,
331  nzslope * cleanup_factor);
332  energy2uplim *= de_psy_factor;
333  if (!(avctx->flags & AV_CODEC_FLAG_QSCALE)) {
334  /** In ABR, we need to priorize less and let rate control do its thing */
335  energy2uplim = sqrtf(energy2uplim);
336  }
337  energy2uplim = FFMAX(0.015625f, FFMIN(1.0f, energy2uplim));
338  uplims[w*16+g] *= av_clipf(rdlambda * energy2uplim, rdmin, rdmax)
339  * sce->ics.group_len[w];
340 
341  energy2uplim = find_form_factor(
342  sce->ics.group_len[w], sce->ics.swb_sizes[g],
343  uplims[w*16+g] / (nzs[g] * sce->ics.swb_sizes[w]),
344  sce->coeffs + start,
345  2.0f);
346  energy2uplim *= de_psy_factor;
347  if (!(avctx->flags & AV_CODEC_FLAG_QSCALE)) {
348  /** In ABR, we need to priorize less and let rate control do its thing */
349  energy2uplim = sqrtf(energy2uplim);
350  }
351  energy2uplim = FFMAX(0.015625f, FFMIN(1.0f, energy2uplim));
352  euplims[w*16+g] *= av_clipf(rdlambda * energy2uplim * sce->ics.group_len[w],
353  0.5f, 1.0f);
354  }
355  start += sce->ics.swb_sizes[g];
356  }
357  }
358 
359  for (i = 0; i < sizeof(maxsf) / sizeof(maxsf[0]); ++i)
360  maxsf[i] = SCALE_MAX_POS;
361 
362  //perform two-loop search
363  //outer loop - improve quality
364  do {
365  //inner loop - quantize spectrum to fit into given number of bits
366  int overdist;
367  int qstep = its ? 1 : 32;
368  do {
369  int changed = 0;
370  prev = -1;
371  recomprd = 0;
372  tbits = 0;
373  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
374  start = w*128;
375  for (g = 0; g < sce->ics.num_swb; g++) {
376  const float *coefs = &sce->coeffs[start];
377  const float *scaled = &s->scoefs[start];
378  int bits = 0;
379  int cb;
380  float dist = 0.0f;
381  float qenergy = 0.0f;
382 
383  if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
384  start += sce->ics.swb_sizes[g];
385  if (sce->can_pns[w*16+g]) {
386  /** PNS isn't free */
387  tbits += ff_pns_bits(sce, w, g);
388  }
389  continue;
390  }
391  cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
392  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
393  int b;
394  float sqenergy;
395  dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
396  scaled + w2*128,
397  sce->ics.swb_sizes[g],
398  sce->sf_idx[w*16+g],
399  cb,
400  1.0f,
401  INFINITY,
402  &b, &sqenergy,
403  0);
404  bits += b;
405  qenergy += sqenergy;
406  }
407  dists[w*16+g] = dist - bits;
408  qenergies[w*16+g] = qenergy;
409  if (prev != -1) {
410  int sfdiff = av_clip(sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO, 0, 2*SCALE_MAX_DIFF);
411  bits += ff_aac_scalefactor_bits[sfdiff];
412  }
413  tbits += bits;
414  start += sce->ics.swb_sizes[g];
415  prev = sce->sf_idx[w*16+g];
416  }
417  }
418  if (tbits > toomanybits) {
419  recomprd = 1;
420  for (i = 0; i < 128; i++) {
421  if (sce->sf_idx[i] < (SCALE_MAX_POS - SCALE_DIV_512)) {
422  int maxsf_i = (tbits > 5800) ? SCALE_MAX_POS : maxsf[i];
423  int new_sf = FFMIN(maxsf_i, sce->sf_idx[i] + qstep);
424  if (new_sf != sce->sf_idx[i]) {
425  sce->sf_idx[i] = new_sf;
426  changed = 1;
427  }
428  }
429  }
430  } else if (tbits < toofewbits) {
431  recomprd = 1;
432  for (i = 0; i < 128; i++) {
433  if (sce->sf_idx[i] > SCALE_ONE_POS) {
434  int new_sf = FFMAX3(minsf[i], SCALE_ONE_POS, sce->sf_idx[i] - qstep);
435  if (new_sf != sce->sf_idx[i]) {
436  sce->sf_idx[i] = new_sf;
437  changed = 1;
438  }
439  }
440  }
441  }
442  qstep >>= 1;
443  if (!qstep && tbits > toomanybits && sce->sf_idx[0] < 217 && changed)
444  qstep = 1;
445  } while (qstep);
446 
447  overdist = 1;
448  fflag = tbits < toofewbits;
449  for (i = 0; i < 2 && (overdist || recomprd); ++i) {
450  if (recomprd) {
451  /** Must recompute distortion */
452  prev = -1;
453  tbits = 0;
454  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
455  start = w*128;
456  for (g = 0; g < sce->ics.num_swb; g++) {
457  const float *coefs = sce->coeffs + start;
458  const float *scaled = s->scoefs + start;
459  int bits = 0;
460  int cb;
461  float dist = 0.0f;
462  float qenergy = 0.0f;
463 
464  if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
465  start += sce->ics.swb_sizes[g];
466  if (sce->can_pns[w*16+g]) {
467  /** PNS isn't free */
468  tbits += ff_pns_bits(sce, w, g);
469  }
470  continue;
471  }
472  cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
473  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
474  int b;
475  float sqenergy;
476  dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
477  scaled + w2*128,
478  sce->ics.swb_sizes[g],
479  sce->sf_idx[w*16+g],
480  cb,
481  1.0f,
482  INFINITY,
483  &b, &sqenergy,
484  0);
485  bits += b;
486  qenergy += sqenergy;
487  }
488  dists[w*16+g] = dist - bits;
489  qenergies[w*16+g] = qenergy;
490  if (prev != -1) {
491  int sfdiff = av_clip(sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO, 0, 2*SCALE_MAX_DIFF);
492  bits += ff_aac_scalefactor_bits[sfdiff];
493  }
494  tbits += bits;
495  start += sce->ics.swb_sizes[g];
496  prev = sce->sf_idx[w*16+g];
497  }
498  }
499  }
500  if (!i && s->options.pns && its > maxits/2 && tbits > toofewbits) {
501  float maxoverdist = 0.0f;
502  float ovrfactor = 1.f+(maxits-its)*16.f/maxits;
503  overdist = recomprd = 0;
504  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
505  for (g = start = 0; g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
506  if (!sce->zeroes[w*16+g] && sce->sf_idx[w*16+g] > SCALE_ONE_POS && dists[w*16+g] > uplims[w*16+g]*ovrfactor) {
507  float ovrdist = dists[w*16+g] / FFMAX(uplims[w*16+g],euplims[w*16+g]);
508  maxoverdist = FFMAX(maxoverdist, ovrdist);
509  overdist++;
510  }
511  }
512  }
513  if (overdist) {
514  /* We have overdistorted bands, trade for zeroes (that can be noise)
515  * Zero the bands in the lowest 1.25% spread-energy-threshold ranking
516  */
517  float minspread = max_spread_thr_r;
518  float maxspread = min_spread_thr_r;
519  float zspread;
520  int zeroable = 0;
521  int zeroed = 0;
522  int maxzeroed, zloop;
523  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
524  for (g = start = 0; g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
525  if (start >= pns_start_pos && !sce->zeroes[w*16+g] && sce->can_pns[w*16+g]) {
526  minspread = FFMIN(minspread, spread_thr_r[w*16+g]);
527  maxspread = FFMAX(maxspread, spread_thr_r[w*16+g]);
528  zeroable++;
529  }
530  }
531  }
532  zspread = (maxspread-minspread) * 0.0125f + minspread;
533  /* Don't PNS everything even if allowed. It suppresses bit starvation signals from RC,
534  * and forced the hand of the later search_for_pns step.
535  * Instead, PNS a fraction of the spread_thr_r range depending on how starved for bits we are,
536  * and leave further PNSing to search_for_pns if worthwhile.
537  */
538  zspread = FFMIN3(min_spread_thr_r * 8.f, zspread,
539  ((toomanybits - tbits) * min_spread_thr_r + (tbits - toofewbits) * max_spread_thr_r) / (toomanybits - toofewbits + 1));
540  maxzeroed = FFMIN(zeroable, FFMAX(1, (zeroable * its + maxits - 1) / (2 * maxits)));
541  for (zloop = 0; zloop < 2; zloop++) {
542  /* Two passes: first distorted stuff - two birds in one shot and all that,
543  * then anything viable. Viable means not zero, but either CB=zero-able
544  * (too high SF), not SF <= 1 (that means we'd be operating at very high
545  * quality, we don't want PNS when doing VHQ), PNS allowed, and within
546  * the lowest ranking percentile.
547  */
548  float loopovrfactor = (zloop) ? 1.0f : ovrfactor;
549  int loopminsf = (zloop) ? (SCALE_ONE_POS - SCALE_DIV_512) : SCALE_ONE_POS;
550  int mcb;
551  for (g = sce->ics.num_swb-1; g > 0 && zeroed < maxzeroed; g--) {
552  if (sce->ics.swb_offset[g] < pns_start_pos)
553  continue;
554  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
555  if (!sce->zeroes[w*16+g] && sce->can_pns[w*16+g] && spread_thr_r[w*16+g] <= zspread
556  && sce->sf_idx[w*16+g] > loopminsf
557  && (dists[w*16+g] > loopovrfactor*uplims[w*16+g] || !(mcb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]))
558  || (mcb <= 1 && dists[w*16+g] > FFMIN(uplims[w*16+g], euplims[w*16+g]))) ) {
559  sce->zeroes[w*16+g] = 1;
560  sce->band_type[w*16+g] = 0;
561  zeroed++;
562  }
563  }
564  }
565  }
566  if (zeroed)
567  recomprd = fflag = 1;
568  } else {
569  overdist = 0;
570  }
571  }
572  }
573 
574  minscaler = SCALE_MAX_POS;
575  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
576  for (g = 0; g < sce->ics.num_swb; g++) {
577  if (!sce->zeroes[w*16+g]) {
578  minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
579  }
580  }
581  }
582 
583  minscaler = nminscaler = av_clip(minscaler, SCALE_ONE_POS - SCALE_DIV_512, SCALE_MAX_POS - SCALE_DIV_512);
584  prev = -1;
585  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
586  /** Start with big steps, end up fine-tunning */
587  int depth = (its > maxits/2) ? ((its > maxits*2/3) ? 1 : 3) : 10;
588  int edepth = depth+2;
589  float uplmax = its / (maxits*0.25f) + 1.0f;
590  uplmax *= (tbits > destbits) ? FFMIN(2.0f, tbits / (float)FFMAX(1,destbits)) : 1.0f;
591  start = w * 128;
592  for (g = 0; g < sce->ics.num_swb; g++) {
593  int prevsc = sce->sf_idx[w*16+g];
594  if (prev < 0 && !sce->zeroes[w*16+g])
595  prev = sce->sf_idx[0];
596  if (!sce->zeroes[w*16+g]) {
597  const float *coefs = sce->coeffs + start;
598  const float *scaled = s->scoefs + start;
599  int cmb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
600  int mindeltasf = FFMAX(0, prev - SCALE_MAX_DIFF);
601  int maxdeltasf = FFMIN(SCALE_MAX_POS - SCALE_DIV_512, prev + SCALE_MAX_DIFF);
602  if ((!cmb || dists[w*16+g] > uplims[w*16+g]) && sce->sf_idx[w*16+g] > FFMAX(mindeltasf, minsf[w*16+g])) {
603  /* Try to make sure there is some energy in every nonzero band
604  * NOTE: This algorithm must be forcibly imbalanced, pushing harder
605  * on holes or more distorted bands at first, otherwise there's
606  * no net gain (since the next iteration will offset all bands
607  * on the opposite direction to compensate for extra bits)
608  */
609  for (i = 0; i < edepth && sce->sf_idx[w*16+g] > mindeltasf; ++i) {
610  int cb, bits;
611  float dist, qenergy;
612  int mb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1);
613  cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
614  dist = qenergy = 0.f;
615  bits = 0;
616  if (!cb) {
617  maxsf[w*16+g] = FFMIN(sce->sf_idx[w*16+g]-1, maxsf[w*16+g]);
618  } else if (i >= depth && dists[w*16+g] < euplims[w*16+g]) {
619  break;
620  }
621  /* !g is the DC band, it's important, since quantization error here
622  * applies to less than a cycle, it creates horrible intermodulation
623  * distortion if it doesn't stick to what psy requests
624  */
625  if (!g && sce->ics.num_windows > 1 && dists[w*16+g] >= euplims[w*16+g])
626  maxsf[w*16+g] = FFMIN(sce->sf_idx[w*16+g], maxsf[w*16+g]);
627  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
628  int b;
629  float sqenergy;
630  dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
631  scaled + w2*128,
632  sce->ics.swb_sizes[g],
633  sce->sf_idx[w*16+g]-1,
634  cb,
635  1.0f,
636  INFINITY,
637  &b, &sqenergy,
638  0);
639  bits += b;
640  qenergy += sqenergy;
641  }
642  sce->sf_idx[w*16+g]--;
643  dists[w*16+g] = dist - bits;
644  qenergies[w*16+g] = qenergy;
645  if (mb && (sce->sf_idx[w*16+g] < mindeltasf || (
646  (dists[w*16+g] < FFMIN(uplmax*uplims[w*16+g], euplims[w*16+g]))
647  && (fabsf(qenergies[w*16+g]-energies[w*16+g]) < euplims[w*16+g])
648  ) )) {
649  break;
650  }
651  }
652  } else if (tbits > toofewbits && sce->sf_idx[w*16+g] < FFMIN(maxdeltasf, maxsf[w*16+g])
653  && (dists[w*16+g] < FFMIN(euplims[w*16+g], uplims[w*16+g]))
654  && (fabsf(qenergies[w*16+g]-energies[w*16+g]) < euplims[w*16+g])
655  ) {
656  /** Um... over target. Save bits for more important stuff. */
657  for (i = 0; i < depth && sce->sf_idx[w*16+g] < maxdeltasf; ++i) {
658  int cb, bits;
659  float dist, qenergy;
660  cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]+1);
661  if (cb > 0) {
662  dist = qenergy = 0.f;
663  bits = 0;
664  for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
665  int b;
666  float sqenergy;
667  dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
668  scaled + w2*128,
669  sce->ics.swb_sizes[g],
670  sce->sf_idx[w*16+g]+1,
671  cb,
672  1.0f,
673  INFINITY,
674  &b, &sqenergy,
675  0);
676  bits += b;
677  qenergy += sqenergy;
678  }
679  dist -= bits;
680  if (dist < FFMIN(euplims[w*16+g], uplims[w*16+g])) {
681  sce->sf_idx[w*16+g]++;
682  dists[w*16+g] = dist;
683  qenergies[w*16+g] = qenergy;
684  } else {
685  break;
686  }
687  } else {
688  maxsf[w*16+g] = FFMIN(sce->sf_idx[w*16+g], maxsf[w*16+g]);
689  break;
690  }
691  }
692  }
693  prev = sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], mindeltasf, maxdeltasf);
694  if (sce->sf_idx[w*16+g] != prevsc)
695  fflag = 1;
696  nminscaler = FFMIN(nminscaler, sce->sf_idx[w*16+g]);
697  sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
698  }
699  start += sce->ics.swb_sizes[g];
700  }
701  }
702 
703  /** SF difference limit violation risk. Must re-clamp. */
704  prev = -1;
705  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
706  for (g = 0; g < sce->ics.num_swb; g++) {
707  if (!sce->zeroes[w*16+g]) {
708  int prevsf = sce->sf_idx[w*16+g];
709  if (prev < 0)
710  prev = prevsf;
711  sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], prev - SCALE_MAX_DIFF, prev + SCALE_MAX_DIFF);
712  sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
713  prev = sce->sf_idx[w*16+g];
714  if (!fflag && prevsf != sce->sf_idx[w*16+g])
715  fflag = 1;
716  }
717  }
718  }
719 
720  its++;
721  } while (fflag && its < maxits);
722 
723  /** Scout out next nonzero bands */
724  ff_init_nextband_map(sce, nextband);
725 
726  prev = -1;
727  for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
728  /** Make sure proper codebooks are set */
729  for (g = 0; g < sce->ics.num_swb; g++) {
730  if (!sce->zeroes[w*16+g]) {
731  sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
732  if (sce->band_type[w*16+g] <= 0) {
733  if (!ff_sfdelta_can_remove_band(sce, nextband, prev, w*16+g)) {
734  /** Cannot zero out, make sure it's not attempted */
735  sce->band_type[w*16+g] = 1;
736  } else {
737  sce->zeroes[w*16+g] = 1;
738  sce->band_type[w*16+g] = 0;
739  }
740  }
741  } else {
742  sce->band_type[w*16+g] = 0;
743  }
744  /** Check that there's no SF delta range violations */
745  if (!sce->zeroes[w*16+g]) {
746  if (prev != -1) {
747  av_unused int sfdiff = sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO;
748  av_assert1(sfdiff >= 0 && sfdiff <= 2*SCALE_MAX_DIFF);
749  } else if (sce->zeroes[0]) {
750  /** Set global gain to something useful */
751  sce->sf_idx[0] = sce->sf_idx[w*16+g];
752  }
753  prev = sce->sf_idx[w*16+g];
754  }
755  }
756  }
757 }
758 
759 #endif /* AVCODEC_AACCODER_TWOLOOP_H */
INFINITY
#define INFINITY
Definition: mathematics.h:118
av_clip
#define av_clip
Definition: common.h:100
SingleChannelElement::can_pns
uint8_t can_pns[128]
band is allowed to PNS (informative)
Definition: aacenc.h:135
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:247
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:134
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:321
SCALE_MAX_POS
#define SCALE_MAX_POS
scalefactor index maximum value
Definition: aac.h:89
IndividualChannelStream::num_swb
int num_swb
number of scalefactor window bands
Definition: aacdec.h:171
SingleChannelElement::coeffs
float coeffs[1024]
coefficients for IMDCT, maybe processed
Definition: aacenc.h:139
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:87
TYPE_CPE
@ TYPE_CPE
Definition: aac.h:41
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:211
s
#define s(width, name)
Definition: cbs_vp9.c:198
AVCodecContext::global_quality
int global_quality
Global quality for codecs which cannot change it per frame.
Definition: avcodec.h:1243
IndividualChannelStream::swb_sizes
const uint8_t * swb_sizes
table of scalefactor band sizes for a particular window
Definition: aacenc.h:98
g
const char * g
Definition: vf_curves.c:128
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:91
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:122
SingleChannelElement::sf_idx
int sf_idx[128]
scalefactor indices
Definition: aacenc.h:133
ff_aac_scalefactor_bits
const uint8_t ff_aac_scalefactor_bits[121]
Definition: aactab.c:204
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:122
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:65
SingleChannelElement::band_type
enum BandType band_type[128]
band types
Definition: aacdec.h:214
mb
#define mb
Definition: vf_colormatrix.c:99
SCALE_MAX_DIFF
#define SCALE_MAX_DIFF
maximum scalefactor difference allowed by standard
Definition: aac.h:90
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:210
i
#define i(width, name, range_min, range_max)
Definition: cbs_h2645.c:256
IndividualChannelStream::num_windows
int num_windows
Definition: aacdec.h:172
SCALE_ONE_POS
#define SCALE_ONE_POS
scalefactor index that corresponds to scale=1.0
Definition: aac.h:88
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:170
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:238
avcodec.h
AACEncContext
AAC encoder context.
Definition: aacenc.h:212
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:168
ff_pns_bits
static int ff_pns_bits(SingleChannelElement *sce, int w, int g)
Definition: aaccoder_twoloop.h:57
ff_quantize_band_cost_cache_init
void ff_quantize_band_cost_cache_init(struct AACEncContext *s)
Definition: aacenc.c:401
aacenc.h
Generated on Sat Aug 31 2024 19:21:28 for FFmpeg by   doxygen 1.8.17