FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
ratecontrol.c
Go to the documentation of this file.
1 /*
2  * Rate control for video encoders
3  *
4  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
5  *
6  * This file is part of FFmpeg.
7  *
8  * FFmpeg is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU Lesser General Public
10  * License as published by the Free Software Foundation; either
11  * version 2.1 of the License, or (at your option) any later version.
12  *
13  * FFmpeg is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16  * Lesser General Public License for more details.
17  *
18  * You should have received a copy of the GNU Lesser General Public
19  * License along with FFmpeg; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  */
22 
23 /**
24  * @file
25  * Rate control for video encoders.
26  */
27 
28 #include "libavutil/attributes.h"
29 #include "avcodec.h"
30 #include "ratecontrol.h"
31 #include "mpegvideo.h"
32 #include "libavutil/eval.h"
33 
34 #undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
35 #include <assert.h>
36 
37 #ifndef M_E
38 #define M_E 2.718281828
39 #endif
40 
41 static int init_pass2(MpegEncContext *s);
42 static double get_qscale(MpegEncContext *s, RateControlEntry *rce,
43  double rate_factor, int frame_num);
44 
46 {
47  snprintf(s->avctx->stats_out, 256,
48  "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d "
49  "fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d;\n",
52  s->pict_type,
54  s->i_tex_bits,
55  s->p_tex_bits,
56  s->mv_bits,
57  s->misc_bits,
58  s->f_code,
59  s->b_code,
62  s->i_count, s->skip_count,
63  s->header_bits);
64 }
65 
66 static double get_fps(AVCodecContext *avctx)
67 {
68  return 1.0 / av_q2d(avctx->time_base) / FFMAX(avctx->ticks_per_frame, 1);
69 }
70 
71 static inline double qp2bits(RateControlEntry *rce, double qp)
72 {
73  if (qp <= 0.0) {
74  av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
75  }
76  return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / qp;
77 }
78 
79 static inline double bits2qp(RateControlEntry *rce, double bits)
80 {
81  if (bits < 0.9) {
82  av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
83  }
84  return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits + 1) / bits;
85 }
86 
88 {
89  RateControlContext *rcc = &s->rc_context;
90  int i, res;
91  static const char * const const_names[] = {
92  "PI",
93  "E",
94  "iTex",
95  "pTex",
96  "tex",
97  "mv",
98  "fCode",
99  "iCount",
100  "mcVar",
101  "var",
102  "isI",
103  "isP",
104  "isB",
105  "avgQP",
106  "qComp",
107 #if 0
108  "lastIQP",
109  "lastPQP",
110  "lastBQP",
111  "nextNonBQP",
112 #endif
113  "avgIITex",
114  "avgPITex",
115  "avgPPTex",
116  "avgBPTex",
117  "avgTex",
118  NULL
119  };
120  static double (* const func1[])(void *, double) = {
121  (void *)bits2qp,
122  (void *)qp2bits,
123  NULL
124  };
125  static const char * const func1_names[] = {
126  "bits2qp",
127  "qp2bits",
128  NULL
129  };
130  emms_c();
131 
133  if (s->avctx->rc_max_rate) {
134  s->avctx->rc_max_available_vbv_use = av_clipf(s->avctx->rc_max_rate/(s->avctx->rc_buffer_size*get_fps(s->avctx)), 1.0/3, 1.0);
135  } else
137  }
138 
139  res = av_expr_parse(&rcc->rc_eq_eval,
140  s->avctx->rc_eq ? s->avctx->rc_eq : "tex^qComp",
141  const_names, func1_names, func1,
142  NULL, NULL, 0, s->avctx);
143  if (res < 0) {
144  av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->avctx->rc_eq);
145  return res;
146  }
147 
148  for (i = 0; i < 5; i++) {
149  rcc->pred[i].coeff = FF_QP2LAMBDA * 7.0;
150  rcc->pred[i].count = 1.0;
151  rcc->pred[i].decay = 0.4;
152 
153  rcc->i_cplx_sum [i] =
154  rcc->p_cplx_sum [i] =
155  rcc->mv_bits_sum[i] =
156  rcc->qscale_sum [i] =
157  rcc->frame_count[i] = 1; // 1 is better because of 1/0 and such
158 
159  rcc->last_qscale_for[i] = FF_QP2LAMBDA * 5;
160  }
162  if (!rcc->buffer_index)
163  rcc->buffer_index = s->avctx->rc_buffer_size * 3 / 4;
164 
165  if (s->flags & CODEC_FLAG_PASS2) {
166  int i;
167  char *p;
168 
169  /* find number of pics */
170  p = s->avctx->stats_in;
171  for (i = -1; p; i++)
172  p = strchr(p + 1, ';');
173  i += s->max_b_frames;
174  if (i <= 0 || i >= INT_MAX / sizeof(RateControlEntry))
175  return -1;
176  rcc->entry = av_mallocz(i * sizeof(RateControlEntry));
177  rcc->num_entries = i;
178 
179  /* init all to skipped p frames
180  * (with b frames we might have a not encoded frame at the end FIXME) */
181  for (i = 0; i < rcc->num_entries; i++) {
182  RateControlEntry *rce = &rcc->entry[i];
183 
185  rce->qscale = rce->new_qscale = FF_QP2LAMBDA * 2;
186  rce->misc_bits = s->mb_num + 10;
187  rce->mb_var_sum = s->mb_num * 100;
188  }
189 
190  /* read stats */
191  p = s->avctx->stats_in;
192  for (i = 0; i < rcc->num_entries - s->max_b_frames; i++) {
193  RateControlEntry *rce;
194  int picture_number;
195  int e;
196  char *next;
197 
198  next = strchr(p, ';');
199  if (next) {
200  (*next) = 0; // sscanf in unbelievably slow on looong strings // FIXME copy / do not write
201  next++;
202  }
203  e = sscanf(p, " in:%d ", &picture_number);
204 
205  assert(picture_number >= 0);
206  assert(picture_number < rcc->num_entries);
207  rce = &rcc->entry[picture_number];
208 
209  e += sscanf(p, " in:%*d out:%*d type:%d q:%f itex:%d ptex:%d mv:%d misc:%d fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d",
210  &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits,
211  &rce->mv_bits, &rce->misc_bits,
212  &rce->f_code, &rce->b_code,
213  &rce->mc_mb_var_sum, &rce->mb_var_sum,
214  &rce->i_count, &rce->skip_count, &rce->header_bits);
215  if (e != 14) {
217  "statistics are damaged at line %d, parser out=%d\n",
218  i, e);
219  return -1;
220  }
221 
222  p = next;
223  }
224 
225  if (init_pass2(s) < 0)
226  return -1;
227 
228  // FIXME maybe move to end
230 #if CONFIG_LIBXVID
231  return ff_xvid_rate_control_init(s);
232 #else
234  "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n");
235  return -1;
236 #endif
237  }
238  }
239 
240  if (!(s->flags & CODEC_FLAG_PASS2)) {
241  rcc->short_term_qsum = 0.001;
242  rcc->short_term_qcount = 0.001;
243 
244  rcc->pass1_rc_eq_output_sum = 0.001;
245  rcc->pass1_wanted_bits = 0.001;
246 
247  if (s->avctx->qblur > 1.0) {
248  av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
249  return -1;
250  }
251  /* init stuff with the user specified complexity */
252  if (s->avctx->rc_initial_cplx) {
253  for (i = 0; i < 60 * 30; i++) {
254  double bits = s->avctx->rc_initial_cplx * (i / 10000.0 + 1.0) * s->mb_num;
255  RateControlEntry rce;
256 
257  if (i % ((s->gop_size + 3) / 4) == 0)
259  else if (i % (s->max_b_frames + 1))
261  else
263 
264  rce.new_pict_type = rce.pict_type;
265  rce.mc_mb_var_sum = bits * s->mb_num / 100000;
266  rce.mb_var_sum = s->mb_num;
267 
268  rce.qscale = FF_QP2LAMBDA * 2;
269  rce.f_code = 2;
270  rce.b_code = 1;
271  rce.misc_bits = 1;
272 
273  if (s->pict_type == AV_PICTURE_TYPE_I) {
274  rce.i_count = s->mb_num;
275  rce.i_tex_bits = bits;
276  rce.p_tex_bits = 0;
277  rce.mv_bits = 0;
278  } else {
279  rce.i_count = 0; // FIXME we do know this approx
280  rce.i_tex_bits = 0;
281  rce.p_tex_bits = bits * 0.9;
282  rce.mv_bits = bits * 0.1;
283  }
284  rcc->i_cplx_sum[rce.pict_type] += rce.i_tex_bits * rce.qscale;
285  rcc->p_cplx_sum[rce.pict_type] += rce.p_tex_bits * rce.qscale;
286  rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
287  rcc->frame_count[rce.pict_type]++;
288 
289  get_qscale(s, &rce, rcc->pass1_wanted_bits / rcc->pass1_rc_eq_output_sum, i);
290 
291  // FIXME misbehaves a little for variable fps
292  rcc->pass1_wanted_bits += s->bit_rate / get_fps(s->avctx);
293  }
294  }
295  }
296 
297  return 0;
298 }
299 
301 {
302  RateControlContext *rcc = &s->rc_context;
303  emms_c();
304 
305  av_expr_free(rcc->rc_eq_eval);
306  av_freep(&rcc->entry);
307 
308 #if CONFIG_LIBXVID
311 #endif
312 }
313 
315 {
316  RateControlContext *rcc = &s->rc_context;
317  const double fps = get_fps(s->avctx);
318  const int buffer_size = s->avctx->rc_buffer_size;
319  const double min_rate = s->avctx->rc_min_rate / fps;
320  const double max_rate = s->avctx->rc_max_rate / fps;
321 
322  av_dlog(s, "%d %f %d %f %f\n",
323  buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
324 
325  if (buffer_size) {
326  int left;
327 
328  rcc->buffer_index -= frame_size;
329  if (rcc->buffer_index < 0) {
330  av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
331  rcc->buffer_index = 0;
332  }
333 
334  left = buffer_size - rcc->buffer_index - 1;
335  rcc->buffer_index += av_clip(left, min_rate, max_rate);
336 
337  if (rcc->buffer_index > buffer_size) {
338  int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
339 
340  if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
341  stuffing = 4;
342  rcc->buffer_index -= 8 * stuffing;
343 
344  if (s->avctx->debug & FF_DEBUG_RC)
345  av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
346 
347  return stuffing;
348  }
349  }
350  return 0;
351 }
352 
353 /**
354  * Modify the bitrate curve from pass1 for one frame.
355  */
357  double rate_factor, int frame_num)
358 {
359  RateControlContext *rcc = &s->rc_context;
360  AVCodecContext *a = s->avctx;
361  const int pict_type = rce->new_pict_type;
362  const double mb_num = s->mb_num;
363  double q, bits;
364  int i;
365 
366  double const_values[] = {
367  M_PI,
368  M_E,
369  rce->i_tex_bits * rce->qscale,
370  rce->p_tex_bits * rce->qscale,
371  (rce->i_tex_bits + rce->p_tex_bits) * (double)rce->qscale,
372  rce->mv_bits / mb_num,
373  rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code) * 0.5 : rce->f_code,
374  rce->i_count / mb_num,
375  rce->mc_mb_var_sum / mb_num,
376  rce->mb_var_sum / mb_num,
380  rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
381  a->qcompress,
382 #if 0
386  rcc->next_non_b_qscale,
387 #endif
392  (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
393  0
394  };
395 
396  bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
397  if (isnan(bits)) {
398  av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq);
399  return -1;
400  }
401 
403  bits *= rate_factor;
404  if (bits < 0.0)
405  bits = 0.0;
406  bits += 1.0; // avoid 1/0 issues
407 
408  /* user override */
409  for (i = 0; i < s->avctx->rc_override_count; i++) {
410  RcOverride *rco = s->avctx->rc_override;
411  if (rco[i].start_frame > frame_num)
412  continue;
413  if (rco[i].end_frame < frame_num)
414  continue;
415 
416  if (rco[i].qscale)
417  bits = qp2bits(rce, rco[i].qscale); // FIXME move at end to really force it?
418  else
419  bits *= rco[i].quality_factor;
420  }
421 
422  q = bits2qp(rce, bits);
423 
424  /* I/B difference */
425  if (pict_type == AV_PICTURE_TYPE_I && s->avctx->i_quant_factor < 0.0)
426  q = -q * s->avctx->i_quant_factor + s->avctx->i_quant_offset;
427  else if (pict_type == AV_PICTURE_TYPE_B && s->avctx->b_quant_factor < 0.0)
428  q = -q * s->avctx->b_quant_factor + s->avctx->b_quant_offset;
429  if (q < 1)
430  q = 1;
431 
432  return q;
433 }
434 
435 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q)
436 {
437  RateControlContext *rcc = &s->rc_context;
438  AVCodecContext *a = s->avctx;
439  const int pict_type = rce->new_pict_type;
440  const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
441  const double last_non_b_q = rcc->last_qscale_for[rcc->last_non_b_pict_type];
442 
443  if (pict_type == AV_PICTURE_TYPE_I &&
445  q = last_p_q * FFABS(a->i_quant_factor) + a->i_quant_offset;
446  else if (pict_type == AV_PICTURE_TYPE_B &&
447  a->b_quant_factor > 0.0)
448  q = last_non_b_q * a->b_quant_factor + a->b_quant_offset;
449  if (q < 1)
450  q = 1;
451 
452  /* last qscale / qdiff stuff */
453  if (rcc->last_non_b_pict_type == pict_type || pict_type != AV_PICTURE_TYPE_I) {
454  double last_q = rcc->last_qscale_for[pict_type];
455  const int maxdiff = FF_QP2LAMBDA * a->max_qdiff;
456 
457  if (q > last_q + maxdiff)
458  q = last_q + maxdiff;
459  else if (q < last_q - maxdiff)
460  q = last_q - maxdiff;
461  }
462 
463  rcc->last_qscale_for[pict_type] = q; // Note we cannot do that after blurring
464 
465  if (pict_type != AV_PICTURE_TYPE_B)
466  rcc->last_non_b_pict_type = pict_type;
467 
468  return q;
469 }
470 
471 /**
472  * Get the qmin & qmax for pict_type.
473  */
474 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type)
475 {
476  int qmin = s->avctx->lmin;
477  int qmax = s->avctx->lmax;
478 
479  assert(qmin <= qmax);
480 
481  switch (pict_type) {
482  case AV_PICTURE_TYPE_B:
483  qmin = (int)(qmin * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
484  qmax = (int)(qmax * FFABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset + 0.5);
485  break;
486  case AV_PICTURE_TYPE_I:
487  qmin = (int)(qmin * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
488  qmax = (int)(qmax * FFABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset + 0.5);
489  break;
490  }
491 
492  qmin = av_clip(qmin, 1, FF_LAMBDA_MAX);
493  qmax = av_clip(qmax, 1, FF_LAMBDA_MAX);
494 
495  if (qmax < qmin)
496  qmax = qmin;
497 
498  *qmin_ret = qmin;
499  *qmax_ret = qmax;
500 }
501 
503  double q, int frame_num)
504 {
505  RateControlContext *rcc = &s->rc_context;
506  const double buffer_size = s->avctx->rc_buffer_size;
507  const double fps = get_fps(s->avctx);
508  const double min_rate = s->avctx->rc_min_rate / fps;
509  const double max_rate = s->avctx->rc_max_rate / fps;
510  const int pict_type = rce->new_pict_type;
511  int qmin, qmax;
512 
513  get_qminmax(&qmin, &qmax, s, pict_type);
514 
515  /* modulation */
516  if (s->avctx->rc_qmod_freq &&
517  frame_num % s->avctx->rc_qmod_freq == 0 &&
518  pict_type == AV_PICTURE_TYPE_P)
519  q *= s->avctx->rc_qmod_amp;
520 
521  /* buffer overflow/underflow protection */
522  if (buffer_size) {
523  double expected_size = rcc->buffer_index;
524  double q_limit;
525 
526  if (min_rate) {
527  double d = 2 * (buffer_size - expected_size) / buffer_size;
528  if (d > 1.0)
529  d = 1.0;
530  else if (d < 0.0001)
531  d = 0.0001;
532  q *= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity);
533 
534  q_limit = bits2qp(rce,
535  FFMAX((min_rate - buffer_size + rcc->buffer_index) *
537 
538  if (q > q_limit) {
539  if (s->avctx->debug & FF_DEBUG_RC)
541  "limiting QP %f -> %f\n", q, q_limit);
542  q = q_limit;
543  }
544  }
545 
546  if (max_rate) {
547  double d = 2 * expected_size / buffer_size;
548  if (d > 1.0)
549  d = 1.0;
550  else if (d < 0.0001)
551  d = 0.0001;
552  q /= pow(d, 1.0 / s->avctx->rc_buffer_aggressivity);
553 
554  q_limit = bits2qp(rce,
555  FFMAX(rcc->buffer_index *
557  1));
558  if (q < q_limit) {
559  if (s->avctx->debug & FF_DEBUG_RC)
561  "limiting QP %f -> %f\n", q, q_limit);
562  q = q_limit;
563  }
564  }
565  }
566  av_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
567  q, max_rate, min_rate, buffer_size, rcc->buffer_index,
569  if (s->avctx->rc_qsquish == 0.0 || qmin == qmax) {
570  if (q < qmin)
571  q = qmin;
572  else if (q > qmax)
573  q = qmax;
574  } else {
575  double min2 = log(qmin);
576  double max2 = log(qmax);
577 
578  q = log(q);
579  q = (q - min2) / (max2 - min2) - 0.5;
580  q *= -4.0;
581  q = 1.0 / (1.0 + exp(q));
582  q = q * (max2 - min2) + min2;
583 
584  q = exp(q);
585  }
586 
587  return q;
588 }
589 
590 // ----------------------------------
591 // 1 Pass Code
592 
593 static double predict_size(Predictor *p, double q, double var)
594 {
595  return p->coeff * var / (q * p->count);
596 }
597 
598 static void update_predictor(Predictor *p, double q, double var, double size)
599 {
600  double new_coeff = size * q / (var + 1);
601  if (var < 10)
602  return;
603 
604  p->count *= p->decay;
605  p->coeff *= p->decay;
606  p->count++;
607  p->coeff += new_coeff;
608 }
609 
610 static void adaptive_quantization(MpegEncContext *s, double q)
611 {
612  int i;
613  const float lumi_masking = s->avctx->lumi_masking / (128.0 * 128.0);
614  const float dark_masking = s->avctx->dark_masking / (128.0 * 128.0);
615  const float temp_cplx_masking = s->avctx->temporal_cplx_masking;
616  const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
617  const float p_masking = s->avctx->p_masking;
618  const float border_masking = s->avctx->border_masking;
619  float bits_sum = 0.0;
620  float cplx_sum = 0.0;
621  float *cplx_tab = s->cplx_tab;
622  float *bits_tab = s->bits_tab;
623  const int qmin = s->avctx->mb_lmin;
624  const int qmax = s->avctx->mb_lmax;
625  Picture *const pic = &s->current_picture;
626  const int mb_width = s->mb_width;
627  const int mb_height = s->mb_height;
628 
629  for (i = 0; i < s->mb_num; i++) {
630  const int mb_xy = s->mb_index2xy[i];
631  float temp_cplx = sqrt(pic->mc_mb_var[mb_xy]); // FIXME merge in pow()
632  float spat_cplx = sqrt(pic->mb_var[mb_xy]);
633  const int lumi = pic->mb_mean[mb_xy];
634  float bits, cplx, factor;
635  int mb_x = mb_xy % s->mb_stride;
636  int mb_y = mb_xy / s->mb_stride;
637  int mb_distance;
638  float mb_factor = 0.0;
639  if (spat_cplx < 4)
640  spat_cplx = 4; // FIXME finetune
641  if (temp_cplx < 4)
642  temp_cplx = 4; // FIXME finetune
643 
644  if ((s->mb_type[mb_xy] & CANDIDATE_MB_TYPE_INTRA)) { // FIXME hq mode
645  cplx = spat_cplx;
646  factor = 1.0 + p_masking;
647  } else {
648  cplx = temp_cplx;
649  factor = pow(temp_cplx, -temp_cplx_masking);
650  }
651  factor *= pow(spat_cplx, -spatial_cplx_masking);
652 
653  if (lumi > 127)
654  factor *= (1.0 - (lumi - 128) * (lumi - 128) * lumi_masking);
655  else
656  factor *= (1.0 - (lumi - 128) * (lumi - 128) * dark_masking);
657 
658  if (mb_x < mb_width / 5) {
659  mb_distance = mb_width / 5 - mb_x;
660  mb_factor = (float)mb_distance / (float)(mb_width / 5);
661  } else if (mb_x > 4 * mb_width / 5) {
662  mb_distance = mb_x - 4 * mb_width / 5;
663  mb_factor = (float)mb_distance / (float)(mb_width / 5);
664  }
665  if (mb_y < mb_height / 5) {
666  mb_distance = mb_height / 5 - mb_y;
667  mb_factor = FFMAX(mb_factor,
668  (float)mb_distance / (float)(mb_height / 5));
669  } else if (mb_y > 4 * mb_height / 5) {
670  mb_distance = mb_y - 4 * mb_height / 5;
671  mb_factor = FFMAX(mb_factor,
672  (float)mb_distance / (float)(mb_height / 5));
673  }
674 
675  factor *= 1.0 - border_masking * mb_factor;
676 
677  if (factor < 0.00001)
678  factor = 0.00001;
679 
680  bits = cplx * factor;
681  cplx_sum += cplx;
682  bits_sum += bits;
683  cplx_tab[i] = cplx;
684  bits_tab[i] = bits;
685  }
686 
687  /* handle qmin/qmax clipping */
688  if (s->flags & CODEC_FLAG_NORMALIZE_AQP) {
689  float factor = bits_sum / cplx_sum;
690  for (i = 0; i < s->mb_num; i++) {
691  float newq = q * cplx_tab[i] / bits_tab[i];
692  newq *= factor;
693 
694  if (newq > qmax) {
695  bits_sum -= bits_tab[i];
696  cplx_sum -= cplx_tab[i] * q / qmax;
697  } else if (newq < qmin) {
698  bits_sum -= bits_tab[i];
699  cplx_sum -= cplx_tab[i] * q / qmin;
700  }
701  }
702  if (bits_sum < 0.001)
703  bits_sum = 0.001;
704  if (cplx_sum < 0.001)
705  cplx_sum = 0.001;
706  }
707 
708  for (i = 0; i < s->mb_num; i++) {
709  const int mb_xy = s->mb_index2xy[i];
710  float newq = q * cplx_tab[i] / bits_tab[i];
711  int intq;
712 
713  if (s->flags & CODEC_FLAG_NORMALIZE_AQP) {
714  newq *= bits_sum / cplx_sum;
715  }
716 
717  intq = (int)(newq + 0.5);
718 
719  if (intq > qmax)
720  intq = qmax;
721  else if (intq < qmin)
722  intq = qmin;
723  s->lambda_table[mb_xy] = intq;
724  }
725 }
726 
728 {
729  RateControlContext *rcc = &s->rc_context;
730  RateControlEntry *rce = &rcc->entry[s->picture_number];
731 
732  s->f_code = rce->f_code;
733  s->b_code = rce->b_code;
734 }
735 
736 // FIXME rd or at least approx for dquant
737 
739 {
740  float q;
741  int qmin, qmax;
742  float br_compensation;
743  double diff;
744  double short_term_q;
745  double fps;
746  int picture_number = s->picture_number;
747  int64_t wanted_bits;
748  RateControlContext *rcc = &s->rc_context;
749  AVCodecContext *a = s->avctx;
750  RateControlEntry local_rce, *rce;
751  double bits;
752  double rate_factor;
753  int var;
754  const int pict_type = s->pict_type;
755  Picture * const pic = &s->current_picture;
756  emms_c();
757 
758 #if CONFIG_LIBXVID
759  if ((s->flags & CODEC_FLAG_PASS2) &&
761  return ff_xvid_rate_estimate_qscale(s, dry_run);
762 #endif
763 
764  get_qminmax(&qmin, &qmax, s, pict_type);
765 
766  fps = get_fps(s->avctx);
767  /* update predictors */
768  if (picture_number > 2 && !dry_run) {
769  const int last_var = s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum
770  : rcc->last_mc_mb_var_sum;
773  rcc->last_qscale,
774  sqrt(last_var),
775  s->frame_bits - s->stuffing_bits);
776  }
777 
778  if (s->flags & CODEC_FLAG_PASS2) {
779  assert(picture_number >= 0);
780  if (picture_number >= rcc->num_entries) {
781  av_log(s, AV_LOG_ERROR, "Input is longer than 2-pass log file\n");
782  return -1;
783  }
784  rce = &rcc->entry[picture_number];
785  wanted_bits = rce->expected_bits;
786  } else {
787  Picture *dts_pic;
788  rce = &local_rce;
789 
790  /* FIXME add a dts field to AVFrame and ensure it is set and use it
791  * here instead of reordering but the reordering is simpler for now
792  * until H.264 B-pyramid must be handled. */
793  if (s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
794  dts_pic = s->current_picture_ptr;
795  else
796  dts_pic = s->last_picture_ptr;
797 
798  if (!dts_pic || dts_pic->f.pts == AV_NOPTS_VALUE)
799  wanted_bits = (uint64_t)(s->bit_rate * (double)picture_number / fps);
800  else
801  wanted_bits = (uint64_t)(s->bit_rate * (double)dts_pic->f.pts / fps);
802  }
803 
804  diff = s->total_bits - wanted_bits;
805  br_compensation = (a->bit_rate_tolerance - diff) / a->bit_rate_tolerance;
806  if (br_compensation <= 0.0)
807  br_compensation = 0.001;
808 
809  var = pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
810 
811  short_term_q = 0; /* avoid warning */
812  if (s->flags & CODEC_FLAG_PASS2) {
813  if (pict_type != AV_PICTURE_TYPE_I)
814  assert(pict_type == rce->new_pict_type);
815 
816  q = rce->new_qscale / br_compensation;
817  av_dlog(s, "%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale,
818  br_compensation, s->frame_bits, var, pict_type);
819  } else {
820  rce->pict_type =
821  rce->new_pict_type = pict_type;
822  rce->mc_mb_var_sum = pic->mc_mb_var_sum;
823  rce->mb_var_sum = pic->mb_var_sum;
824  rce->qscale = FF_QP2LAMBDA * 2;
825  rce->f_code = s->f_code;
826  rce->b_code = s->b_code;
827  rce->misc_bits = 1;
828 
829  bits = predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
830  if (pict_type == AV_PICTURE_TYPE_I) {
831  rce->i_count = s->mb_num;
832  rce->i_tex_bits = bits;
833  rce->p_tex_bits = 0;
834  rce->mv_bits = 0;
835  } else {
836  rce->i_count = 0; // FIXME we do know this approx
837  rce->i_tex_bits = 0;
838  rce->p_tex_bits = bits * 0.9;
839  rce->mv_bits = bits * 0.1;
840  }
841  rcc->i_cplx_sum[pict_type] += rce->i_tex_bits * rce->qscale;
842  rcc->p_cplx_sum[pict_type] += rce->p_tex_bits * rce->qscale;
843  rcc->mv_bits_sum[pict_type] += rce->mv_bits;
844  rcc->frame_count[pict_type]++;
845 
846  bits = rce->i_tex_bits + rce->p_tex_bits;
847  rate_factor = rcc->pass1_wanted_bits /
848  rcc->pass1_rc_eq_output_sum * br_compensation;
849 
850  q = get_qscale(s, rce, rate_factor, picture_number);
851  if (q < 0)
852  return -1;
853 
854  assert(q > 0.0);
855  q = get_diff_limited_q(s, rce, q);
856  assert(q > 0.0);
857 
858  // FIXME type dependent blur like in 2-pass
859  if (pict_type == AV_PICTURE_TYPE_P || s->intra_only) {
860  rcc->short_term_qsum *= a->qblur;
861  rcc->short_term_qcount *= a->qblur;
862 
863  rcc->short_term_qsum += q;
864  rcc->short_term_qcount++;
865  q = short_term_q = rcc->short_term_qsum / rcc->short_term_qcount;
866  }
867  assert(q > 0.0);
868 
869  q = modify_qscale(s, rce, q, picture_number);
870 
871  rcc->pass1_wanted_bits += s->bit_rate / fps;
872 
873  assert(q > 0.0);
874  }
875 
876  if (s->avctx->debug & FF_DEBUG_RC) {
878  "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f "
879  "size:%d var:%d/%d br:%d fps:%d\n",
880  av_get_picture_type_char(pict_type),
881  qmin, q, qmax, picture_number,
882  (int)wanted_bits / 1000, (int)s->total_bits / 1000,
883  br_compensation, short_term_q, s->frame_bits,
884  pic->mb_var_sum, pic->mc_mb_var_sum,
885  s->bit_rate / 1000, (int)fps);
886  }
887 
888  if (q < qmin)
889  q = qmin;
890  else if (q > qmax)
891  q = qmax;
892 
893  if (s->adaptive_quant)
894  adaptive_quantization(s, q);
895  else
896  q = (int)(q + 0.5);
897 
898  if (!dry_run) {
899  rcc->last_qscale = q;
900  rcc->last_mc_mb_var_sum = pic->mc_mb_var_sum;
901  rcc->last_mb_var_sum = pic->mb_var_sum;
902  }
903  return q;
904 }
905 
906 // ----------------------------------------------
907 // 2-Pass code
908 
910 {
911  RateControlContext *rcc = &s->rc_context;
912  AVCodecContext *a = s->avctx;
913  int i, toobig;
914  double fps = get_fps(s->avctx);
915  double complexity[5] = { 0 }; // approximate bits at quant=1
916  uint64_t const_bits[5] = { 0 }; // quantizer independent bits
917  uint64_t all_const_bits;
918  uint64_t all_available_bits = (uint64_t)(s->bit_rate *
919  (double)rcc->num_entries / fps);
920  double rate_factor = 0;
921  double step;
922  const int filter_size = (int)(a->qblur * 4) | 1;
923  double expected_bits = 0; // init to silence gcc warning
924  double *qscale, *blurred_qscale, qscale_sum;
925 
926  /* find complexity & const_bits & decide the pict_types */
927  for (i = 0; i < rcc->num_entries; i++) {
928  RateControlEntry *rce = &rcc->entry[i];
929 
930  rce->new_pict_type = rce->pict_type;
931  rcc->i_cplx_sum[rce->pict_type] += rce->i_tex_bits * rce->qscale;
932  rcc->p_cplx_sum[rce->pict_type] += rce->p_tex_bits * rce->qscale;
933  rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
934  rcc->frame_count[rce->pict_type]++;
935 
936  complexity[rce->new_pict_type] += (rce->i_tex_bits + rce->p_tex_bits) *
937  (double)rce->qscale;
938  const_bits[rce->new_pict_type] += rce->mv_bits + rce->misc_bits;
939  }
940 
941  all_const_bits = const_bits[AV_PICTURE_TYPE_I] +
942  const_bits[AV_PICTURE_TYPE_P] +
943  const_bits[AV_PICTURE_TYPE_B];
944 
945  if (all_available_bits < all_const_bits) {
946  av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
947  return -1;
948  }
949 
950  qscale = av_malloc(sizeof(double) * rcc->num_entries);
951  blurred_qscale = av_malloc(sizeof(double) * rcc->num_entries);
952  toobig = 0;
953 
954  for (step = 256 * 256; step > 0.0000001; step *= 0.5) {
955  expected_bits = 0;
956  rate_factor += step;
957 
958  rcc->buffer_index = s->avctx->rc_buffer_size / 2;
959 
960  /* find qscale */
961  for (i = 0; i < rcc->num_entries; i++) {
962  RateControlEntry *rce = &rcc->entry[i];
963 
964  qscale[i] = get_qscale(s, &rcc->entry[i], rate_factor, i);
965  rcc->last_qscale_for[rce->pict_type] = qscale[i];
966  }
967  assert(filter_size % 2 == 1);
968 
969  /* fixed I/B QP relative to P mode */
970  for (i = FFMAX(0, rcc->num_entries - 300); i < rcc->num_entries; i++) {
971  RateControlEntry *rce = &rcc->entry[i];
972 
973  qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
974  }
975 
976  for (i = rcc->num_entries - 1; i >= 0; i--) {
977  RateControlEntry *rce = &rcc->entry[i];
978 
979  qscale[i] = get_diff_limited_q(s, rce, qscale[i]);
980  }
981 
982  /* smooth curve */
983  for (i = 0; i < rcc->num_entries; i++) {
984  RateControlEntry *rce = &rcc->entry[i];
985  const int pict_type = rce->new_pict_type;
986  int j;
987  double q = 0.0, sum = 0.0;
988 
989  for (j = 0; j < filter_size; j++) {
990  int index = i + j - filter_size / 2;
991  double d = index - i;
992  double coeff = a->qblur == 0 ? 1.0 : exp(-d * d / (a->qblur * a->qblur));
993 
994  if (index < 0 || index >= rcc->num_entries)
995  continue;
996  if (pict_type != rcc->entry[index].new_pict_type)
997  continue;
998  q += qscale[index] * coeff;
999  sum += coeff;
1000  }
1001  blurred_qscale[i] = q / sum;
1002  }
1003 
1004  /* find expected bits */
1005  for (i = 0; i < rcc->num_entries; i++) {
1006  RateControlEntry *rce = &rcc->entry[i];
1007  double bits;
1008 
1009  rce->new_qscale = modify_qscale(s, rce, blurred_qscale[i], i);
1010 
1011  bits = qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
1012  bits += 8 * ff_vbv_update(s, bits);
1013 
1014  rce->expected_bits = expected_bits;
1015  expected_bits += bits;
1016  }
1017 
1018  av_dlog(s->avctx,
1019  "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
1020  expected_bits, (int)all_available_bits, rate_factor);
1021  if (expected_bits > all_available_bits) {
1022  rate_factor -= step;
1023  ++toobig;
1024  }
1025  }
1026  av_free(qscale);
1027  av_free(blurred_qscale);
1028 
1029  /* check bitrate calculations and print info */
1030  qscale_sum = 0.0;
1031  for (i = 0; i < rcc->num_entries; i++) {
1032  av_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
1033  i,
1034  rcc->entry[i].new_qscale,
1035  rcc->entry[i].new_qscale / FF_QP2LAMBDA);
1036  qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA,
1037  s->avctx->qmin, s->avctx->qmax);
1038  }
1039  assert(toobig <= 40);
1041  "[lavc rc] requested bitrate: %d bps expected bitrate: %d bps\n",
1042  s->bit_rate,
1043  (int)(expected_bits / ((double)all_available_bits / s->bit_rate)));
1045  "[lavc rc] estimated target average qp: %.3f\n",
1046  (float)qscale_sum / rcc->num_entries);
1047  if (toobig == 0) {
1048  av_log(s->avctx, AV_LOG_INFO,
1049  "[lavc rc] Using all of requested bitrate is not "
1050  "necessary for this video with these parameters.\n");
1051  } else if (toobig == 40) {
1053  "[lavc rc] Error: bitrate too low for this video "
1054  "with these parameters.\n");
1055  return -1;
1056  } else if (fabs(expected_bits / all_available_bits - 1.0) > 0.01) {
1058  "[lavc rc] Error: 2pass curve failed to converge\n");
1059  return -1;
1060  }
1061 
1062  return 0;
1063 }