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 Libav.
7  *
8  * Libav 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  * Libav 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 Libav; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  */
22 
28 #include "libavutil/intmath.h"
29 #include "avcodec.h"
30 #include "dsputil.h"
31 #include "ratecontrol.h"
32 #include "mpegvideo.h"
33 #include "libavutil/eval.h"
34 
35 #undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
36 #include <assert.h>
37 
38 #ifndef M_E
39 #define M_E 2.718281828
40 #endif
41 
42 static int init_pass2(MpegEncContext *s);
43 static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num);
44 
46  snprintf(s->avctx->stats_out, 256, "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d;\n",
50 }
51 
52 static inline double qp2bits(RateControlEntry *rce, double qp){
53  if(qp<=0.0){
54  av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
55  }
56  return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ qp;
57 }
58 
59 static inline double bits2qp(RateControlEntry *rce, double bits){
60  if(bits<0.9){
61  av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
62  }
63  return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ bits;
64 }
65 
67 {
69  int i, res;
70  static const char * const const_names[]={
71  "PI",
72  "E",
73  "iTex",
74  "pTex",
75  "tex",
76  "mv",
77  "fCode",
78  "iCount",
79  "mcVar",
80  "var",
81  "isI",
82  "isP",
83  "isB",
84  "avgQP",
85  "qComp",
86 /* "lastIQP",
87  "lastPQP",
88  "lastBQP",
89  "nextNonBQP",*/
90  "avgIITex",
91  "avgPITex",
92  "avgPPTex",
93  "avgBPTex",
94  "avgTex",
95  NULL
96  };
97  static double (* const func1[])(void *, double)={
98  (void *)bits2qp,
99  (void *)qp2bits,
100  NULL
101  };
102  static const char * const func1_names[]={
103  "bits2qp",
104  "qp2bits",
105  NULL
106  };
107  emms_c();
108 
109  res = av_expr_parse(&rcc->rc_eq_eval, s->avctx->rc_eq ? s->avctx->rc_eq : "tex^qComp", const_names, func1_names, func1, NULL, NULL, 0, s->avctx);
110  if (res < 0) {
111  av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->avctx->rc_eq);
112  return res;
113  }
114 
115  for(i=0; i<5; i++){
116  rcc->pred[i].coeff= FF_QP2LAMBDA * 7.0;
117  rcc->pred[i].count= 1.0;
118 
119  rcc->pred[i].decay= 0.4;
120  rcc->i_cplx_sum [i]=
121  rcc->p_cplx_sum [i]=
122  rcc->mv_bits_sum[i]=
123  rcc->qscale_sum [i]=
124  rcc->frame_count[i]= 1; // 1 is better because of 1/0 and such
125  rcc->last_qscale_for[i]=FF_QP2LAMBDA * 5;
126  }
128 
129  if(s->flags&CODEC_FLAG_PASS2){
130  int i;
131  char *p;
132 
133  /* find number of pics */
134  p= s->avctx->stats_in;
135  for(i=-1; p; i++){
136  p= strchr(p+1, ';');
137  }
138  i+= s->max_b_frames;
139  if(i<=0 || i>=INT_MAX / sizeof(RateControlEntry))
140  return -1;
141  rcc->entry = av_mallocz(i*sizeof(RateControlEntry));
142  rcc->num_entries= i;
143 
144  /* init all to skipped p frames (with b frames we might have a not encoded frame at the end FIXME) */
145  for(i=0; i<rcc->num_entries; i++){
146  RateControlEntry *rce= &rcc->entry[i];
148  rce->qscale= rce->new_qscale=FF_QP2LAMBDA * 2;
149  rce->misc_bits= s->mb_num + 10;
150  rce->mb_var_sum= s->mb_num*100;
151  }
152 
153  /* read stats */
154  p= s->avctx->stats_in;
155  for(i=0; i<rcc->num_entries - s->max_b_frames; i++){
156  RateControlEntry *rce;
157  int picture_number;
158  int e;
159  char *next;
160 
161  next= strchr(p, ';');
162  if(next){
163  (*next)=0; //sscanf in unbelievably slow on looong strings //FIXME copy / do not write
164  next++;
165  }
166  e= sscanf(p, " in:%d ", &picture_number);
167 
168  assert(picture_number >= 0);
169  assert(picture_number < rcc->num_entries);
170  rce= &rcc->entry[picture_number];
171 
172  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",
173  &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits, &rce->mv_bits, &rce->misc_bits,
174  &rce->f_code, &rce->b_code, &rce->mc_mb_var_sum, &rce->mb_var_sum, &rce->i_count, &rce->skip_count, &rce->header_bits);
175  if(e!=14){
176  av_log(s->avctx, AV_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
177  return -1;
178  }
179 
180  p= next;
181  }
182 
183  if(init_pass2(s) < 0) return -1;
184 
185  //FIXME maybe move to end
187 #if CONFIG_LIBXVID
188  return ff_xvid_rate_control_init(s);
189 #else
190  av_log(s->avctx, AV_LOG_ERROR, "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n");
191  return -1;
192 #endif
193  }
194  }
195 
196  if(!(s->flags&CODEC_FLAG_PASS2)){
197 
198  rcc->short_term_qsum=0.001;
199  rcc->short_term_qcount=0.001;
200 
201  rcc->pass1_rc_eq_output_sum= 0.001;
202  rcc->pass1_wanted_bits=0.001;
203 
204  if(s->avctx->qblur > 1.0){
205  av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
206  return -1;
207  }
208  /* init stuff with the user specified complexity */
209  if(s->avctx->rc_initial_cplx){
210  for(i=0; i<60*30; i++){
211  double bits= s->avctx->rc_initial_cplx * (i/10000.0 + 1.0)*s->mb_num;
212  RateControlEntry rce;
213 
214  if (i%((s->gop_size+3)/4)==0) rce.pict_type= AV_PICTURE_TYPE_I;
215  else if(i%(s->max_b_frames+1)) rce.pict_type= AV_PICTURE_TYPE_B;
216  else rce.pict_type= AV_PICTURE_TYPE_P;
217 
218  rce.new_pict_type= rce.pict_type;
219  rce.mc_mb_var_sum= bits*s->mb_num/100000;
220  rce.mb_var_sum = s->mb_num;
221  rce.qscale = FF_QP2LAMBDA * 2;
222  rce.f_code = 2;
223  rce.b_code = 1;
224  rce.misc_bits= 1;
225 
226  if(s->pict_type== AV_PICTURE_TYPE_I){
227  rce.i_count = s->mb_num;
228  rce.i_tex_bits= bits;
229  rce.p_tex_bits= 0;
230  rce.mv_bits= 0;
231  }else{
232  rce.i_count = 0; //FIXME we do know this approx
233  rce.i_tex_bits= 0;
234  rce.p_tex_bits= bits*0.9;
235  rce.mv_bits= bits*0.1;
236  }
237  rcc->i_cplx_sum [rce.pict_type] += rce.i_tex_bits*rce.qscale;
238  rcc->p_cplx_sum [rce.pict_type] += rce.p_tex_bits*rce.qscale;
239  rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
240  rcc->frame_count[rce.pict_type] ++;
241 
242  get_qscale(s, &rce, rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum, i);
243  rcc->pass1_wanted_bits+= s->bit_rate/(1/av_q2d(s->avctx->time_base)); //FIXME misbehaves a little for variable fps
244  }
245  }
246 
247  }
248 
249  return 0;
250 }
251 
253 {
254  RateControlContext *rcc= &s->rc_context;
255  emms_c();
256 
257  av_expr_free(rcc->rc_eq_eval);
258  av_freep(&rcc->entry);
259 
260 #if CONFIG_LIBXVID
263 #endif
264 }
265 
266 int ff_vbv_update(MpegEncContext *s, int frame_size){
267  RateControlContext *rcc= &s->rc_context;
268  const double fps= 1/av_q2d(s->avctx->time_base);
269  const int buffer_size= s->avctx->rc_buffer_size;
270  const double min_rate= s->avctx->rc_min_rate/fps;
271  const double max_rate= s->avctx->rc_max_rate/fps;
272 
273 //printf("%d %f %d %f %f\n", buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
274  if(buffer_size){
275  int left;
276 
277  rcc->buffer_index-= frame_size;
278  if(rcc->buffer_index < 0){
279  av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
280  rcc->buffer_index= 0;
281  }
282 
283  left= buffer_size - rcc->buffer_index - 1;
284  rcc->buffer_index += av_clip(left, min_rate, max_rate);
285 
286  if(rcc->buffer_index > buffer_size){
287  int stuffing= ceil((rcc->buffer_index - buffer_size)/8);
288 
289  if(stuffing < 4 && s->codec_id == CODEC_ID_MPEG4)
290  stuffing=4;
291  rcc->buffer_index -= 8*stuffing;
292 
293  if(s->avctx->debug & FF_DEBUG_RC)
294  av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
295 
296  return stuffing;
297  }
298  }
299  return 0;
300 }
301 
305 static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num){
306  RateControlContext *rcc= &s->rc_context;
307  AVCodecContext *a= s->avctx;
308  double q, bits;
309  const int pict_type= rce->new_pict_type;
310  const double mb_num= s->mb_num;
311  int i;
312 
313  double const_values[]={
314  M_PI,
315  M_E,
316  rce->i_tex_bits*rce->qscale,
317  rce->p_tex_bits*rce->qscale,
318  (rce->i_tex_bits + rce->p_tex_bits)*(double)rce->qscale,
319  rce->mv_bits/mb_num,
320  rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code)*0.5 : rce->f_code,
321  rce->i_count/mb_num,
322  rce->mc_mb_var_sum/mb_num,
323  rce->mb_var_sum/mb_num,
327  rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
328  a->qcompress,
329 /* rcc->last_qscale_for[AV_PICTURE_TYPE_I],
330  rcc->last_qscale_for[AV_PICTURE_TYPE_P],
331  rcc->last_qscale_for[AV_PICTURE_TYPE_B],
332  rcc->next_non_b_qscale,*/
337  (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
338  0
339  };
340 
341  bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
342  if (isnan(bits)) {
343  av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq);
344  return -1;
345  }
346 
348  bits*=rate_factor;
349  if(bits<0.0) bits=0.0;
350  bits+= 1.0; //avoid 1/0 issues
351 
352  /* user override */
353  for(i=0; i<s->avctx->rc_override_count; i++){
354  RcOverride *rco= s->avctx->rc_override;
355  if(rco[i].start_frame > frame_num) continue;
356  if(rco[i].end_frame < frame_num) continue;
357 
358  if(rco[i].qscale)
359  bits= qp2bits(rce, rco[i].qscale); //FIXME move at end to really force it?
360  else
361  bits*= rco[i].quality_factor;
362  }
363 
364  q= bits2qp(rce, bits);
365 
366  /* I/B difference */
367  if (pict_type==AV_PICTURE_TYPE_I && s->avctx->i_quant_factor<0.0)
368  q= -q*s->avctx->i_quant_factor + s->avctx->i_quant_offset;
369  else if(pict_type==AV_PICTURE_TYPE_B && s->avctx->b_quant_factor<0.0)
370  q= -q*s->avctx->b_quant_factor + s->avctx->b_quant_offset;
371  if(q<1) q=1;
372 
373  return q;
374 }
375 
376 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q){
377  RateControlContext *rcc= &s->rc_context;
378  AVCodecContext *a= s->avctx;
379  const int pict_type= rce->new_pict_type;
380  const double last_p_q = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
381  const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
382 
383  if (pict_type==AV_PICTURE_TYPE_I && (a->i_quant_factor>0.0 || rcc->last_non_b_pict_type==AV_PICTURE_TYPE_P))
384  q= last_p_q *FFABS(a->i_quant_factor) + a->i_quant_offset;
385  else if(pict_type==AV_PICTURE_TYPE_B && a->b_quant_factor>0.0)
386  q= last_non_b_q* a->b_quant_factor + a->b_quant_offset;
387  if(q<1) q=1;
388 
389  /* last qscale / qdiff stuff */
390  if(rcc->last_non_b_pict_type==pict_type || pict_type!=AV_PICTURE_TYPE_I){
391  double last_q= rcc->last_qscale_for[pict_type];
392  const int maxdiff= FF_QP2LAMBDA * a->max_qdiff;
393 
394  if (q > last_q + maxdiff) q= last_q + maxdiff;
395  else if(q < last_q - maxdiff) q= last_q - maxdiff;
396  }
397 
398  rcc->last_qscale_for[pict_type]= q; //Note we cannot do that after blurring
399 
400  if(pict_type!=AV_PICTURE_TYPE_B)
401  rcc->last_non_b_pict_type= pict_type;
402 
403  return q;
404 }
405 
409 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type){
410  int qmin= s->avctx->lmin;
411  int qmax= s->avctx->lmax;
412 
413  assert(qmin <= qmax);
414 
415  if(pict_type==AV_PICTURE_TYPE_B){
416  qmin= (int)(qmin*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
417  qmax= (int)(qmax*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
418  }else if(pict_type==AV_PICTURE_TYPE_I){
419  qmin= (int)(qmin*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
420  qmax= (int)(qmax*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
421  }
422 
423  qmin= av_clip(qmin, 1, FF_LAMBDA_MAX);
424  qmax= av_clip(qmax, 1, FF_LAMBDA_MAX);
425 
426  if(qmax<qmin) qmax= qmin;
427 
428  *qmin_ret= qmin;
429  *qmax_ret= qmax;
430 }
431 
432 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce, double q, int frame_num){
433  RateControlContext *rcc= &s->rc_context;
434  int qmin, qmax;
435  const int pict_type= rce->new_pict_type;
436  const double buffer_size= s->avctx->rc_buffer_size;
437  const double fps= 1/av_q2d(s->avctx->time_base);
438  const double min_rate= s->avctx->rc_min_rate / fps;
439  const double max_rate= s->avctx->rc_max_rate / fps;
440 
441  get_qminmax(&qmin, &qmax, s, pict_type);
442 
443  /* modulation */
444  if(s->avctx->rc_qmod_freq && frame_num%s->avctx->rc_qmod_freq==0 && pict_type==AV_PICTURE_TYPE_P)
445  q*= s->avctx->rc_qmod_amp;
446 
447 //printf("q:%f\n", q);
448  /* buffer overflow/underflow protection */
449  if(buffer_size){
450  double expected_size= rcc->buffer_index;
451  double q_limit;
452 
453  if(min_rate){
454  double d= 2*(buffer_size - expected_size)/buffer_size;
455  if(d>1.0) d=1.0;
456  else if(d<0.0001) d=0.0001;
457  q*= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
458 
459  q_limit= bits2qp(rce, FFMAX((min_rate - buffer_size + rcc->buffer_index) * s->avctx->rc_min_vbv_overflow_use, 1));
460  if(q > q_limit){
461  if(s->avctx->debug&FF_DEBUG_RC){
462  av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
463  }
464  q= q_limit;
465  }
466  }
467 
468  if(max_rate){
469  double d= 2*expected_size/buffer_size;
470  if(d>1.0) d=1.0;
471  else if(d<0.0001) d=0.0001;
472  q/= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
473 
474  q_limit= bits2qp(rce, FFMAX(rcc->buffer_index * s->avctx->rc_max_available_vbv_use, 1));
475  if(q < q_limit){
476  if(s->avctx->debug&FF_DEBUG_RC){
477  av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
478  }
479  q= q_limit;
480  }
481  }
482  }
483 //printf("q:%f max:%f min:%f size:%f index:%d bits:%f agr:%f\n", q,max_rate, min_rate, buffer_size, rcc->buffer_index, bits, s->avctx->rc_buffer_aggressivity);
484  if(s->avctx->rc_qsquish==0.0 || qmin==qmax){
485  if (q<qmin) q=qmin;
486  else if(q>qmax) q=qmax;
487  }else{
488  double min2= log(qmin);
489  double max2= log(qmax);
490 
491  q= log(q);
492  q= (q - min2)/(max2-min2) - 0.5;
493  q*= -4.0;
494  q= 1.0/(1.0 + exp(q));
495  q= q*(max2-min2) + min2;
496 
497  q= exp(q);
498  }
499 
500  return q;
501 }
502 
503 //----------------------------------
504 // 1 Pass Code
505 
506 static double predict_size(Predictor *p, double q, double var)
507 {
508  return p->coeff*var / (q*p->count);
509 }
510 
511 /*
512 static double predict_qp(Predictor *p, double size, double var)
513 {
514 //printf("coeff:%f, count:%f, var:%f, size:%f//\n", p->coeff, p->count, var, size);
515  return p->coeff*var / (size*p->count);
516 }
517 */
518 
519 static void update_predictor(Predictor *p, double q, double var, double size)
520 {
521  double new_coeff= size*q / (var + 1);
522  if(var<10) return;
523 
524  p->count*= p->decay;
525  p->coeff*= p->decay;
526  p->count++;
527  p->coeff+= new_coeff;
528 }
529 
530 static void adaptive_quantization(MpegEncContext *s, double q){
531  int i;
532  const float lumi_masking= s->avctx->lumi_masking / (128.0*128.0);
533  const float dark_masking= s->avctx->dark_masking / (128.0*128.0);
534  const float temp_cplx_masking= s->avctx->temporal_cplx_masking;
535  const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
536  const float p_masking = s->avctx->p_masking;
537  const float border_masking = s->avctx->border_masking;
538  float bits_sum= 0.0;
539  float cplx_sum= 0.0;
540  float cplx_tab[s->mb_num];
541  float bits_tab[s->mb_num];
542  const int qmin= s->avctx->mb_lmin;
543  const int qmax= s->avctx->mb_lmax;
544  Picture * const pic= &s->current_picture;
545  const int mb_width = s->mb_width;
546  const int mb_height = s->mb_height;
547 
548  for(i=0; i<s->mb_num; i++){
549  const int mb_xy= s->mb_index2xy[i];
550  float temp_cplx= sqrt(pic->mc_mb_var[mb_xy]); //FIXME merge in pow()
551  float spat_cplx= sqrt(pic->mb_var[mb_xy]);
552  const int lumi= pic->mb_mean[mb_xy];
553  float bits, cplx, factor;
554  int mb_x = mb_xy % s->mb_stride;
555  int mb_y = mb_xy / s->mb_stride;
556  int mb_distance;
557  float mb_factor = 0.0;
558 #if 0
559  if(spat_cplx < q/3) spat_cplx= q/3; //FIXME finetune
560  if(temp_cplx < q/3) temp_cplx= q/3; //FIXME finetune
561 #endif
562  if(spat_cplx < 4) spat_cplx= 4; //FIXME finetune
563  if(temp_cplx < 4) temp_cplx= 4; //FIXME finetune
564 
565  if((s->mb_type[mb_xy]&CANDIDATE_MB_TYPE_INTRA)){//FIXME hq mode
566  cplx= spat_cplx;
567  factor= 1.0 + p_masking;
568  }else{
569  cplx= temp_cplx;
570  factor= pow(temp_cplx, - temp_cplx_masking);
571  }
572  factor*=pow(spat_cplx, - spatial_cplx_masking);
573 
574  if(lumi>127)
575  factor*= (1.0 - (lumi-128)*(lumi-128)*lumi_masking);
576  else
577  factor*= (1.0 - (lumi-128)*(lumi-128)*dark_masking);
578 
579  if(mb_x < mb_width/5){
580  mb_distance = mb_width/5 - mb_x;
581  mb_factor = (float)mb_distance / (float)(mb_width/5);
582  }else if(mb_x > 4*mb_width/5){
583  mb_distance = mb_x - 4*mb_width/5;
584  mb_factor = (float)mb_distance / (float)(mb_width/5);
585  }
586  if(mb_y < mb_height/5){
587  mb_distance = mb_height/5 - mb_y;
588  mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
589  }else if(mb_y > 4*mb_height/5){
590  mb_distance = mb_y - 4*mb_height/5;
591  mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
592  }
593 
594  factor*= 1.0 - border_masking*mb_factor;
595 
596  if(factor<0.00001) factor= 0.00001;
597 
598  bits= cplx*factor;
599  cplx_sum+= cplx;
600  bits_sum+= bits;
601  cplx_tab[i]= cplx;
602  bits_tab[i]= bits;
603  }
604 
605  /* handle qmin/qmax clipping */
607  float factor= bits_sum/cplx_sum;
608  for(i=0; i<s->mb_num; i++){
609  float newq= q*cplx_tab[i]/bits_tab[i];
610  newq*= factor;
611 
612  if (newq > qmax){
613  bits_sum -= bits_tab[i];
614  cplx_sum -= cplx_tab[i]*q/qmax;
615  }
616  else if(newq < qmin){
617  bits_sum -= bits_tab[i];
618  cplx_sum -= cplx_tab[i]*q/qmin;
619  }
620  }
621  if(bits_sum < 0.001) bits_sum= 0.001;
622  if(cplx_sum < 0.001) cplx_sum= 0.001;
623  }
624 
625  for(i=0; i<s->mb_num; i++){
626  const int mb_xy= s->mb_index2xy[i];
627  float newq= q*cplx_tab[i]/bits_tab[i];
628  int intq;
629 
631  newq*= bits_sum/cplx_sum;
632  }
633 
634  intq= (int)(newq + 0.5);
635 
636  if (intq > qmax) intq= qmax;
637  else if(intq < qmin) intq= qmin;
638 //if(i%s->mb_width==0) printf("\n");
639 //printf("%2d%3d ", intq, ff_sqrt(s->mc_mb_var[i]));
640  s->lambda_table[mb_xy]= intq;
641  }
642 }
643 
645  RateControlContext *rcc= &s->rc_context;
646  int picture_number= s->picture_number;
647  RateControlEntry *rce;
648 
649  rce= &rcc->entry[picture_number];
650  s->f_code= rce->f_code;
651  s->b_code= rce->b_code;
652 }
653 
654 //FIXME rd or at least approx for dquant
655 
657 {
658  float q;
659  int qmin, qmax;
660  float br_compensation;
661  double diff;
662  double short_term_q;
663  double fps;
664  int picture_number= s->picture_number;
665  int64_t wanted_bits;
666  RateControlContext *rcc= &s->rc_context;
667  AVCodecContext *a= s->avctx;
668  RateControlEntry local_rce, *rce;
669  double bits;
670  double rate_factor;
671  int var;
672  const int pict_type= s->pict_type;
673  Picture * const pic= &s->current_picture;
674  emms_c();
675 
676 #if CONFIG_LIBXVID
678  return ff_xvid_rate_estimate_qscale(s, dry_run);
679 #endif
680 
681  get_qminmax(&qmin, &qmax, s, pict_type);
682 
683  fps= 1/av_q2d(s->avctx->time_base);
684 //printf("input_pic_num:%d pic_num:%d frame_rate:%d\n", s->input_picture_number, s->picture_number, s->frame_rate);
685  /* update predictors */
686  if(picture_number>2 && !dry_run){
687  const int last_var= s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum;
688  update_predictor(&rcc->pred[s->last_pict_type], rcc->last_qscale, sqrt(last_var), s->frame_bits);
689  }
690 
691  if(s->flags&CODEC_FLAG_PASS2){
692  assert(picture_number>=0);
693  assert(picture_number<rcc->num_entries);
694  rce= &rcc->entry[picture_number];
695  wanted_bits= rce->expected_bits;
696  }else{
697  Picture *dts_pic;
698  rce= &local_rce;
699 
700  //FIXME add a dts field to AVFrame and ensure its set and use it here instead of reordering
701  //but the reordering is simpler for now until h.264 b pyramid must be handeld
702  if(s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
703  dts_pic= s->current_picture_ptr;
704  else
705  dts_pic= s->last_picture_ptr;
706 
707 //if(dts_pic)
708 // av_log(NULL, AV_LOG_ERROR, "%Ld %Ld %Ld %d\n", s->current_picture_ptr->pts, s->user_specified_pts, dts_pic->pts, picture_number);
709 
710  if (!dts_pic || dts_pic->f.pts == AV_NOPTS_VALUE)
711  wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps);
712  else
713  wanted_bits = (uint64_t)(s->bit_rate*(double)dts_pic->f.pts / fps);
714  }
715 
716  diff= s->total_bits - wanted_bits;
717  br_compensation= (a->bit_rate_tolerance - diff)/a->bit_rate_tolerance;
718  if(br_compensation<=0.0) br_compensation=0.001;
719 
720  var= pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
721 
722  short_term_q = 0; /* avoid warning */
723  if(s->flags&CODEC_FLAG_PASS2){
724  if(pict_type!=AV_PICTURE_TYPE_I)
725  assert(pict_type == rce->new_pict_type);
726 
727  q= rce->new_qscale / br_compensation;
728 //printf("%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale, br_compensation, s->frame_bits, var, pict_type);
729  }else{
730  rce->pict_type=
731  rce->new_pict_type= pict_type;
732  rce->mc_mb_var_sum= pic->mc_mb_var_sum;
733  rce->mb_var_sum = pic-> mb_var_sum;
734  rce->qscale = FF_QP2LAMBDA * 2;
735  rce->f_code = s->f_code;
736  rce->b_code = s->b_code;
737  rce->misc_bits= 1;
738 
739  bits= predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
740  if(pict_type== AV_PICTURE_TYPE_I){
741  rce->i_count = s->mb_num;
742  rce->i_tex_bits= bits;
743  rce->p_tex_bits= 0;
744  rce->mv_bits= 0;
745  }else{
746  rce->i_count = 0; //FIXME we do know this approx
747  rce->i_tex_bits= 0;
748  rce->p_tex_bits= bits*0.9;
749 
750  rce->mv_bits= bits*0.1;
751  }
752  rcc->i_cplx_sum [pict_type] += rce->i_tex_bits*rce->qscale;
753  rcc->p_cplx_sum [pict_type] += rce->p_tex_bits*rce->qscale;
754  rcc->mv_bits_sum[pict_type] += rce->mv_bits;
755  rcc->frame_count[pict_type] ++;
756 
757  bits= rce->i_tex_bits + rce->p_tex_bits;
758  rate_factor= rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum * br_compensation;
759 
760  q= get_qscale(s, rce, rate_factor, picture_number);
761  if (q < 0)
762  return -1;
763 
764  assert(q>0.0);
765 //printf("%f ", q);
766  q= get_diff_limited_q(s, rce, q);
767 //printf("%f ", q);
768  assert(q>0.0);
769 
770  if(pict_type==AV_PICTURE_TYPE_P || s->intra_only){ //FIXME type dependent blur like in 2-pass
771  rcc->short_term_qsum*=a->qblur;
772  rcc->short_term_qcount*=a->qblur;
773 
774  rcc->short_term_qsum+= q;
775  rcc->short_term_qcount++;
776 //printf("%f ", q);
777  q= short_term_q= rcc->short_term_qsum/rcc->short_term_qcount;
778 //printf("%f ", q);
779  }
780  assert(q>0.0);
781 
782  q= modify_qscale(s, rce, q, picture_number);
783 
784  rcc->pass1_wanted_bits+= s->bit_rate/fps;
785 
786  assert(q>0.0);
787  }
788 
789  if(s->avctx->debug&FF_DEBUG_RC){
790  av_log(s->avctx, AV_LOG_DEBUG, "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f size:%d var:%d/%d br:%d fps:%d\n",
791  av_get_picture_type_char(pict_type), qmin, q, qmax, picture_number, (int)wanted_bits/1000, (int)s->total_bits/1000,
792  br_compensation, short_term_q, s->frame_bits, pic->mb_var_sum, pic->mc_mb_var_sum, s->bit_rate/1000, (int)fps
793  );
794  }
795 
796  if (q<qmin) q=qmin;
797  else if(q>qmax) q=qmax;
798 
799  if(s->adaptive_quant)
800  adaptive_quantization(s, q);
801  else
802  q= (int)(q + 0.5);
803 
804  if(!dry_run){
805  rcc->last_qscale= q;
807  rcc->last_mb_var_sum= pic->mb_var_sum;
808  }
809  return q;
810 }
811 
812 //----------------------------------------------
813 // 2-Pass code
814 
816 {
817  RateControlContext *rcc= &s->rc_context;
818  AVCodecContext *a= s->avctx;
819  int i, toobig;
820  double fps= 1/av_q2d(s->avctx->time_base);
821  double complexity[5]={0,0,0,0,0}; // aproximate bits at quant=1
822  uint64_t const_bits[5]={0,0,0,0,0}; // quantizer independent bits
823  uint64_t all_const_bits;
824  uint64_t all_available_bits= (uint64_t)(s->bit_rate*(double)rcc->num_entries/fps);
825  double rate_factor=0;
826  double step;
827  //int last_i_frame=-10000000;
828  const int filter_size= (int)(a->qblur*4) | 1;
829  double expected_bits;
830  double *qscale, *blurred_qscale, qscale_sum;
831 
832  /* find complexity & const_bits & decide the pict_types */
833  for(i=0; i<rcc->num_entries; i++){
834  RateControlEntry *rce= &rcc->entry[i];
835 
836  rce->new_pict_type= rce->pict_type;
837  rcc->i_cplx_sum [rce->pict_type] += rce->i_tex_bits*rce->qscale;
838  rcc->p_cplx_sum [rce->pict_type] += rce->p_tex_bits*rce->qscale;
839  rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
840  rcc->frame_count[rce->pict_type] ++;
841 
842  complexity[rce->new_pict_type]+= (rce->i_tex_bits+ rce->p_tex_bits)*(double)rce->qscale;
843  const_bits[rce->new_pict_type]+= rce->mv_bits + rce->misc_bits;
844  }
845  all_const_bits= const_bits[AV_PICTURE_TYPE_I] + const_bits[AV_PICTURE_TYPE_P] + const_bits[AV_PICTURE_TYPE_B];
846 
847  if(all_available_bits < all_const_bits){
848  av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
849  return -1;
850  }
851 
852  qscale= av_malloc(sizeof(double)*rcc->num_entries);
853  blurred_qscale= av_malloc(sizeof(double)*rcc->num_entries);
854  toobig = 0;
855 
856  for(step=256*256; step>0.0000001; step*=0.5){
857  expected_bits=0;
858  rate_factor+= step;
859 
860  rcc->buffer_index= s->avctx->rc_buffer_size/2;
861 
862  /* find qscale */
863  for(i=0; i<rcc->num_entries; i++){
864  RateControlEntry *rce= &rcc->entry[i];
865  qscale[i]= get_qscale(s, &rcc->entry[i], rate_factor, i);
866  rcc->last_qscale_for[rce->pict_type] = qscale[i];
867  }
868  assert(filter_size%2==1);
869 
870  /* fixed I/B QP relative to P mode */
871  for(i=rcc->num_entries-1; i>=0; i--){
872  RateControlEntry *rce= &rcc->entry[i];
873 
874  qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
875  }
876 
877  /* smooth curve */
878  for(i=0; i<rcc->num_entries; i++){
879  RateControlEntry *rce= &rcc->entry[i];
880  const int pict_type= rce->new_pict_type;
881  int j;
882  double q=0.0, sum=0.0;
883 
884  for(j=0; j<filter_size; j++){
885  int index= i+j-filter_size/2;
886  double d= index-i;
887  double coeff= a->qblur==0 ? 1.0 : exp(-d*d/(a->qblur * a->qblur));
888 
889  if(index < 0 || index >= rcc->num_entries) continue;
890  if(pict_type != rcc->entry[index].new_pict_type) continue;
891  q+= qscale[index] * coeff;
892  sum+= coeff;
893  }
894  blurred_qscale[i]= q/sum;
895  }
896 
897  /* find expected bits */
898  for(i=0; i<rcc->num_entries; i++){
899  RateControlEntry *rce= &rcc->entry[i];
900  double bits;
901  rce->new_qscale= modify_qscale(s, rce, blurred_qscale[i], i);
902  bits= qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
903 //printf("%d %f\n", rce->new_bits, blurred_qscale[i]);
904  bits += 8*ff_vbv_update(s, bits);
905 
906  rce->expected_bits= expected_bits;
907  expected_bits += bits;
908  }
909 
910  /*
911  av_log(s->avctx, AV_LOG_INFO,
912  "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
913  expected_bits, (int)all_available_bits, rate_factor);
914  */
915  if(expected_bits > all_available_bits) {
916  rate_factor-= step;
917  ++toobig;
918  }
919  }
920  av_free(qscale);
921  av_free(blurred_qscale);
922 
923  /* check bitrate calculations and print info */
924  qscale_sum = 0.0;
925  for(i=0; i<rcc->num_entries; i++){
926  /* av_log(s->avctx, AV_LOG_DEBUG, "[lavc rc] entry[%d].new_qscale = %.3f qp = %.3f\n",
927  i, rcc->entry[i].new_qscale, rcc->entry[i].new_qscale / FF_QP2LAMBDA); */
928  qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA, s->avctx->qmin, s->avctx->qmax);
929  }
930  assert(toobig <= 40);
932  "[lavc rc] requested bitrate: %d bps expected bitrate: %d bps\n",
933  s->bit_rate,
934  (int)(expected_bits / ((double)all_available_bits/s->bit_rate)));
936  "[lavc rc] estimated target average qp: %.3f\n",
937  (float)qscale_sum / rcc->num_entries);
938  if (toobig == 0) {
940  "[lavc rc] Using all of requested bitrate is not "
941  "necessary for this video with these parameters.\n");
942  } else if (toobig == 40) {
944  "[lavc rc] Error: bitrate too low for this video "
945  "with these parameters.\n");
946  return -1;
947  } else if (fabs(expected_bits/all_available_bits - 1.0) > 0.01) {
949  "[lavc rc] Error: 2pass curve failed to converge\n");
950  return -1;
951  }
952 
953  return 0;
954 }