per file doxy
[ffmpeg.git] / libavcodec / ratecontrol.c
1 /*
2  * Rate control for video encoders
3  *
4  * Copyright (c) 2002 Michael Niedermayer <michaelni@gmx.at>
5  *
6  * This library is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2 of the License, or (at your option) any later version.
10  *
11  * This library is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with this library; if not, write to the Free Software
18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
19  */
20
21 /**
22  * @file ratecontrol.c
23  * Rate control for video encoders.
24  */ 
25
26 #include "avcodec.h"
27 #include "dsputil.h"
28 #include "mpegvideo.h"
29
30 #undef NDEBUG // allways check asserts, the speed effect is far too small to disable them
31 #include <assert.h>
32
33 #ifndef M_E
34 #define M_E 2.718281828
35 #endif
36
37 static int init_pass2(MpegEncContext *s);
38 static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num);
39
40 void ff_write_pass1_stats(MpegEncContext *s){
41     sprintf(s->avctx->stats_out, "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;\n",
42             s->picture_number, s->input_picture_number - s->max_b_frames, s->pict_type, 
43             s->frame_qscale, s->i_tex_bits, s->p_tex_bits, s->mv_bits, s->misc_bits, 
44             s->f_code, s->b_code, s->current_picture.mc_mb_var_sum, s->current_picture.mb_var_sum, s->i_count);
45 }
46
47 int ff_rate_control_init(MpegEncContext *s)
48 {
49     RateControlContext *rcc= &s->rc_context;
50     int i;
51     emms_c();
52
53     for(i=0; i<5; i++){
54         rcc->pred[i].coeff= 7.0;
55         rcc->pred[i].count= 1.0;
56     
57         rcc->pred[i].decay= 0.4;
58         rcc->i_cplx_sum [i]=
59         rcc->p_cplx_sum [i]=
60         rcc->mv_bits_sum[i]=
61         rcc->qscale_sum [i]=
62         rcc->frame_count[i]= 1; // 1 is better cuz of 1/0 and such
63         rcc->last_qscale_for[i]=5;
64     }
65     rcc->buffer_index= s->avctx->rc_buffer_size/2;
66
67     if(s->flags&CODEC_FLAG_PASS2){
68         int i;
69         char *p;
70
71         /* find number of pics */
72         p= s->avctx->stats_in;
73         for(i=-1; p; i++){
74             p= strchr(p+1, ';');
75         }
76         i+= s->max_b_frames;
77         rcc->entry = (RateControlEntry*)av_mallocz(i*sizeof(RateControlEntry));
78         rcc->num_entries= i;
79         
80         /* init all to skiped p frames (with b frames we might have a not encoded frame at the end FIXME) */
81         for(i=0; i<rcc->num_entries; i++){
82             RateControlEntry *rce= &rcc->entry[i];
83             rce->pict_type= rce->new_pict_type=P_TYPE;
84             rce->qscale= rce->new_qscale=2;
85             rce->misc_bits= s->mb_num + 10;
86             rce->mb_var_sum= s->mb_num*100;
87         }        
88         
89         /* read stats */
90         p= s->avctx->stats_in;
91         for(i=0; i<rcc->num_entries - s->max_b_frames; i++){
92             RateControlEntry *rce;
93             int picture_number;
94             int e;
95             char *next;
96
97             next= strchr(p, ';');
98             if(next){
99                 (*next)=0; //sscanf in unbelieavle slow on looong strings //FIXME copy / dont write
100                 next++;
101             }
102             e= sscanf(p, " in:%d ", &picture_number);
103
104             assert(picture_number >= 0);
105             assert(picture_number < rcc->num_entries);
106             rce= &rcc->entry[picture_number];
107
108             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",
109                    &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits, &rce->mv_bits, &rce->misc_bits, 
110                    &rce->f_code, &rce->b_code, &rce->mc_mb_var_sum, &rce->mb_var_sum, &rce->i_count);
111             if(e!=12){
112                 fprintf(stderr, "statistics are damaged at line %d, parser out=%d\n", i, e);
113                 return -1;
114             }
115             p= next;
116         }
117         
118         if(init_pass2(s) < 0) return -1;
119     }
120      
121     if(!(s->flags&CODEC_FLAG_PASS2)){
122
123         rcc->short_term_qsum=0.001;
124         rcc->short_term_qcount=0.001;
125     
126         rcc->pass1_rc_eq_output_sum= 0.001;
127         rcc->pass1_wanted_bits=0.001;
128         
129         /* init stuff with the user specified complexity */
130         if(s->avctx->rc_initial_cplx){
131             for(i=0; i<60*30; i++){
132                 double bits= s->avctx->rc_initial_cplx * (i/10000.0 + 1.0)*s->mb_num;
133                 RateControlEntry rce;
134                 double q;
135                 
136                 if     (i%((s->gop_size+3)/4)==0) rce.pict_type= I_TYPE;
137                 else if(i%(s->max_b_frames+1))    rce.pict_type= B_TYPE;
138                 else                              rce.pict_type= P_TYPE;
139
140                 rce.new_pict_type= rce.pict_type;
141                 rce.mc_mb_var_sum= bits*s->mb_num/100000;
142                 rce.mb_var_sum   = s->mb_num;
143                 rce.qscale   = 2;
144                 rce.f_code   = 2;
145                 rce.b_code   = 1;
146                 rce.misc_bits= 1;
147
148                 if(s->pict_type== I_TYPE){
149                     rce.i_count   = s->mb_num;
150                     rce.i_tex_bits= bits;
151                     rce.p_tex_bits= 0;
152                     rce.mv_bits= 0;
153                 }else{
154                     rce.i_count   = 0; //FIXME we do know this approx
155                     rce.i_tex_bits= 0;
156                     rce.p_tex_bits= bits*0.9;
157                     rce.mv_bits= bits*0.1;
158                 }
159                 rcc->i_cplx_sum [rce.pict_type] += rce.i_tex_bits*rce.qscale;
160                 rcc->p_cplx_sum [rce.pict_type] += rce.p_tex_bits*rce.qscale;
161                 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
162                 rcc->frame_count[rce.pict_type] ++;
163
164                 bits= rce.i_tex_bits + rce.p_tex_bits;
165
166                 q= get_qscale(s, &rce, rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum, i);
167                 rcc->pass1_wanted_bits+= s->bit_rate/(s->frame_rate / (double)FRAME_RATE_BASE);
168             }
169         }
170
171     }
172     
173     return 0;
174 }
175
176 void ff_rate_control_uninit(MpegEncContext *s)
177 {
178     RateControlContext *rcc= &s->rc_context;
179     emms_c();
180
181     av_freep(&rcc->entry);
182 }
183
184 static inline double qp2bits(RateControlEntry *rce, double qp){
185     if(qp<=0.0){
186         fprintf(stderr, "qp<=0.0\n");
187     }
188     return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ qp;
189 }
190
191 static inline double bits2qp(RateControlEntry *rce, double bits){
192     if(bits<0.9){
193         fprintf(stderr, "bits<0.9\n");
194     }
195     return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ bits;
196 }
197     
198 static void update_rc_buffer(MpegEncContext *s, int frame_size){
199     RateControlContext *rcc= &s->rc_context;
200     const double fps= (double)s->frame_rate / FRAME_RATE_BASE;
201     const double buffer_size= s->avctx->rc_buffer_size;
202     const double min_rate= s->avctx->rc_min_rate/fps;
203     const double max_rate= s->avctx->rc_max_rate/fps;
204
205     if(buffer_size){
206         rcc->buffer_index-= frame_size;
207         if(rcc->buffer_index < buffer_size/2 /*FIXME /2 */ || min_rate==0){
208             rcc->buffer_index+= max_rate;
209             if(rcc->buffer_index >= buffer_size)
210                 rcc->buffer_index= buffer_size-1;
211         }else{
212             rcc->buffer_index+= min_rate;
213         }
214         
215         if(rcc->buffer_index < 0)
216             fprintf(stderr, "rc buffer underflow\n");
217         if(rcc->buffer_index >= s->avctx->rc_buffer_size)
218             fprintf(stderr, "rc buffer overflow\n");
219     }
220 }
221
222 /**
223  * modifies the bitrate curve from pass1 for one frame
224  */
225 static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num){
226     RateControlContext *rcc= &s->rc_context;
227     double q, bits;
228     const int pict_type= rce->new_pict_type;
229     const double mb_num= s->mb_num;  
230     int i;
231
232     double const_values[]={
233         M_PI,
234         M_E,
235         rce->i_tex_bits*rce->qscale,
236         rce->p_tex_bits*rce->qscale,
237         (rce->i_tex_bits + rce->p_tex_bits)*(double)rce->qscale,
238         rce->mv_bits/mb_num,
239         rce->pict_type == B_TYPE ? (rce->f_code + rce->b_code)*0.5 : rce->f_code,
240         rce->i_count/mb_num,
241         rce->mc_mb_var_sum/mb_num,
242         rce->mb_var_sum/mb_num,
243         rce->pict_type == I_TYPE,
244         rce->pict_type == P_TYPE,
245         rce->pict_type == B_TYPE,
246         rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
247         s->qcompress,
248 /*        rcc->last_qscale_for[I_TYPE],
249         rcc->last_qscale_for[P_TYPE],
250         rcc->last_qscale_for[B_TYPE],
251         rcc->next_non_b_qscale,*/
252         rcc->i_cplx_sum[I_TYPE] / (double)rcc->frame_count[I_TYPE],
253         rcc->i_cplx_sum[P_TYPE] / (double)rcc->frame_count[P_TYPE],
254         rcc->p_cplx_sum[P_TYPE] / (double)rcc->frame_count[P_TYPE],
255         rcc->p_cplx_sum[B_TYPE] / (double)rcc->frame_count[B_TYPE],
256         (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
257         0
258     };
259     static const char *const_names[]={
260         "PI",
261         "E",
262         "iTex",
263         "pTex",
264         "tex",
265         "mv",
266         "fCode",
267         "iCount",
268         "mcVar",
269         "var",
270         "isI",
271         "isP",
272         "isB",
273         "avgQP",
274         "qComp",
275 /*        "lastIQP",
276         "lastPQP",
277         "lastBQP",
278         "nextNonBQP",*/
279         "avgIITex",
280         "avgPITex",
281         "avgPPTex",
282         "avgBPTex",
283         "avgTex",
284         NULL
285     };
286     static double (*func1[])(void *, double)={
287         (void *)bits2qp,
288         (void *)qp2bits,
289         NULL
290     };
291     static const char *func1_names[]={
292         "bits2qp",
293         "qp2bits",
294         NULL
295     };
296
297     bits= ff_eval(s->avctx->rc_eq, const_values, const_names, func1, func1_names, NULL, NULL, rce);
298     
299     rcc->pass1_rc_eq_output_sum+= bits;
300     bits*=rate_factor;
301     if(bits<0.0) bits=0.0;
302     bits+= 1.0; //avoid 1/0 issues
303     
304     /* user override */
305     for(i=0; i<s->avctx->rc_override_count; i++){
306         RcOverride *rco= s->avctx->rc_override;
307         if(rco[i].start_frame > frame_num) continue;
308         if(rco[i].end_frame   < frame_num) continue;
309     
310         if(rco[i].qscale) 
311             bits= qp2bits(rce, rco[i].qscale); //FIXME move at end to really force it?
312         else
313             bits*= rco[i].quality_factor;
314     }
315
316     q= bits2qp(rce, bits);
317     
318     /* I/B difference */
319     if     (pict_type==I_TYPE && s->avctx->i_quant_factor<0.0)
320         q= -q*s->avctx->i_quant_factor + s->avctx->i_quant_offset;
321     else if(pict_type==B_TYPE && s->avctx->b_quant_factor<0.0)
322         q= -q*s->avctx->b_quant_factor + s->avctx->b_quant_offset;
323         
324     return q;
325 }
326
327 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q){
328     RateControlContext *rcc= &s->rc_context;
329     AVCodecContext *a= s->avctx;
330     const int pict_type= rce->new_pict_type;
331     const double last_p_q    = rcc->last_qscale_for[P_TYPE];
332     const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
333     
334     if     (pict_type==I_TYPE && (a->i_quant_factor>0.0 || rcc->last_non_b_pict_type==P_TYPE))
335         q= last_p_q    *ABS(a->i_quant_factor) + a->i_quant_offset;
336     else if(pict_type==B_TYPE && a->b_quant_factor>0.0)
337         q= last_non_b_q*    a->b_quant_factor  + a->b_quant_offset;
338
339     /* last qscale / qdiff stuff */
340     if(rcc->last_non_b_pict_type==pict_type || pict_type!=I_TYPE){
341         double last_q= rcc->last_qscale_for[pict_type];
342
343         if     (q > last_q + a->max_qdiff) q= last_q + a->max_qdiff;
344         else if(q < last_q - a->max_qdiff) q= last_q - a->max_qdiff;
345     }
346
347     rcc->last_qscale_for[pict_type]= q; //Note we cant do that after blurring
348     
349     if(pict_type!=B_TYPE)
350         rcc->last_non_b_pict_type= pict_type;
351
352     return q;
353 }
354
355 /**
356  * gets the qmin & qmax for pict_type
357  */
358 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type){
359     int qmin= s->qmin;                                                       
360     int qmax= s->qmax;
361
362     if(pict_type==B_TYPE){
363         qmin= (int)(qmin*ABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
364         qmax= (int)(qmax*ABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
365     }else if(pict_type==I_TYPE){
366         qmin= (int)(qmin*ABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
367         qmax= (int)(qmax*ABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
368     }
369
370     if(qmin<1) qmin=1;
371     if(qmin==1 && s->qmin>1) qmin=2; //avoid qmin=1 unless the user wants qmin=1
372
373     if(qmin<3 && s->max_qcoeff<=128 && pict_type==I_TYPE) qmin=3; //reduce cliping problems
374
375     if(qmax>31) qmax=31;
376     if(qmax<=qmin) qmax= qmin= (qmax+qmin+1)>>1;
377     
378     *qmin_ret= qmin;
379     *qmax_ret= qmax;
380 }
381
382 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce, double q, int frame_num){
383     RateControlContext *rcc= &s->rc_context;
384     int qmin, qmax;
385     double bits;
386     const int pict_type= rce->new_pict_type;
387     const double buffer_size= s->avctx->rc_buffer_size;
388     const double min_rate= s->avctx->rc_min_rate;
389     const double max_rate= s->avctx->rc_max_rate;
390     
391     get_qminmax(&qmin, &qmax, s, pict_type);
392
393     /* modulation */
394     if(s->avctx->rc_qmod_freq && frame_num%s->avctx->rc_qmod_freq==0 && pict_type==P_TYPE)
395         q*= s->avctx->rc_qmod_amp;
396
397     bits= qp2bits(rce, q);
398 //printf("q:%f\n", q);
399     /* buffer overflow/underflow protection */
400     if(buffer_size){
401         double expected_size= rcc->buffer_index;
402
403         if(min_rate){
404             double d= 2*(buffer_size - expected_size)/buffer_size;
405             if(d>1.0) d=1.0;
406             else if(d<0.0001) d=0.0001;
407             q*= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
408
409             q= FFMIN(q, bits2qp(rce, FFMAX((min_rate - buffer_size + rcc->buffer_index)*2, 1)));
410         }
411
412         if(max_rate){
413             double d= 2*expected_size/buffer_size;
414             if(d>1.0) d=1.0;
415             else if(d<0.0001) d=0.0001;
416             q/= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
417
418             q= FFMAX(q, bits2qp(rce, FFMAX(rcc->buffer_index/2, 1)));
419         }
420     }
421 //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);
422     if(s->avctx->rc_qsquish==0.0 || qmin==qmax){
423         if     (q<qmin) q=qmin;
424         else if(q>qmax) q=qmax;
425     }else{
426         double min2= log(qmin);
427         double max2= log(qmax);
428         
429         q= log(q);
430         q= (q - min2)/(max2-min2) - 0.5;
431         q*= -4.0;
432         q= 1.0/(1.0 + exp(q));
433         q= q*(max2-min2) + min2;
434         
435         q= exp(q);
436     }
437     
438     return q;
439 }
440
441 //----------------------------------
442 // 1 Pass Code
443
444 static double predict_size(Predictor *p, double q, double var)
445 {
446      return p->coeff*var / (q*p->count);
447 }
448
449 /*
450 static double predict_qp(Predictor *p, double size, double var)
451 {
452 //printf("coeff:%f, count:%f, var:%f, size:%f//\n", p->coeff, p->count, var, size);
453      return p->coeff*var / (size*p->count);
454 }
455 */
456
457 static void update_predictor(Predictor *p, double q, double var, double size)
458 {
459     double new_coeff= size*q / (var + 1);
460     if(var<10) return;
461
462     p->count*= p->decay;
463     p->coeff*= p->decay;
464     p->count++;
465     p->coeff+= new_coeff;
466 }
467
468 static void adaptive_quantization(MpegEncContext *s, double q){
469     int i;
470     const float lumi_masking= s->avctx->lumi_masking / (128.0*128.0);
471     const float dark_masking= s->avctx->dark_masking / (128.0*128.0);
472     const float temp_cplx_masking= s->avctx->temporal_cplx_masking;
473     const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
474     const float p_masking = s->avctx->p_masking;
475     float bits_sum= 0.0;
476     float cplx_sum= 0.0;
477     float cplx_tab[s->mb_num];
478     float bits_tab[s->mb_num];
479     const int qmin= s->avctx->mb_qmin;
480     const int qmax= s->avctx->mb_qmax;
481     Picture * const pic= &s->current_picture;
482     
483     for(i=0; i<s->mb_num; i++){
484         float temp_cplx= sqrt(pic->mc_mb_var[i]);
485         float spat_cplx= sqrt(pic->mb_var[i]);
486         const int lumi= pic->mb_mean[i];
487         float bits, cplx, factor;
488         
489         if(spat_cplx < q/3) spat_cplx= q/3; //FIXME finetune
490         if(temp_cplx < q/3) temp_cplx= q/3; //FIXME finetune
491         
492         if((s->mb_type[i]&MB_TYPE_INTRA)){//FIXME hq mode 
493             cplx= spat_cplx;
494             factor= 1.0 + p_masking;
495         }else{
496             cplx= temp_cplx;
497             factor= pow(temp_cplx, - temp_cplx_masking);
498         }
499         factor*=pow(spat_cplx, - spatial_cplx_masking);
500
501         if(lumi>127)
502             factor*= (1.0 - (lumi-128)*(lumi-128)*lumi_masking);
503         else
504             factor*= (1.0 - (lumi-128)*(lumi-128)*dark_masking);
505         
506         if(factor<0.00001) factor= 0.00001;
507         
508         bits= cplx*factor;
509         cplx_sum+= cplx;
510         bits_sum+= bits;
511         cplx_tab[i]= cplx;
512         bits_tab[i]= bits;
513     }
514
515     /* handle qmin/qmax cliping */
516     if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
517         for(i=0; i<s->mb_num; i++){
518             float newq= q*cplx_tab[i]/bits_tab[i];
519             newq*= bits_sum/cplx_sum;
520
521             if     (newq > qmax){
522                 bits_sum -= bits_tab[i];
523                 cplx_sum -= cplx_tab[i]*q/qmax;
524             }
525             else if(newq < qmin){
526                 bits_sum -= bits_tab[i];
527                 cplx_sum -= cplx_tab[i]*q/qmin;
528             }
529         }
530     }
531    
532     for(i=0; i<s->mb_num; i++){
533         float newq= q*cplx_tab[i]/bits_tab[i];
534         int intq;
535
536         if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
537             newq*= bits_sum/cplx_sum;
538         }
539
540         if(i && ABS(pic->qscale_table[i-1] - newq)<0.75)
541             intq= pic->qscale_table[i-1];
542         else
543             intq= (int)(newq + 0.5);
544
545         if     (intq > qmax) intq= qmax;
546         else if(intq < qmin) intq= qmin;
547 //if(i%s->mb_width==0) printf("\n");
548 //printf("%2d%3d ", intq, ff_sqrt(s->mc_mb_var[i]));
549         pic->qscale_table[i]= intq;
550     }
551 }
552
553 float ff_rate_estimate_qscale(MpegEncContext *s)
554 {
555     float q;
556     int qmin, qmax;
557     float br_compensation;
558     double diff;
559     double short_term_q;
560     double fps;
561     int picture_number= s->picture_number;
562     int64_t wanted_bits;
563     RateControlContext *rcc= &s->rc_context;
564     RateControlEntry local_rce, *rce;
565     double bits;
566     double rate_factor;
567     int var;
568     const int pict_type= s->pict_type;
569     Picture * const pic= &s->current_picture;
570     emms_c();
571
572     get_qminmax(&qmin, &qmax, s, pict_type);
573
574     fps= (double)s->frame_rate / FRAME_RATE_BASE;
575 //printf("input_pic_num:%d pic_num:%d frame_rate:%d\n", s->input_picture_number, s->picture_number, s->frame_rate);
576         /* update predictors */
577     if(picture_number>2){
578         const int last_var= s->last_pict_type == I_TYPE ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum;
579         update_predictor(&rcc->pred[s->last_pict_type], rcc->last_qscale, sqrt(last_var), s->frame_bits);
580     }
581
582     if(s->flags&CODEC_FLAG_PASS2){
583         assert(picture_number>=0);
584         assert(picture_number<rcc->num_entries);
585         rce= &rcc->entry[picture_number];
586         wanted_bits= rce->expected_bits;
587     }else{
588         rce= &local_rce;
589         wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps);
590     }
591
592     diff= s->total_bits - wanted_bits;
593     br_compensation= (s->bit_rate_tolerance - diff)/s->bit_rate_tolerance;
594     if(br_compensation<=0.0) br_compensation=0.001;
595
596     var= pict_type == I_TYPE ? pic->mb_var_sum : pic->mc_mb_var_sum;
597     
598     if(s->flags&CODEC_FLAG_PASS2){
599         if(pict_type!=I_TYPE)
600             assert(pict_type == rce->new_pict_type);
601
602         q= rce->new_qscale / br_compensation;
603 //printf("%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale, br_compensation, s->frame_bits, var, pict_type);
604     }else{
605         rce->pict_type= 
606         rce->new_pict_type= pict_type;
607         rce->mc_mb_var_sum= pic->mc_mb_var_sum;
608         rce->mb_var_sum   = pic->   mb_var_sum;
609         rce->qscale   = 2;
610         rce->f_code   = s->f_code;
611         rce->b_code   = s->b_code;
612         rce->misc_bits= 1;
613
614         if(picture_number>0)
615             update_rc_buffer(s, s->frame_bits);
616
617         bits= predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
618         if(pict_type== I_TYPE){
619             rce->i_count   = s->mb_num;
620             rce->i_tex_bits= bits;
621             rce->p_tex_bits= 0;
622             rce->mv_bits= 0;
623         }else{
624             rce->i_count   = 0; //FIXME we do know this approx
625             rce->i_tex_bits= 0;
626             rce->p_tex_bits= bits*0.9;
627             
628             rce->mv_bits= bits*0.1;
629         }
630         rcc->i_cplx_sum [pict_type] += rce->i_tex_bits*rce->qscale;
631         rcc->p_cplx_sum [pict_type] += rce->p_tex_bits*rce->qscale;
632         rcc->mv_bits_sum[pict_type] += rce->mv_bits;
633         rcc->frame_count[pict_type] ++;
634
635         bits= rce->i_tex_bits + rce->p_tex_bits;
636         rate_factor= rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum * br_compensation;
637     
638         q= get_qscale(s, rce, rate_factor, picture_number);
639
640         assert(q>0.0);
641 //printf("%f ", q);
642         q= get_diff_limited_q(s, rce, q);
643 //printf("%f ", q);
644         assert(q>0.0);
645
646         if(pict_type==P_TYPE || s->intra_only){ //FIXME type dependant blur like in 2-pass
647             rcc->short_term_qsum*=s->qblur;
648             rcc->short_term_qcount*=s->qblur;
649
650             rcc->short_term_qsum+= q;
651             rcc->short_term_qcount++;
652 //printf("%f ", q);
653             q= short_term_q= rcc->short_term_qsum/rcc->short_term_qcount;
654 //printf("%f ", q);
655         }
656         assert(q>0.0);
657         
658         q= modify_qscale(s, rce, q, picture_number);
659
660         rcc->pass1_wanted_bits+= s->bit_rate/fps;
661
662         assert(q>0.0);
663     }
664
665     if(s->avctx->debug&FF_DEBUG_RC){
666         printf("%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",
667         ff_get_pict_type_char(pict_type), qmin, q, qmax, picture_number, (int)wanted_bits/1000, (int)s->total_bits/1000,
668         br_compensation, short_term_q, s->frame_bits, pic->mb_var_sum, pic->mc_mb_var_sum, s->bit_rate/1000, (int)fps
669         );
670     }
671
672     if     (q<qmin) q=qmin; 
673     else if(q>qmax) q=qmax;
674
675     if(s->adaptive_quant)
676         adaptive_quantization(s, q);
677     else
678         q= (int)(q + 0.5);
679     
680     rcc->last_qscale= q;
681     rcc->last_mc_mb_var_sum= pic->mc_mb_var_sum;
682     rcc->last_mb_var_sum= pic->mb_var_sum;
683 #if 0
684 {
685     static int mvsum=0, texsum=0;
686     mvsum += s->mv_bits;
687     texsum += s->i_tex_bits + s->p_tex_bits;
688     printf("%d %d//\n\n", mvsum, texsum);
689 }
690 #endif
691     return q;
692 }
693
694 //----------------------------------------------
695 // 2-Pass code
696
697 static int init_pass2(MpegEncContext *s)
698 {
699     RateControlContext *rcc= &s->rc_context;
700     int i;
701     double fps= (double)s->frame_rate / FRAME_RATE_BASE;
702     double complexity[5]={0,0,0,0,0};   // aproximate bits at quant=1
703     double avg_quantizer[5];
704     uint64_t const_bits[5]={0,0,0,0,0}; // quantizer idependant bits
705     uint64_t available_bits[5];
706     uint64_t all_const_bits;
707     uint64_t all_available_bits= (uint64_t)(s->bit_rate*(double)rcc->num_entries/fps);
708     double rate_factor=0;
709     double step;
710     //int last_i_frame=-10000000;
711     const int filter_size= (int)(s->qblur*4) | 1;  
712     double expected_bits;
713     double *qscale, *blured_qscale;
714
715     /* find complexity & const_bits & decide the pict_types */
716     for(i=0; i<rcc->num_entries; i++){
717         RateControlEntry *rce= &rcc->entry[i];
718         
719         rce->new_pict_type= rce->pict_type;
720         rcc->i_cplx_sum [rce->pict_type] += rce->i_tex_bits*rce->qscale;
721         rcc->p_cplx_sum [rce->pict_type] += rce->p_tex_bits*rce->qscale;
722         rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
723         rcc->frame_count[rce->pict_type] ++;
724
725         complexity[rce->new_pict_type]+= (rce->i_tex_bits+ rce->p_tex_bits)*(double)rce->qscale;
726         const_bits[rce->new_pict_type]+= rce->mv_bits + rce->misc_bits;
727     }
728     all_const_bits= const_bits[I_TYPE] + const_bits[P_TYPE] + const_bits[B_TYPE];
729     
730     if(all_available_bits < all_const_bits){
731         fprintf(stderr, "requested bitrate is to low\n");
732         return -1;
733     }
734     
735     /* find average quantizers */
736     avg_quantizer[P_TYPE]=0;
737     for(step=256*256; step>0.0000001; step*=0.5){
738         double expected_bits=0;
739         avg_quantizer[P_TYPE]+= step;
740         
741         avg_quantizer[I_TYPE]= avg_quantizer[P_TYPE]*ABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset;
742         avg_quantizer[B_TYPE]= avg_quantizer[P_TYPE]*ABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset;
743         
744         expected_bits= 
745             + all_const_bits 
746             + complexity[I_TYPE]/avg_quantizer[I_TYPE]
747             + complexity[P_TYPE]/avg_quantizer[P_TYPE]
748             + complexity[B_TYPE]/avg_quantizer[B_TYPE];
749             
750         if(expected_bits < all_available_bits) avg_quantizer[P_TYPE]-= step;
751 //printf("%f %lld %f\n", expected_bits, all_available_bits, avg_quantizer[P_TYPE]);
752     }
753 //printf("qp_i:%f, qp_p:%f, qp_b:%f\n", avg_quantizer[I_TYPE],avg_quantizer[P_TYPE],avg_quantizer[B_TYPE]);
754
755     for(i=0; i<5; i++){
756         available_bits[i]= const_bits[i] + complexity[i]/avg_quantizer[i];
757     }
758 //printf("%lld %lld %lld %lld\n", available_bits[I_TYPE], available_bits[P_TYPE], available_bits[B_TYPE], all_available_bits);
759         
760     qscale= av_malloc(sizeof(double)*rcc->num_entries);
761     blured_qscale= av_malloc(sizeof(double)*rcc->num_entries);
762
763     for(step=256*256; step>0.0000001; step*=0.5){
764         expected_bits=0;
765         rate_factor+= step;
766         
767         rcc->buffer_index= s->avctx->rc_buffer_size/2;
768
769         /* find qscale */
770         for(i=0; i<rcc->num_entries; i++){
771             qscale[i]= get_qscale(s, &rcc->entry[i], rate_factor, i);
772         }
773         assert(filter_size%2==1);
774
775         /* fixed I/B QP relative to P mode */
776         for(i=rcc->num_entries-1; i>=0; i--){
777             RateControlEntry *rce= &rcc->entry[i];
778             
779             qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
780         }
781
782         /* smooth curve */
783         for(i=0; i<rcc->num_entries; i++){
784             RateControlEntry *rce= &rcc->entry[i];
785             const int pict_type= rce->new_pict_type;
786             int j;
787             double q=0.0, sum=0.0;
788         
789             for(j=0; j<filter_size; j++){
790                 int index= i+j-filter_size/2;
791                 double d= index-i;
792                 double coeff= s->qblur==0 ? 1.0 : exp(-d*d/(s->qblur * s->qblur));
793             
794                 if(index < 0 || index >= rcc->num_entries) continue;
795                 if(pict_type != rcc->entry[index].new_pict_type) continue;
796                 q+= qscale[index] * coeff;
797                 sum+= coeff;
798             }
799             blured_qscale[i]= q/sum;
800         }
801     
802         /* find expected bits */
803         for(i=0; i<rcc->num_entries; i++){
804             RateControlEntry *rce= &rcc->entry[i];
805             double bits;
806             rce->new_qscale= modify_qscale(s, rce, blured_qscale[i], i);
807             bits= qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
808 //printf("%d %f\n", rce->new_bits, blured_qscale[i]);
809             update_rc_buffer(s, bits);
810
811             rce->expected_bits= expected_bits;
812             expected_bits += bits;
813         }
814
815 //        printf("%f %d %f\n", expected_bits, (int)all_available_bits, rate_factor);
816         if(expected_bits > all_available_bits) rate_factor-= step;
817     }
818     av_free(qscale);
819     av_free(blured_qscale);
820
821     if(abs(expected_bits/all_available_bits - 1.0) > 0.01 ){
822         fprintf(stderr, "Error: 2pass curve failed to converge\n");
823         return -1;
824     }
825
826     return 0;
827 }