renaming variable
[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 #include <math.h>
21 #include "common.h"
22 #include "avcodec.h"
23 #include "dsputil.h"
24 #include "mpegvideo.h"
25
26 #undef NDEBUG // allways check asserts, the speed effect is far too small to disable them
27 #include <assert.h>
28
29 #ifndef M_PI
30 #define M_PI 3.14159265358979323846
31 #endif
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->mc_mb_var_sum, s->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     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         bits2qp,
288         qp2bits,
289         NULL
290     };
291     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         if     (q > last_q + a->max_qdiff) q= last_q + a->max_qdiff;
343         else if(q < last_q - a->max_qdiff) q= last_q - a->max_qdiff;
344     }
345
346     rcc->last_qscale_for[pict_type]= q; //Note we cant do that after blurring
347     
348     if(pict_type!=B_TYPE)
349         rcc->last_non_b_pict_type= pict_type;
350
351     return q;
352 }
353
354 /**
355  * gets the qmin & qmax for pict_type
356  */
357 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type){
358     int qmin= s->qmin;                                                       
359     int qmax= s->qmax;
360
361     if(pict_type==B_TYPE){
362         qmin= (int)(qmin*ABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
363         qmax= (int)(qmax*ABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
364     }else if(pict_type==I_TYPE){
365         qmin= (int)(qmin*ABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
366         qmax= (int)(qmax*ABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
367     }
368
369     if(qmin<1) qmin=1;
370     if(qmin==1 && s->qmin>1) qmin=2; //avoid qmin=1 unless the user wants qmin=1
371
372     if(qmin<3 && s->max_qcoeff<=128 && pict_type==I_TYPE) qmin=3; //reduce cliping problems
373
374     if(qmax>31) qmax=31;
375     if(qmax<=qmin) qmax= qmin= (qmax+qmin+1)>>1;
376     
377     *qmin_ret= qmin;
378     *qmax_ret= qmax;
379 }
380
381 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce, double q, int frame_num){
382     RateControlContext *rcc= &s->rc_context;
383     int qmin, qmax;
384     double bits;
385     const int pict_type= rce->new_pict_type;
386     const double buffer_size= s->avctx->rc_buffer_size;
387     const double min_rate= s->avctx->rc_min_rate;
388     const double max_rate= s->avctx->rc_max_rate;
389     
390     get_qminmax(&qmin, &qmax, s, pict_type);
391
392     /* modulation */
393     if(s->avctx->rc_qmod_freq && frame_num%s->avctx->rc_qmod_freq==0 && pict_type==P_TYPE)
394         q*= s->avctx->rc_qmod_amp;
395
396     bits= qp2bits(rce, q);
397 //printf("q:%f\n", q);
398     /* buffer overflow/underflow protection */
399     if(buffer_size){
400         double expected_size= rcc->buffer_index;
401
402         if(min_rate){
403             double d= 2*(buffer_size - expected_size)/buffer_size;
404             if(d>1.0) d=1.0;
405             else if(d<0.0001) d=0.0001;
406             q*= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
407
408             q= MIN(q, bits2qp(rce, MAX((min_rate - buffer_size + rcc->buffer_index)*2, 1)));
409         }
410
411         if(max_rate){
412             double d= 2*expected_size/buffer_size;
413             if(d>1.0) d=1.0;
414             else if(d<0.0001) d=0.0001;
415             q/= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
416
417             q= MAX(q, bits2qp(rce, MAX(rcc->buffer_index/2, 1)));
418         }
419     }
420 //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);
421     if(s->avctx->rc_qsquish==0.0 || qmin==qmax){
422         if     (q<qmin) q=qmin;
423         else if(q>qmax) q=qmax;
424     }else{
425         double min2= log(qmin);
426         double max2= log(qmax);
427         
428         q= log(q);
429         q= (q - min2)/(max2-min2) - 0.5;
430         q*= -4.0;
431         q= 1.0/(1.0 + exp(q));
432         q= q*(max2-min2) + min2;
433         
434         q= exp(q);
435     }
436     
437     return q;
438 }
439
440 //----------------------------------
441 // 1 Pass Code
442
443 static double predict_size(Predictor *p, double q, double var)
444 {
445      return p->coeff*var / (q*p->count);
446 }
447
448 static double predict_qp(Predictor *p, double size, double var)
449 {
450 //printf("coeff:%f, count:%f, var:%f, size:%f//\n", p->coeff, p->count, var, size);
451      return p->coeff*var / (size*p->count);
452 }
453
454 static void update_predictor(Predictor *p, double q, double var, double size)
455 {
456     double new_coeff= size*q / (var + 1);
457     if(var<10) return;
458
459     p->count*= p->decay;
460     p->coeff*= p->decay;
461     p->count++;
462     p->coeff+= new_coeff;
463 }
464
465 static void adaptive_quantization(MpegEncContext *s, double q){
466     int i;
467     const float lumi_masking= s->avctx->lumi_masking / (128.0*128.0);
468     const float dark_masking= s->avctx->dark_masking / (128.0*128.0);
469     const float temp_cplx_masking= s->avctx->temporal_cplx_masking;
470     const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
471     const float p_masking = s->avctx->p_masking;
472     float bits_sum= 0.0;
473     float cplx_sum= 0.0;
474     float cplx_tab[s->mb_num];
475     float bits_tab[s->mb_num];
476     const int qmin= 2; //s->avctx->mb_qmin;
477     const int qmax= 31; //s->avctx->mb_qmax;
478     
479     for(i=0; i<s->mb_num; i++){
480         float temp_cplx= sqrt(s->mc_mb_var[i]);
481         float spat_cplx= sqrt(s->mb_var[i]);
482         const int lumi= s->mb_mean[i];
483         float bits, cplx, factor;
484         
485         if(spat_cplx < q/3) spat_cplx= q/3; //FIXME finetune
486         if(temp_cplx < q/3) temp_cplx= q/3; //FIXME finetune
487         
488         if((s->mb_type[i]&MB_TYPE_INTRA)){//FIXME hq mode 
489             cplx= spat_cplx;
490             factor= 1.0 + p_masking;
491         }else{
492             cplx= temp_cplx;
493             factor= pow(temp_cplx, - temp_cplx_masking);
494         }
495         factor*=pow(spat_cplx, - spatial_cplx_masking);
496
497         if(lumi>127)
498             factor*= (1.0 - (lumi-128)*(lumi-128)*lumi_masking);
499         else
500             factor*= (1.0 - (lumi-128)*(lumi-128)*dark_masking);
501         
502         if(factor<0.00001) factor= 0.00001;
503         
504         bits= cplx*factor;
505         cplx_sum+= cplx;
506         bits_sum+= bits;
507         cplx_tab[i]= cplx;
508         bits_tab[i]= bits;
509     }
510
511     /* handle qmin/qmax cliping */
512     if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
513         for(i=0; i<s->mb_num; i++){
514             float newq= q*cplx_tab[i]/bits_tab[i];
515             newq*= bits_sum/cplx_sum;
516
517             if     (newq > qmax){
518                 bits_sum -= bits_tab[i];
519                 cplx_sum -= cplx_tab[i]*q/qmax;
520             }
521             else if(newq < qmin){
522                 bits_sum -= bits_tab[i];
523                 cplx_sum -= cplx_tab[i]*q/qmin;
524             }
525         }
526     }
527    
528     for(i=0; i<s->mb_num; i++){
529         float newq= q*cplx_tab[i]/bits_tab[i];
530         int intq;
531
532         if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
533             newq*= bits_sum/cplx_sum;
534         }
535
536         if(i && ABS(s->qscale_table[i-1] - newq)<0.75)
537             intq= s->qscale_table[i-1];
538         else
539             intq= (int)(newq + 0.5);
540
541         if     (intq > qmax) intq= qmax;
542         else if(intq < qmin) intq= qmin;
543 //if(i%s->mb_width==0) printf("\n");
544 //printf("%2d%3d ", intq, ff_sqrt(s->mc_mb_var[i]));
545         s->qscale_table[i]= intq;
546     }
547 }
548
549 float ff_rate_estimate_qscale(MpegEncContext *s)
550 {
551     float q;
552     int qmin, qmax;
553     float br_compensation;
554     double diff;
555     double short_term_q;
556     double fps;
557     int picture_number= s->picture_number;
558     int64_t wanted_bits;
559     RateControlContext *rcc= &s->rc_context;
560     RateControlEntry local_rce, *rce;
561     double bits;
562     double rate_factor;
563     int var;
564     const int pict_type= s->pict_type;
565     emms_c();
566
567     get_qminmax(&qmin, &qmax, s, pict_type);
568
569     fps= (double)s->frame_rate / FRAME_RATE_BASE;
570 //printf("input_pic_num:%d pic_num:%d frame_rate:%d\n", s->input_picture_number, s->picture_number, s->frame_rate);
571         /* update predictors */
572     if(picture_number>2){
573         const int last_var= s->last_pict_type == I_TYPE ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum;
574         update_predictor(&rcc->pred[s->last_pict_type], rcc->last_qscale, sqrt(last_var), s->frame_bits);
575     }
576
577     if(s->flags&CODEC_FLAG_PASS2){
578         assert(picture_number>=0);
579         assert(picture_number<rcc->num_entries);
580         rce= &rcc->entry[picture_number];
581         wanted_bits= rce->expected_bits;
582     }else{
583         rce= &local_rce;
584         wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps);
585     }
586
587     diff= s->total_bits - wanted_bits;
588     br_compensation= (s->bit_rate_tolerance - diff)/s->bit_rate_tolerance;
589     if(br_compensation<=0.0) br_compensation=0.001;
590
591     var= pict_type == I_TYPE ? s->mb_var_sum : s->mc_mb_var_sum;
592     
593     if(s->flags&CODEC_FLAG_PASS2){
594         if(pict_type!=I_TYPE)
595             assert(pict_type == rce->new_pict_type);
596
597         q= rce->new_qscale / br_compensation;
598 //printf("%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale, br_compensation, s->frame_bits, var, pict_type);
599     }else{
600         rce->pict_type= 
601         rce->new_pict_type= pict_type;
602         rce->mc_mb_var_sum= s->mc_mb_var_sum;
603         rce->mb_var_sum   = s->   mb_var_sum;
604         rce->qscale   = 2;
605         rce->f_code   = s->f_code;
606         rce->b_code   = s->b_code;
607         rce->misc_bits= 1;
608
609         if(picture_number>0)
610             update_rc_buffer(s, s->frame_bits);
611
612         bits= predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
613         if(pict_type== I_TYPE){
614             rce->i_count   = s->mb_num;
615             rce->i_tex_bits= bits;
616             rce->p_tex_bits= 0;
617             rce->mv_bits= 0;
618         }else{
619             rce->i_count   = 0; //FIXME we do know this approx
620             rce->i_tex_bits= 0;
621             rce->p_tex_bits= bits*0.9;
622             
623             rce->mv_bits= bits*0.1;
624         }
625         rcc->i_cplx_sum [pict_type] += rce->i_tex_bits*rce->qscale;
626         rcc->p_cplx_sum [pict_type] += rce->p_tex_bits*rce->qscale;
627         rcc->mv_bits_sum[pict_type] += rce->mv_bits;
628         rcc->frame_count[pict_type] ++;
629
630         bits= rce->i_tex_bits + rce->p_tex_bits;
631         rate_factor= rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum * br_compensation;
632     
633         q= get_qscale(s, rce, rate_factor, picture_number);
634
635         assert(q>0.0);
636 //printf("%f ", q);
637         q= get_diff_limited_q(s, rce, q);
638 //printf("%f ", q);
639         assert(q>0.0);
640
641         if(pict_type==P_TYPE || s->intra_only){ //FIXME type dependant blur like in 2-pass
642             rcc->short_term_qsum*=s->qblur;
643             rcc->short_term_qcount*=s->qblur;
644
645             rcc->short_term_qsum+= q;
646             rcc->short_term_qcount++;
647 //printf("%f ", q);
648             q= short_term_q= rcc->short_term_qsum/rcc->short_term_qcount;
649 //printf("%f ", q);
650         }
651         assert(q>0.0);
652         
653         q= modify_qscale(s, rce, q, picture_number);
654
655         rcc->pass1_wanted_bits+= s->bit_rate/fps;
656
657         assert(q>0.0);
658     }
659 //printf("qmin:%d, qmax:%d, q:%f\n", qmin, qmax, q);
660     
661
662     if     (q<qmin) q=qmin; 
663     else if(q>qmax) q=qmax;
664         
665 //    printf("%f %d %d %d\n", q, picture_number, (int)wanted_bits, (int)s->total_bits);
666     
667 //printf("%f %f %f\n", q, br_compensation, short_term_q);
668    
669 //printf("q:%d diff:%d comp:%f st_q:%f last_size:%d type:%d\n", qscale, (int)diff, br_compensation, 
670 //       short_term_q, s->frame_bits, pict_type);
671 //printf("%d %d\n", s->bit_rate, (int)fps);
672
673     if(s->adaptive_quant)
674         adaptive_quantization(s, q);
675     else
676         q= (int)(q + 0.5);
677     
678     rcc->last_qscale= q;
679     rcc->last_mc_mb_var_sum= s->mc_mb_var_sum;
680     rcc->last_mb_var_sum= s->mb_var_sum;
681     return q;
682 }
683
684 //----------------------------------------------
685 // 2-Pass code
686
687 static int init_pass2(MpegEncContext *s)
688 {
689     RateControlContext *rcc= &s->rc_context;
690     int i;
691     double fps= (double)s->frame_rate / FRAME_RATE_BASE;
692     double complexity[5]={0,0,0,0,0};   // aproximate bits at quant=1
693     double avg_quantizer[5];
694     uint64_t const_bits[5]={0,0,0,0,0}; // quantizer idependant bits
695     uint64_t available_bits[5];
696     uint64_t all_const_bits;
697     uint64_t all_available_bits= (uint64_t)(s->bit_rate*(double)rcc->num_entries/fps);
698     double rate_factor=0;
699     double step;
700     int last_i_frame=-10000000;
701     const int filter_size= (int)(s->qblur*4) | 1;  
702     double expected_bits;
703     double *qscale, *blured_qscale;
704
705     /* find complexity & const_bits & decide the pict_types */
706     for(i=0; i<rcc->num_entries; i++){
707         RateControlEntry *rce= &rcc->entry[i];
708         
709         if(s->b_frame_strategy==0 || s->max_b_frames==0){
710             rce->new_pict_type= rce->pict_type;
711         }else{
712             int j;
713             int next_non_b_type=P_TYPE;
714
715             switch(rce->pict_type){
716             case I_TYPE:
717                 if(i-last_i_frame>s->gop_size/2){ //FIXME this is not optimal
718                     rce->new_pict_type= I_TYPE;
719                     last_i_frame= i;
720                 }else{
721                     rce->new_pict_type= P_TYPE; // will be caught by the scene detection anyway
722                 }
723                 break;
724             case P_TYPE:
725                 rce->new_pict_type= P_TYPE;
726                 break;
727             case B_TYPE:
728                 for(j=i+1; j<i+s->max_b_frames+2 && j<rcc->num_entries; j++){
729                     if(rcc->entry[j].pict_type != B_TYPE){
730                         next_non_b_type= rcc->entry[j].pict_type;
731                         break;
732                     }
733                 }
734                 if(next_non_b_type==I_TYPE)
735                     rce->new_pict_type= P_TYPE;
736                 else
737                     rce->new_pict_type= B_TYPE;
738                 break;
739             }
740         }
741         rcc->i_cplx_sum [rce->pict_type] += rce->i_tex_bits*rce->qscale;
742         rcc->p_cplx_sum [rce->pict_type] += rce->p_tex_bits*rce->qscale;
743         rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
744         rcc->frame_count[rce->pict_type] ++;
745
746         complexity[rce->new_pict_type]+= (rce->i_tex_bits+ rce->p_tex_bits)*(double)rce->qscale;
747         const_bits[rce->new_pict_type]+= rce->mv_bits + rce->misc_bits;
748     }
749     all_const_bits= const_bits[I_TYPE] + const_bits[P_TYPE] + const_bits[B_TYPE];
750     
751     if(all_available_bits < all_const_bits){
752         fprintf(stderr, "requested bitrate is to low\n");
753         return -1;
754     }
755     
756     /* find average quantizers */
757     avg_quantizer[P_TYPE]=0;
758     for(step=256*256; step>0.0000001; step*=0.5){
759         double expected_bits=0;
760         avg_quantizer[P_TYPE]+= step;
761         
762         avg_quantizer[I_TYPE]= avg_quantizer[P_TYPE]*ABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset;
763         avg_quantizer[B_TYPE]= avg_quantizer[P_TYPE]*ABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset;
764         
765         expected_bits= 
766             + all_const_bits 
767             + complexity[I_TYPE]/avg_quantizer[I_TYPE]
768             + complexity[P_TYPE]/avg_quantizer[P_TYPE]
769             + complexity[B_TYPE]/avg_quantizer[B_TYPE];
770             
771         if(expected_bits < all_available_bits) avg_quantizer[P_TYPE]-= step;
772 //printf("%f %lld %f\n", expected_bits, all_available_bits, avg_quantizer[P_TYPE]);
773     }
774 //printf("qp_i:%f, qp_p:%f, qp_b:%f\n", avg_quantizer[I_TYPE],avg_quantizer[P_TYPE],avg_quantizer[B_TYPE]);
775
776     for(i=0; i<5; i++){
777         available_bits[i]= const_bits[i] + complexity[i]/avg_quantizer[i];
778     }
779 //printf("%lld %lld %lld %lld\n", available_bits[I_TYPE], available_bits[P_TYPE], available_bits[B_TYPE], all_available_bits);
780         
781     qscale= malloc(sizeof(double)*rcc->num_entries);
782     blured_qscale= malloc(sizeof(double)*rcc->num_entries);
783
784     for(step=256*256; step>0.0000001; step*=0.5){
785         expected_bits=0;
786         rate_factor+= step;
787         
788         rcc->buffer_index= s->avctx->rc_buffer_size/2;
789
790         /* find qscale */
791         for(i=0; i<rcc->num_entries; i++){
792             qscale[i]= get_qscale(s, &rcc->entry[i], rate_factor, i);
793         }
794         assert(filter_size%2==1);
795
796         /* fixed I/B QP relative to P mode */
797         for(i=rcc->num_entries-1; i>=0; i--){
798             RateControlEntry *rce= &rcc->entry[i];
799             
800             qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
801         }
802
803         /* smooth curve */
804         for(i=0; i<rcc->num_entries; i++){
805             RateControlEntry *rce= &rcc->entry[i];
806             const int pict_type= rce->new_pict_type;
807             int j;
808             double q=0.0, sum=0.0;
809         
810             for(j=0; j<filter_size; j++){
811                 int index= i+j-filter_size/2;
812                 double d= index-i;
813                 double coeff= s->qblur==0 ? 1.0 : exp(-d*d/(s->qblur * s->qblur));
814             
815                 if(index < 0 || index >= rcc->num_entries) continue;
816                 if(pict_type != rcc->entry[index].new_pict_type) continue;
817                 q+= qscale[index] * coeff;
818                 sum+= coeff;
819             }
820             blured_qscale[i]= q/sum;
821         }
822     
823         /* find expected bits */
824         for(i=0; i<rcc->num_entries; i++){
825             RateControlEntry *rce= &rcc->entry[i];
826             double bits;
827             rce->new_qscale= modify_qscale(s, rce, blured_qscale[i], i);
828             bits= qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
829 //printf("%d %f\n", rce->new_bits, blured_qscale[i]);
830             update_rc_buffer(s, bits);
831
832             rce->expected_bits= expected_bits;
833             expected_bits += bits;
834         }
835
836 //        printf("%f %d %f\n", expected_bits, (int)all_available_bits, rate_factor);
837         if(expected_bits > all_available_bits) rate_factor-= step;
838     }
839     free(qscale);
840     free(blured_qscale);
841
842     if(abs(expected_bits/all_available_bits - 1.0) > 0.01 ){
843         fprintf(stderr, "Error: 2pass curve failed to converge\n");
844         return -1;
845     }
846
847     return 0;
848 }