fixing 2pass assert failure
[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->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     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     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= FFMIN(q, bits2qp(rce, FFMAX((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= FFMAX(q, bits2qp(rce, FFMAX(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     Picture * const pic= &s->current_picture;
479     
480     for(i=0; i<s->mb_num; i++){
481         float temp_cplx= sqrt(pic->mc_mb_var[i]);
482         float spat_cplx= sqrt(pic->mb_var[i]);
483         const int lumi= pic->mb_mean[i];
484         float bits, cplx, factor;
485         
486         if(spat_cplx < q/3) spat_cplx= q/3; //FIXME finetune
487         if(temp_cplx < q/3) temp_cplx= q/3; //FIXME finetune
488         
489         if((s->mb_type[i]&MB_TYPE_INTRA)){//FIXME hq mode 
490             cplx= spat_cplx;
491             factor= 1.0 + p_masking;
492         }else{
493             cplx= temp_cplx;
494             factor= pow(temp_cplx, - temp_cplx_masking);
495         }
496         factor*=pow(spat_cplx, - spatial_cplx_masking);
497
498         if(lumi>127)
499             factor*= (1.0 - (lumi-128)*(lumi-128)*lumi_masking);
500         else
501             factor*= (1.0 - (lumi-128)*(lumi-128)*dark_masking);
502         
503         if(factor<0.00001) factor= 0.00001;
504         
505         bits= cplx*factor;
506         cplx_sum+= cplx;
507         bits_sum+= bits;
508         cplx_tab[i]= cplx;
509         bits_tab[i]= bits;
510     }
511
512     /* handle qmin/qmax cliping */
513     if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
514         for(i=0; i<s->mb_num; i++){
515             float newq= q*cplx_tab[i]/bits_tab[i];
516             newq*= bits_sum/cplx_sum;
517
518             if     (newq > qmax){
519                 bits_sum -= bits_tab[i];
520                 cplx_sum -= cplx_tab[i]*q/qmax;
521             }
522             else if(newq < qmin){
523                 bits_sum -= bits_tab[i];
524                 cplx_sum -= cplx_tab[i]*q/qmin;
525             }
526         }
527     }
528    
529     for(i=0; i<s->mb_num; i++){
530         float newq= q*cplx_tab[i]/bits_tab[i];
531         int intq;
532
533         if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
534             newq*= bits_sum/cplx_sum;
535         }
536
537         if(i && ABS(pic->qscale_table[i-1] - newq)<0.75)
538             intq= pic->qscale_table[i-1];
539         else
540             intq= (int)(newq + 0.5);
541
542         if     (intq > qmax) intq= qmax;
543         else if(intq < qmin) intq= qmin;
544 //if(i%s->mb_width==0) printf("\n");
545 //printf("%2d%3d ", intq, ff_sqrt(s->mc_mb_var[i]));
546         pic->qscale_table[i]= intq;
547     }
548 }
549
550 float ff_rate_estimate_qscale(MpegEncContext *s)
551 {
552     float q;
553     int qmin, qmax;
554     float br_compensation;
555     double diff;
556     double short_term_q;
557     double fps;
558     int picture_number= s->picture_number;
559     int64_t wanted_bits;
560     RateControlContext *rcc= &s->rc_context;
561     RateControlEntry local_rce, *rce;
562     double bits;
563     double rate_factor;
564     int var;
565     const int pict_type= s->pict_type;
566     Picture * const pic= &s->current_picture;
567     emms_c();
568
569     get_qminmax(&qmin, &qmax, s, pict_type);
570
571     fps= (double)s->frame_rate / FRAME_RATE_BASE;
572 //printf("input_pic_num:%d pic_num:%d frame_rate:%d\n", s->input_picture_number, s->picture_number, s->frame_rate);
573         /* update predictors */
574     if(picture_number>2){
575         const int last_var= s->last_pict_type == I_TYPE ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum;
576         update_predictor(&rcc->pred[s->last_pict_type], rcc->last_qscale, sqrt(last_var), s->frame_bits);
577     }
578
579     if(s->flags&CODEC_FLAG_PASS2){
580         assert(picture_number>=0);
581         assert(picture_number<rcc->num_entries);
582         rce= &rcc->entry[picture_number];
583         wanted_bits= rce->expected_bits;
584     }else{
585         rce= &local_rce;
586         wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps);
587     }
588
589     diff= s->total_bits - wanted_bits;
590     br_compensation= (s->bit_rate_tolerance - diff)/s->bit_rate_tolerance;
591     if(br_compensation<=0.0) br_compensation=0.001;
592
593     var= pict_type == I_TYPE ? pic->mb_var_sum : pic->mc_mb_var_sum;
594     
595     if(s->flags&CODEC_FLAG_PASS2){
596         if(pict_type!=I_TYPE)
597             assert(pict_type == rce->new_pict_type);
598
599         q= rce->new_qscale / br_compensation;
600 //printf("%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale, br_compensation, s->frame_bits, var, pict_type);
601     }else{
602         rce->pict_type= 
603         rce->new_pict_type= pict_type;
604         rce->mc_mb_var_sum= pic->mc_mb_var_sum;
605         rce->mb_var_sum   = pic->   mb_var_sum;
606         rce->qscale   = 2;
607         rce->f_code   = s->f_code;
608         rce->b_code   = s->b_code;
609         rce->misc_bits= 1;
610
611         if(picture_number>0)
612             update_rc_buffer(s, s->frame_bits);
613
614         bits= predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
615         if(pict_type== I_TYPE){
616             rce->i_count   = s->mb_num;
617             rce->i_tex_bits= bits;
618             rce->p_tex_bits= 0;
619             rce->mv_bits= 0;
620         }else{
621             rce->i_count   = 0; //FIXME we do know this approx
622             rce->i_tex_bits= 0;
623             rce->p_tex_bits= bits*0.9;
624             
625             rce->mv_bits= bits*0.1;
626         }
627         rcc->i_cplx_sum [pict_type] += rce->i_tex_bits*rce->qscale;
628         rcc->p_cplx_sum [pict_type] += rce->p_tex_bits*rce->qscale;
629         rcc->mv_bits_sum[pict_type] += rce->mv_bits;
630         rcc->frame_count[pict_type] ++;
631
632         bits= rce->i_tex_bits + rce->p_tex_bits;
633         rate_factor= rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum * br_compensation;
634     
635         q= get_qscale(s, rce, rate_factor, picture_number);
636
637         assert(q>0.0);
638 //printf("%f ", q);
639         q= get_diff_limited_q(s, rce, q);
640 //printf("%f ", q);
641         assert(q>0.0);
642
643         if(pict_type==P_TYPE || s->intra_only){ //FIXME type dependant blur like in 2-pass
644             rcc->short_term_qsum*=s->qblur;
645             rcc->short_term_qcount*=s->qblur;
646
647             rcc->short_term_qsum+= q;
648             rcc->short_term_qcount++;
649 //printf("%f ", q);
650             q= short_term_q= rcc->short_term_qsum/rcc->short_term_qcount;
651 //printf("%f ", q);
652         }
653         assert(q>0.0);
654         
655         q= modify_qscale(s, rce, q, picture_number);
656
657         rcc->pass1_wanted_bits+= s->bit_rate/fps;
658
659         assert(q>0.0);
660     }
661 //printf("qmin:%d, qmax:%d, q:%f\n", qmin, qmax, q);
662     
663
664     if     (q<qmin) q=qmin; 
665     else if(q>qmax) q=qmax;
666         
667 //    printf("%f %d %d %d\n", q, picture_number, (int)wanted_bits, (int)s->total_bits);
668        
669 //printf("diff:%d comp:%f st_q:%f last_size:%d type:%d\n", (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= pic->mc_mb_var_sum;
680     rcc->last_mb_var_sum= pic->mb_var_sum;
681 #if 0
682 {
683     static int mvsum=0, texsum=0;
684     mvsum += s->mv_bits;
685     texsum += s->i_tex_bits + s->p_tex_bits;
686     printf("%d %d//\n\n", mvsum, texsum);
687 }
688 #endif
689     return q;
690 }
691
692 //----------------------------------------------
693 // 2-Pass code
694
695 static int init_pass2(MpegEncContext *s)
696 {
697     RateControlContext *rcc= &s->rc_context;
698     int i;
699     double fps= (double)s->frame_rate / FRAME_RATE_BASE;
700     double complexity[5]={0,0,0,0,0};   // aproximate bits at quant=1
701     double avg_quantizer[5];
702     uint64_t const_bits[5]={0,0,0,0,0}; // quantizer idependant bits
703     uint64_t available_bits[5];
704     uint64_t all_const_bits;
705     uint64_t all_available_bits= (uint64_t)(s->bit_rate*(double)rcc->num_entries/fps);
706     double rate_factor=0;
707     double step;
708     int last_i_frame=-10000000;
709     const int filter_size= (int)(s->qblur*4) | 1;  
710     double expected_bits;
711     double *qscale, *blured_qscale;
712
713     /* find complexity & const_bits & decide the pict_types */
714     for(i=0; i<rcc->num_entries; i++){
715         RateControlEntry *rce= &rcc->entry[i];
716         
717         rce->new_pict_type= rce->pict_type;
718         rcc->i_cplx_sum [rce->pict_type] += rce->i_tex_bits*rce->qscale;
719         rcc->p_cplx_sum [rce->pict_type] += rce->p_tex_bits*rce->qscale;
720         rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
721         rcc->frame_count[rce->pict_type] ++;
722
723         complexity[rce->new_pict_type]+= (rce->i_tex_bits+ rce->p_tex_bits)*(double)rce->qscale;
724         const_bits[rce->new_pict_type]+= rce->mv_bits + rce->misc_bits;
725     }
726     all_const_bits= const_bits[I_TYPE] + const_bits[P_TYPE] + const_bits[B_TYPE];
727     
728     if(all_available_bits < all_const_bits){
729         fprintf(stderr, "requested bitrate is to low\n");
730         return -1;
731     }
732     
733     /* find average quantizers */
734     avg_quantizer[P_TYPE]=0;
735     for(step=256*256; step>0.0000001; step*=0.5){
736         double expected_bits=0;
737         avg_quantizer[P_TYPE]+= step;
738         
739         avg_quantizer[I_TYPE]= avg_quantizer[P_TYPE]*ABS(s->avctx->i_quant_factor) + s->avctx->i_quant_offset;
740         avg_quantizer[B_TYPE]= avg_quantizer[P_TYPE]*ABS(s->avctx->b_quant_factor) + s->avctx->b_quant_offset;
741         
742         expected_bits= 
743             + all_const_bits 
744             + complexity[I_TYPE]/avg_quantizer[I_TYPE]
745             + complexity[P_TYPE]/avg_quantizer[P_TYPE]
746             + complexity[B_TYPE]/avg_quantizer[B_TYPE];
747             
748         if(expected_bits < all_available_bits) avg_quantizer[P_TYPE]-= step;
749 //printf("%f %lld %f\n", expected_bits, all_available_bits, avg_quantizer[P_TYPE]);
750     }
751 //printf("qp_i:%f, qp_p:%f, qp_b:%f\n", avg_quantizer[I_TYPE],avg_quantizer[P_TYPE],avg_quantizer[B_TYPE]);
752
753     for(i=0; i<5; i++){
754         available_bits[i]= const_bits[i] + complexity[i]/avg_quantizer[i];
755     }
756 //printf("%lld %lld %lld %lld\n", available_bits[I_TYPE], available_bits[P_TYPE], available_bits[B_TYPE], all_available_bits);
757         
758     qscale= malloc(sizeof(double)*rcc->num_entries);
759     blured_qscale= malloc(sizeof(double)*rcc->num_entries);
760
761     for(step=256*256; step>0.0000001; step*=0.5){
762         expected_bits=0;
763         rate_factor+= step;
764         
765         rcc->buffer_index= s->avctx->rc_buffer_size/2;
766
767         /* find qscale */
768         for(i=0; i<rcc->num_entries; i++){
769             qscale[i]= get_qscale(s, &rcc->entry[i], rate_factor, i);
770         }
771         assert(filter_size%2==1);
772
773         /* fixed I/B QP relative to P mode */
774         for(i=rcc->num_entries-1; i>=0; i--){
775             RateControlEntry *rce= &rcc->entry[i];
776             
777             qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
778         }
779
780         /* smooth curve */
781         for(i=0; i<rcc->num_entries; i++){
782             RateControlEntry *rce= &rcc->entry[i];
783             const int pict_type= rce->new_pict_type;
784             int j;
785             double q=0.0, sum=0.0;
786         
787             for(j=0; j<filter_size; j++){
788                 int index= i+j-filter_size/2;
789                 double d= index-i;
790                 double coeff= s->qblur==0 ? 1.0 : exp(-d*d/(s->qblur * s->qblur));
791             
792                 if(index < 0 || index >= rcc->num_entries) continue;
793                 if(pict_type != rcc->entry[index].new_pict_type) continue;
794                 q+= qscale[index] * coeff;
795                 sum+= coeff;
796             }
797             blured_qscale[i]= q/sum;
798         }
799     
800         /* find expected bits */
801         for(i=0; i<rcc->num_entries; i++){
802             RateControlEntry *rce= &rcc->entry[i];
803             double bits;
804             rce->new_qscale= modify_qscale(s, rce, blured_qscale[i], i);
805             bits= qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
806 //printf("%d %f\n", rce->new_bits, blured_qscale[i]);
807             update_rc_buffer(s, bits);
808
809             rce->expected_bits= expected_bits;
810             expected_bits += bits;
811         }
812
813 //        printf("%f %d %f\n", expected_bits, (int)all_available_bits, rate_factor);
814         if(expected_bits > all_available_bits) rate_factor-= step;
815     }
816     free(qscale);
817     free(blured_qscale);
818
819     if(abs(expected_bits/all_available_bits - 1.0) > 0.01 ){
820         fprintf(stderr, "Error: 2pass curve failed to converge\n");
821         return -1;
822     }
823
824     return 0;
825 }