CONFIG_XVID --> CONFIG_LIBXVID
[ffmpeg.git] / libavcodec / ratecontrol.c
1 /*
2  * Rate control for video encoders
3  *
4  * Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
5  *
6  * This file is part of FFmpeg.
7  *
8  * FFmpeg is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU Lesser General Public
10  * License as published by the Free Software Foundation; either
11  * version 2.1 of the License, or (at your option) any later version.
12  *
13  * FFmpeg is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * Lesser General Public License for more details.
17  *
18  * You should have received a copy of the GNU Lesser General Public
19  * License along with FFmpeg; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  */
22
23 /**
24  * @file ratecontrol.c
25  * Rate control for video encoders.
26  */
27
28 #include "avcodec.h"
29 #include "dsputil.h"
30 #include "ratecontrol.h"
31 #include "mpegvideo.h"
32 #include "eval.h"
33
34 #undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
35 #include <assert.h>
36
37 #ifndef M_E
38 #define M_E 2.718281828
39 #endif
40
41 static int init_pass2(MpegEncContext *s);
42 static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num);
43
44 void ff_write_pass1_stats(MpegEncContext *s){
45     snprintf(s->avctx->stats_out, 256, "in:%d out:%d type:%d q:%d itex:%d ptex:%d mv:%d misc:%d fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d;\n",
46             s->current_picture_ptr->display_picture_number, s->current_picture_ptr->coded_picture_number, s->pict_type,
47             s->current_picture.quality, s->i_tex_bits, s->p_tex_bits, s->mv_bits, s->misc_bits,
48             s->f_code, s->b_code, s->current_picture.mc_mb_var_sum, s->current_picture.mb_var_sum, s->i_count, s->skip_count, s->header_bits);
49 }
50
51 static inline double qp2bits(RateControlEntry *rce, double qp){
52     if(qp<=0.0){
53         av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
54     }
55     return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ qp;
56 }
57
58 static inline double bits2qp(RateControlEntry *rce, double bits){
59     if(bits<0.9){
60         av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
61     }
62     return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ bits;
63 }
64
65 int ff_rate_control_init(MpegEncContext *s)
66 {
67     RateControlContext *rcc= &s->rc_context;
68     int i;
69     char *error = NULL;
70     static const char *const_names[]={
71         "PI",
72         "E",
73         "iTex",
74         "pTex",
75         "tex",
76         "mv",
77         "fCode",
78         "iCount",
79         "mcVar",
80         "var",
81         "isI",
82         "isP",
83         "isB",
84         "avgQP",
85         "qComp",
86 /*        "lastIQP",
87         "lastPQP",
88         "lastBQP",
89         "nextNonBQP",*/
90         "avgIITex",
91         "avgPITex",
92         "avgPPTex",
93         "avgBPTex",
94         "avgTex",
95         NULL
96     };
97     static double (*func1[])(void *, double)={
98         (void *)bits2qp,
99         (void *)qp2bits,
100         NULL
101     };
102     static const char *func1_names[]={
103         "bits2qp",
104         "qp2bits",
105         NULL
106     };
107     emms_c();
108
109     rcc->rc_eq_eval = ff_parse(s->avctx->rc_eq, const_names, func1, func1_names, NULL, NULL, &error);
110     if (!rcc->rc_eq_eval) {
111         av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\": %s\n", s->avctx->rc_eq, error? error : "");
112         return -1;
113     }
114
115     for(i=0; i<5; i++){
116         rcc->pred[i].coeff= FF_QP2LAMBDA * 7.0;
117         rcc->pred[i].count= 1.0;
118
119         rcc->pred[i].decay= 0.4;
120         rcc->i_cplx_sum [i]=
121         rcc->p_cplx_sum [i]=
122         rcc->mv_bits_sum[i]=
123         rcc->qscale_sum [i]=
124         rcc->frame_count[i]= 1; // 1 is better cuz of 1/0 and such
125         rcc->last_qscale_for[i]=FF_QP2LAMBDA * 5;
126     }
127     rcc->buffer_index= s->avctx->rc_initial_buffer_occupancy;
128
129     if(s->flags&CODEC_FLAG_PASS2){
130         int i;
131         char *p;
132
133         /* find number of pics */
134         p= s->avctx->stats_in;
135         for(i=-1; p; i++){
136             p= strchr(p+1, ';');
137         }
138         i+= s->max_b_frames;
139         if(i<=0 || i>=INT_MAX / sizeof(RateControlEntry))
140             return -1;
141         rcc->entry = (RateControlEntry*)av_mallocz(i*sizeof(RateControlEntry));
142         rcc->num_entries= i;
143
144         /* init all to skipped p frames (with b frames we might have a not encoded frame at the end FIXME) */
145         for(i=0; i<rcc->num_entries; i++){
146             RateControlEntry *rce= &rcc->entry[i];
147             rce->pict_type= rce->new_pict_type=P_TYPE;
148             rce->qscale= rce->new_qscale=FF_QP2LAMBDA * 2;
149             rce->misc_bits= s->mb_num + 10;
150             rce->mb_var_sum= s->mb_num*100;
151         }
152
153         /* read stats */
154         p= s->avctx->stats_in;
155         for(i=0; i<rcc->num_entries - s->max_b_frames; i++){
156             RateControlEntry *rce;
157             int picture_number;
158             int e;
159             char *next;
160
161             next= strchr(p, ';');
162             if(next){
163                 (*next)=0; //sscanf in unbelievably slow on looong strings //FIXME copy / do not write
164                 next++;
165             }
166             e= sscanf(p, " in:%d ", &picture_number);
167
168             assert(picture_number >= 0);
169             assert(picture_number < rcc->num_entries);
170             rce= &rcc->entry[picture_number];
171
172             e+=sscanf(p, " in:%*d out:%*d type:%d q:%f itex:%d ptex:%d mv:%d misc:%d fcode:%d bcode:%d mc-var:%d var:%d icount:%d skipcount:%d hbits:%d",
173                    &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits, &rce->mv_bits, &rce->misc_bits,
174                    &rce->f_code, &rce->b_code, &rce->mc_mb_var_sum, &rce->mb_var_sum, &rce->i_count, &rce->skip_count, &rce->header_bits);
175             if(e!=14){
176                 av_log(s->avctx, AV_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
177                 return -1;
178             }
179
180             p= next;
181         }
182
183         if(init_pass2(s) < 0) return -1;
184
185         //FIXME maybe move to end
186         if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) {
187 #ifdef CONFIG_LIBXVID
188             return ff_xvid_rate_control_init(s);
189 #else
190             av_log(s->avctx, AV_LOG_ERROR, "XviD ratecontrol requires libavcodec compiled with XviD support\n");
191             return -1;
192 #endif
193         }
194     }
195
196     if(!(s->flags&CODEC_FLAG_PASS2)){
197
198         rcc->short_term_qsum=0.001;
199         rcc->short_term_qcount=0.001;
200
201         rcc->pass1_rc_eq_output_sum= 0.001;
202         rcc->pass1_wanted_bits=0.001;
203
204         if(s->avctx->qblur > 1.0){
205             av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
206             return -1;
207         }
208         /* init stuff with the user specified complexity */
209         if(s->avctx->rc_initial_cplx){
210             for(i=0; i<60*30; i++){
211                 double bits= s->avctx->rc_initial_cplx * (i/10000.0 + 1.0)*s->mb_num;
212                 RateControlEntry rce;
213                 double q;
214
215                 if     (i%((s->gop_size+3)/4)==0) rce.pict_type= I_TYPE;
216                 else if(i%(s->max_b_frames+1))    rce.pict_type= B_TYPE;
217                 else                              rce.pict_type= P_TYPE;
218
219                 rce.new_pict_type= rce.pict_type;
220                 rce.mc_mb_var_sum= bits*s->mb_num/100000;
221                 rce.mb_var_sum   = s->mb_num;
222                 rce.qscale   = FF_QP2LAMBDA * 2;
223                 rce.f_code   = 2;
224                 rce.b_code   = 1;
225                 rce.misc_bits= 1;
226
227                 if(s->pict_type== I_TYPE){
228                     rce.i_count   = s->mb_num;
229                     rce.i_tex_bits= bits;
230                     rce.p_tex_bits= 0;
231                     rce.mv_bits= 0;
232                 }else{
233                     rce.i_count   = 0; //FIXME we do know this approx
234                     rce.i_tex_bits= 0;
235                     rce.p_tex_bits= bits*0.9;
236                     rce.mv_bits= bits*0.1;
237                 }
238                 rcc->i_cplx_sum [rce.pict_type] += rce.i_tex_bits*rce.qscale;
239                 rcc->p_cplx_sum [rce.pict_type] += rce.p_tex_bits*rce.qscale;
240                 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
241                 rcc->frame_count[rce.pict_type] ++;
242
243                 bits= rce.i_tex_bits + rce.p_tex_bits;
244
245                 q= get_qscale(s, &rce, rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum, i);
246                 rcc->pass1_wanted_bits+= s->bit_rate/(1/av_q2d(s->avctx->time_base)); //FIXME missbehaves a little for variable fps
247             }
248         }
249
250     }
251
252     return 0;
253 }
254
255 void ff_rate_control_uninit(MpegEncContext *s)
256 {
257     RateControlContext *rcc= &s->rc_context;
258     emms_c();
259
260     ff_eval_free(rcc->rc_eq_eval);
261     av_freep(&rcc->entry);
262
263 #ifdef CONFIG_LIBXVID
264     if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
265         ff_xvid_rate_control_uninit(s);
266 #endif
267 }
268
269 int ff_vbv_update(MpegEncContext *s, int frame_size){
270     RateControlContext *rcc= &s->rc_context;
271     const double fps= 1/av_q2d(s->avctx->time_base);
272     const int buffer_size= s->avctx->rc_buffer_size;
273     const double min_rate= s->avctx->rc_min_rate/fps;
274     const double max_rate= s->avctx->rc_max_rate/fps;
275
276 //printf("%d %f %d %f %f\n", buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
277     if(buffer_size){
278         int left;
279
280         rcc->buffer_index-= frame_size;
281         if(rcc->buffer_index < 0){
282             av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
283             rcc->buffer_index= 0;
284         }
285
286         left= buffer_size - rcc->buffer_index - 1;
287         rcc->buffer_index += av_clip(left, min_rate, max_rate);
288
289         if(rcc->buffer_index > buffer_size){
290             int stuffing= ceil((rcc->buffer_index - buffer_size)/8);
291
292             if(stuffing < 4 && s->codec_id == CODEC_ID_MPEG4)
293                 stuffing=4;
294             rcc->buffer_index -= 8*stuffing;
295
296             if(s->avctx->debug & FF_DEBUG_RC)
297                 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
298
299             return stuffing;
300         }
301     }
302     return 0;
303 }
304
305 /**
306  * modifies the bitrate curve from pass1 for one frame
307  */
308 static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num){
309     RateControlContext *rcc= &s->rc_context;
310     AVCodecContext *a= s->avctx;
311     double q, bits;
312     const int pict_type= rce->new_pict_type;
313     const double mb_num= s->mb_num;
314     int i;
315
316     double const_values[]={
317         M_PI,
318         M_E,
319         rce->i_tex_bits*rce->qscale,
320         rce->p_tex_bits*rce->qscale,
321         (rce->i_tex_bits + rce->p_tex_bits)*(double)rce->qscale,
322         rce->mv_bits/mb_num,
323         rce->pict_type == B_TYPE ? (rce->f_code + rce->b_code)*0.5 : rce->f_code,
324         rce->i_count/mb_num,
325         rce->mc_mb_var_sum/mb_num,
326         rce->mb_var_sum/mb_num,
327         rce->pict_type == I_TYPE,
328         rce->pict_type == P_TYPE,
329         rce->pict_type == B_TYPE,
330         rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
331         a->qcompress,
332 /*        rcc->last_qscale_for[I_TYPE],
333         rcc->last_qscale_for[P_TYPE],
334         rcc->last_qscale_for[B_TYPE],
335         rcc->next_non_b_qscale,*/
336         rcc->i_cplx_sum[I_TYPE] / (double)rcc->frame_count[I_TYPE],
337         rcc->i_cplx_sum[P_TYPE] / (double)rcc->frame_count[P_TYPE],
338         rcc->p_cplx_sum[P_TYPE] / (double)rcc->frame_count[P_TYPE],
339         rcc->p_cplx_sum[B_TYPE] / (double)rcc->frame_count[B_TYPE],
340         (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
341         0
342     };
343
344     bits= ff_parse_eval(rcc->rc_eq_eval, const_values, rce);
345     if (isnan(bits)) {
346         av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq);
347         return -1;
348     }
349
350     rcc->pass1_rc_eq_output_sum+= bits;
351     bits*=rate_factor;
352     if(bits<0.0) bits=0.0;
353     bits+= 1.0; //avoid 1/0 issues
354
355     /* user override */
356     for(i=0; i<s->avctx->rc_override_count; i++){
357         RcOverride *rco= s->avctx->rc_override;
358         if(rco[i].start_frame > frame_num) continue;
359         if(rco[i].end_frame   < frame_num) continue;
360
361         if(rco[i].qscale)
362             bits= qp2bits(rce, rco[i].qscale); //FIXME move at end to really force it?
363         else
364             bits*= rco[i].quality_factor;
365     }
366
367     q= bits2qp(rce, bits);
368
369     /* I/B difference */
370     if     (pict_type==I_TYPE && s->avctx->i_quant_factor<0.0)
371         q= -q*s->avctx->i_quant_factor + s->avctx->i_quant_offset;
372     else if(pict_type==B_TYPE && s->avctx->b_quant_factor<0.0)
373         q= -q*s->avctx->b_quant_factor + s->avctx->b_quant_offset;
374
375     return q;
376 }
377
378 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q){
379     RateControlContext *rcc= &s->rc_context;
380     AVCodecContext *a= s->avctx;
381     const int pict_type= rce->new_pict_type;
382     const double last_p_q    = rcc->last_qscale_for[P_TYPE];
383     const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
384
385     if     (pict_type==I_TYPE && (a->i_quant_factor>0.0 || rcc->last_non_b_pict_type==P_TYPE))
386         q= last_p_q    *FFABS(a->i_quant_factor) + a->i_quant_offset;
387     else if(pict_type==B_TYPE && a->b_quant_factor>0.0)
388         q= last_non_b_q*    a->b_quant_factor  + a->b_quant_offset;
389
390     /* last qscale / qdiff stuff */
391     if(rcc->last_non_b_pict_type==pict_type || pict_type!=I_TYPE){
392         double last_q= rcc->last_qscale_for[pict_type];
393         const int maxdiff= FF_QP2LAMBDA * a->max_qdiff;
394
395         if     (q > last_q + maxdiff) q= last_q + maxdiff;
396         else if(q < last_q - maxdiff) q= last_q - maxdiff;
397     }
398
399     rcc->last_qscale_for[pict_type]= q; //Note we cannot do that after blurring
400
401     if(pict_type!=B_TYPE)
402         rcc->last_non_b_pict_type= pict_type;
403
404     return q;
405 }
406
407 /**
408  * gets the qmin & qmax for pict_type
409  */
410 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type){
411     int qmin= s->avctx->lmin;
412     int qmax= s->avctx->lmax;
413
414     assert(qmin <= qmax);
415
416     if(pict_type==B_TYPE){
417         qmin= (int)(qmin*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
418         qmax= (int)(qmax*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
419     }else if(pict_type==I_TYPE){
420         qmin= (int)(qmin*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
421         qmax= (int)(qmax*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
422     }
423
424     qmin= av_clip(qmin, 1, FF_LAMBDA_MAX);
425     qmax= av_clip(qmax, 1, FF_LAMBDA_MAX);
426
427     if(qmax<qmin) qmax= qmin;
428
429     *qmin_ret= qmin;
430     *qmax_ret= qmax;
431 }
432
433 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce, double q, int frame_num){
434     RateControlContext *rcc= &s->rc_context;
435     int qmin, qmax;
436     double bits;
437     const int pict_type= rce->new_pict_type;
438     const double buffer_size= s->avctx->rc_buffer_size;
439     const double fps= 1/av_q2d(s->avctx->time_base);
440     const double min_rate= s->avctx->rc_min_rate / fps;
441     const double max_rate= s->avctx->rc_max_rate / fps;
442
443     get_qminmax(&qmin, &qmax, s, pict_type);
444
445     /* modulation */
446     if(s->avctx->rc_qmod_freq && frame_num%s->avctx->rc_qmod_freq==0 && pict_type==P_TYPE)
447         q*= s->avctx->rc_qmod_amp;
448
449     bits= qp2bits(rce, q);
450 //printf("q:%f\n", q);
451     /* buffer overflow/underflow protection */
452     if(buffer_size){
453         double expected_size= rcc->buffer_index;
454         double q_limit;
455
456         if(min_rate){
457             double d= 2*(buffer_size - expected_size)/buffer_size;
458             if(d>1.0) d=1.0;
459             else if(d<0.0001) d=0.0001;
460             q*= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
461
462             q_limit= bits2qp(rce, FFMAX((min_rate - buffer_size + rcc->buffer_index)*3, 1));
463             if(q > q_limit){
464                 if(s->avctx->debug&FF_DEBUG_RC){
465                     av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
466                 }
467                 q= q_limit;
468             }
469         }
470
471         if(max_rate){
472             double d= 2*expected_size/buffer_size;
473             if(d>1.0) d=1.0;
474             else if(d<0.0001) d=0.0001;
475             q/= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
476
477             q_limit= bits2qp(rce, FFMAX(rcc->buffer_index/3, 1));
478             if(q < q_limit){
479                 if(s->avctx->debug&FF_DEBUG_RC){
480                     av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
481                 }
482                 q= q_limit;
483             }
484         }
485     }
486 //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);
487     if(s->avctx->rc_qsquish==0.0 || qmin==qmax){
488         if     (q<qmin) q=qmin;
489         else if(q>qmax) q=qmax;
490     }else{
491         double min2= log(qmin);
492         double max2= log(qmax);
493
494         q= log(q);
495         q= (q - min2)/(max2-min2) - 0.5;
496         q*= -4.0;
497         q= 1.0/(1.0 + exp(q));
498         q= q*(max2-min2) + min2;
499
500         q= exp(q);
501     }
502
503     return q;
504 }
505
506 //----------------------------------
507 // 1 Pass Code
508
509 static double predict_size(Predictor *p, double q, double var)
510 {
511      return p->coeff*var / (q*p->count);
512 }
513
514 /*
515 static double predict_qp(Predictor *p, double size, double var)
516 {
517 //printf("coeff:%f, count:%f, var:%f, size:%f//\n", p->coeff, p->count, var, size);
518      return p->coeff*var / (size*p->count);
519 }
520 */
521
522 static void update_predictor(Predictor *p, double q, double var, double size)
523 {
524     double new_coeff= size*q / (var + 1);
525     if(var<10) return;
526
527     p->count*= p->decay;
528     p->coeff*= p->decay;
529     p->count++;
530     p->coeff+= new_coeff;
531 }
532
533 static void adaptive_quantization(MpegEncContext *s, double q){
534     int i;
535     const float lumi_masking= s->avctx->lumi_masking / (128.0*128.0);
536     const float dark_masking= s->avctx->dark_masking / (128.0*128.0);
537     const float temp_cplx_masking= s->avctx->temporal_cplx_masking;
538     const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
539     const float p_masking = s->avctx->p_masking;
540     const float border_masking = s->avctx->border_masking;
541     float bits_sum= 0.0;
542     float cplx_sum= 0.0;
543     float cplx_tab[s->mb_num];
544     float bits_tab[s->mb_num];
545     const int qmin= s->avctx->mb_lmin;
546     const int qmax= s->avctx->mb_lmax;
547     Picture * const pic= &s->current_picture;
548     const int mb_width = s->mb_width;
549     const int mb_height = s->mb_height;
550
551     for(i=0; i<s->mb_num; i++){
552         const int mb_xy= s->mb_index2xy[i];
553         float temp_cplx= sqrt(pic->mc_mb_var[mb_xy]); //FIXME merge in pow()
554         float spat_cplx= sqrt(pic->mb_var[mb_xy]);
555         const int lumi= pic->mb_mean[mb_xy];
556         float bits, cplx, factor;
557         int mb_x = mb_xy % s->mb_stride;
558         int mb_y = mb_xy / s->mb_stride;
559         int mb_distance;
560         float mb_factor = 0.0;
561 #if 0
562         if(spat_cplx < q/3) spat_cplx= q/3; //FIXME finetune
563         if(temp_cplx < q/3) temp_cplx= q/3; //FIXME finetune
564 #endif
565         if(spat_cplx < 4) spat_cplx= 4; //FIXME finetune
566         if(temp_cplx < 4) temp_cplx= 4; //FIXME finetune
567
568         if((s->mb_type[mb_xy]&CANDIDATE_MB_TYPE_INTRA)){//FIXME hq mode
569             cplx= spat_cplx;
570             factor= 1.0 + p_masking;
571         }else{
572             cplx= temp_cplx;
573             factor= pow(temp_cplx, - temp_cplx_masking);
574         }
575         factor*=pow(spat_cplx, - spatial_cplx_masking);
576
577         if(lumi>127)
578             factor*= (1.0 - (lumi-128)*(lumi-128)*lumi_masking);
579         else
580             factor*= (1.0 - (lumi-128)*(lumi-128)*dark_masking);
581
582         if(mb_x < mb_width/5){
583             mb_distance = mb_width/5 - mb_x;
584             mb_factor = (float)mb_distance / (float)(mb_width/5);
585         }else if(mb_x > 4*mb_width/5){
586             mb_distance = mb_x - 4*mb_width/5;
587             mb_factor = (float)mb_distance / (float)(mb_width/5);
588         }
589         if(mb_y < mb_height/5){
590             mb_distance = mb_height/5 - mb_y;
591             mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
592         }else if(mb_y > 4*mb_height/5){
593             mb_distance = mb_y - 4*mb_height/5;
594             mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
595         }
596
597         factor*= 1.0 - border_masking*mb_factor;
598
599         if(factor<0.00001) factor= 0.00001;
600
601         bits= cplx*factor;
602         cplx_sum+= cplx;
603         bits_sum+= bits;
604         cplx_tab[i]= cplx;
605         bits_tab[i]= bits;
606     }
607
608     /* handle qmin/qmax clipping */
609     if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
610         float factor= bits_sum/cplx_sum;
611         for(i=0; i<s->mb_num; i++){
612             float newq= q*cplx_tab[i]/bits_tab[i];
613             newq*= factor;
614
615             if     (newq > qmax){
616                 bits_sum -= bits_tab[i];
617                 cplx_sum -= cplx_tab[i]*q/qmax;
618             }
619             else if(newq < qmin){
620                 bits_sum -= bits_tab[i];
621                 cplx_sum -= cplx_tab[i]*q/qmin;
622             }
623         }
624         if(bits_sum < 0.001) bits_sum= 0.001;
625         if(cplx_sum < 0.001) cplx_sum= 0.001;
626     }
627
628     for(i=0; i<s->mb_num; i++){
629         const int mb_xy= s->mb_index2xy[i];
630         float newq= q*cplx_tab[i]/bits_tab[i];
631         int intq;
632
633         if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
634             newq*= bits_sum/cplx_sum;
635         }
636
637         intq= (int)(newq + 0.5);
638
639         if     (intq > qmax) intq= qmax;
640         else if(intq < qmin) intq= qmin;
641 //if(i%s->mb_width==0) printf("\n");
642 //printf("%2d%3d ", intq, ff_sqrt(s->mc_mb_var[i]));
643         s->lambda_table[mb_xy]= intq;
644     }
645 }
646
647 void ff_get_2pass_fcode(MpegEncContext *s){
648     RateControlContext *rcc= &s->rc_context;
649     int picture_number= s->picture_number;
650     RateControlEntry *rce;
651
652     rce= &rcc->entry[picture_number];
653     s->f_code= rce->f_code;
654     s->b_code= rce->b_code;
655 }
656
657 //FIXME rd or at least approx for dquant
658
659 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
660 {
661     float q;
662     int qmin, qmax;
663     float br_compensation;
664     double diff;
665     double short_term_q;
666     double fps;
667     int picture_number= s->picture_number;
668     int64_t wanted_bits;
669     RateControlContext *rcc= &s->rc_context;
670     AVCodecContext *a= s->avctx;
671     RateControlEntry local_rce, *rce;
672     double bits;
673     double rate_factor;
674     int var;
675     const int pict_type= s->pict_type;
676     Picture * const pic= &s->current_picture;
677     emms_c();
678
679 #ifdef CONFIG_LIBXVID
680     if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
681         return ff_xvid_rate_estimate_qscale(s, dry_run);
682 #endif
683
684     get_qminmax(&qmin, &qmax, s, pict_type);
685
686     fps= 1/av_q2d(s->avctx->time_base);
687 //printf("input_pic_num:%d pic_num:%d frame_rate:%d\n", s->input_picture_number, s->picture_number, s->frame_rate);
688         /* update predictors */
689     if(picture_number>2 && !dry_run){
690         const int last_var= s->last_pict_type == I_TYPE ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum;
691         update_predictor(&rcc->pred[s->last_pict_type], rcc->last_qscale, sqrt(last_var), s->frame_bits);
692     }
693
694     if(s->flags&CODEC_FLAG_PASS2){
695         assert(picture_number>=0);
696         assert(picture_number<rcc->num_entries);
697         rce= &rcc->entry[picture_number];
698         wanted_bits= rce->expected_bits;
699     }else{
700         rce= &local_rce;
701         wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps);
702     }
703
704     diff= s->total_bits - wanted_bits;
705     br_compensation= (a->bit_rate_tolerance - diff)/a->bit_rate_tolerance;
706     if(br_compensation<=0.0) br_compensation=0.001;
707
708     var= pict_type == I_TYPE ? pic->mb_var_sum : pic->mc_mb_var_sum;
709
710     short_term_q = 0; /* avoid warning */
711     if(s->flags&CODEC_FLAG_PASS2){
712         if(pict_type!=I_TYPE)
713             assert(pict_type == rce->new_pict_type);
714
715         q= rce->new_qscale / br_compensation;
716 //printf("%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale, br_compensation, s->frame_bits, var, pict_type);
717     }else{
718         rce->pict_type=
719         rce->new_pict_type= pict_type;
720         rce->mc_mb_var_sum= pic->mc_mb_var_sum;
721         rce->mb_var_sum   = pic->   mb_var_sum;
722         rce->qscale   = FF_QP2LAMBDA * 2;
723         rce->f_code   = s->f_code;
724         rce->b_code   = s->b_code;
725         rce->misc_bits= 1;
726
727         bits= predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
728         if(pict_type== I_TYPE){
729             rce->i_count   = s->mb_num;
730             rce->i_tex_bits= bits;
731             rce->p_tex_bits= 0;
732             rce->mv_bits= 0;
733         }else{
734             rce->i_count   = 0; //FIXME we do know this approx
735             rce->i_tex_bits= 0;
736             rce->p_tex_bits= bits*0.9;
737
738             rce->mv_bits= bits*0.1;
739         }
740         rcc->i_cplx_sum [pict_type] += rce->i_tex_bits*rce->qscale;
741         rcc->p_cplx_sum [pict_type] += rce->p_tex_bits*rce->qscale;
742         rcc->mv_bits_sum[pict_type] += rce->mv_bits;
743         rcc->frame_count[pict_type] ++;
744
745         bits= rce->i_tex_bits + rce->p_tex_bits;
746         rate_factor= rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum * br_compensation;
747
748         q= get_qscale(s, rce, rate_factor, picture_number);
749         if (q < 0)
750             return -1;
751
752         assert(q>0.0);
753 //printf("%f ", q);
754         q= get_diff_limited_q(s, rce, q);
755 //printf("%f ", q);
756         assert(q>0.0);
757
758         if(pict_type==P_TYPE || s->intra_only){ //FIXME type dependent blur like in 2-pass
759             rcc->short_term_qsum*=a->qblur;
760             rcc->short_term_qcount*=a->qblur;
761
762             rcc->short_term_qsum+= q;
763             rcc->short_term_qcount++;
764 //printf("%f ", q);
765             q= short_term_q= rcc->short_term_qsum/rcc->short_term_qcount;
766 //printf("%f ", q);
767         }
768         assert(q>0.0);
769
770         q= modify_qscale(s, rce, q, picture_number);
771
772         rcc->pass1_wanted_bits+= s->bit_rate/fps;
773
774         assert(q>0.0);
775     }
776
777     if(s->avctx->debug&FF_DEBUG_RC){
778         av_log(s->avctx, AV_LOG_DEBUG, "%c qp:%d<%2.1f<%d %d want:%d total:%d comp:%f st_q:%2.2f size:%d var:%d/%d br:%d fps:%d\n",
779         av_get_pict_type_char(pict_type), qmin, q, qmax, picture_number, (int)wanted_bits/1000, (int)s->total_bits/1000,
780         br_compensation, short_term_q, s->frame_bits, pic->mb_var_sum, pic->mc_mb_var_sum, s->bit_rate/1000, (int)fps
781         );
782     }
783
784     if     (q<qmin) q=qmin;
785     else if(q>qmax) q=qmax;
786
787     if(s->adaptive_quant)
788         adaptive_quantization(s, q);
789     else
790         q= (int)(q + 0.5);
791
792     if(!dry_run){
793         rcc->last_qscale= q;
794         rcc->last_mc_mb_var_sum= pic->mc_mb_var_sum;
795         rcc->last_mb_var_sum= pic->mb_var_sum;
796     }
797 #if 0
798 {
799     static int mvsum=0, texsum=0;
800     mvsum += s->mv_bits;
801     texsum += s->i_tex_bits + s->p_tex_bits;
802     printf("%d %d//\n\n", mvsum, texsum);
803 }
804 #endif
805     return q;
806 }
807
808 //----------------------------------------------
809 // 2-Pass code
810
811 static int init_pass2(MpegEncContext *s)
812 {
813     RateControlContext *rcc= &s->rc_context;
814     AVCodecContext *a= s->avctx;
815     int i, toobig;
816     double fps= 1/av_q2d(s->avctx->time_base);
817     double complexity[5]={0,0,0,0,0};   // aproximate bits at quant=1
818     uint64_t const_bits[5]={0,0,0,0,0}; // quantizer independent bits
819     uint64_t all_const_bits;
820     uint64_t all_available_bits= (uint64_t)(s->bit_rate*(double)rcc->num_entries/fps);
821     double rate_factor=0;
822     double step;
823     //int last_i_frame=-10000000;
824     const int filter_size= (int)(a->qblur*4) | 1;
825     double expected_bits;
826     double *qscale, *blured_qscale, qscale_sum;
827
828     /* find complexity & const_bits & decide the pict_types */
829     for(i=0; i<rcc->num_entries; i++){
830         RateControlEntry *rce= &rcc->entry[i];
831
832         rce->new_pict_type= rce->pict_type;
833         rcc->i_cplx_sum [rce->pict_type] += rce->i_tex_bits*rce->qscale;
834         rcc->p_cplx_sum [rce->pict_type] += rce->p_tex_bits*rce->qscale;
835         rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
836         rcc->frame_count[rce->pict_type] ++;
837
838         complexity[rce->new_pict_type]+= (rce->i_tex_bits+ rce->p_tex_bits)*(double)rce->qscale;
839         const_bits[rce->new_pict_type]+= rce->mv_bits + rce->misc_bits;
840     }
841     all_const_bits= const_bits[I_TYPE] + const_bits[P_TYPE] + const_bits[B_TYPE];
842
843     if(all_available_bits < all_const_bits){
844         av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
845         return -1;
846     }
847
848     qscale= av_malloc(sizeof(double)*rcc->num_entries);
849     blured_qscale= av_malloc(sizeof(double)*rcc->num_entries);
850     toobig = 0;
851
852     for(step=256*256; step>0.0000001; step*=0.5){
853         expected_bits=0;
854         rate_factor+= step;
855
856         rcc->buffer_index= s->avctx->rc_buffer_size/2;
857
858         /* find qscale */
859         for(i=0; i<rcc->num_entries; i++){
860             qscale[i]= get_qscale(s, &rcc->entry[i], rate_factor, i);
861         }
862         assert(filter_size%2==1);
863
864         /* fixed I/B QP relative to P mode */
865         for(i=rcc->num_entries-1; i>=0; i--){
866             RateControlEntry *rce= &rcc->entry[i];
867
868             qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
869         }
870
871         /* smooth curve */
872         for(i=0; i<rcc->num_entries; i++){
873             RateControlEntry *rce= &rcc->entry[i];
874             const int pict_type= rce->new_pict_type;
875             int j;
876             double q=0.0, sum=0.0;
877
878             for(j=0; j<filter_size; j++){
879                 int index= i+j-filter_size/2;
880                 double d= index-i;
881                 double coeff= a->qblur==0 ? 1.0 : exp(-d*d/(a->qblur * a->qblur));
882
883                 if(index < 0 || index >= rcc->num_entries) continue;
884                 if(pict_type != rcc->entry[index].new_pict_type) continue;
885                 q+= qscale[index] * coeff;
886                 sum+= coeff;
887             }
888             blured_qscale[i]= q/sum;
889         }
890
891         /* find expected bits */
892         for(i=0; i<rcc->num_entries; i++){
893             RateControlEntry *rce= &rcc->entry[i];
894             double bits;
895             rce->new_qscale= modify_qscale(s, rce, blured_qscale[i], i);
896             bits= qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
897 //printf("%d %f\n", rce->new_bits, blured_qscale[i]);
898             bits += 8*ff_vbv_update(s, bits);
899
900             rce->expected_bits= expected_bits;
901             expected_bits += bits;
902         }
903
904         /*
905         av_log(s->avctx, AV_LOG_INFO,
906             "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
907             expected_bits, (int)all_available_bits, rate_factor);
908         */
909         if(expected_bits > all_available_bits) {
910             rate_factor-= step;
911             ++toobig;
912         }
913     }
914     av_free(qscale);
915     av_free(blured_qscale);
916
917     /* check bitrate calculations and print info */
918     qscale_sum = 0.0;
919     for(i=0; i<rcc->num_entries; i++){
920         /* av_log(s->avctx, AV_LOG_DEBUG, "[lavc rc] entry[%d].new_qscale = %.3f  qp = %.3f\n",
921             i, rcc->entry[i].new_qscale, rcc->entry[i].new_qscale / FF_QP2LAMBDA); */
922         qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA, s->avctx->qmin, s->avctx->qmax);
923     }
924     assert(toobig <= 40);
925     av_log(s->avctx, AV_LOG_DEBUG,
926         "[lavc rc] requested bitrate: %d bps  expected bitrate: %d bps\n",
927         s->bit_rate,
928         (int)(expected_bits / ((double)all_available_bits/s->bit_rate)));
929     av_log(s->avctx, AV_LOG_DEBUG,
930         "[lavc rc] estimated target average qp: %.3f\n",
931         (float)qscale_sum / rcc->num_entries);
932     if (toobig == 0) {
933         av_log(s->avctx, AV_LOG_INFO,
934             "[lavc rc] Using all of requested bitrate is not "
935             "necessary for this video with these parameters.\n");
936     } else if (toobig == 40) {
937         av_log(s->avctx, AV_LOG_ERROR,
938             "[lavc rc] Error: bitrate too low for this video "
939             "with these parameters.\n");
940         return -1;
941     } else if (fabs(expected_bits/all_available_bits - 1.0) > 0.01) {
942         av_log(s->avctx, AV_LOG_ERROR,
943             "[lavc rc] Error: 2pass curve failed to converge\n");
944         return -1;
945     }
946
947     return 0;
948 }