Add av_ prefix to clip functions
[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 // allways 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 unbelieavle slow on looong strings //FIXME copy / dont 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_XVID
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         /* init stuff with the user specified complexity */
205         if(s->avctx->rc_initial_cplx){
206             for(i=0; i<60*30; i++){
207                 double bits= s->avctx->rc_initial_cplx * (i/10000.0 + 1.0)*s->mb_num;
208                 RateControlEntry rce;
209                 double q;
210
211                 if     (i%((s->gop_size+3)/4)==0) rce.pict_type= I_TYPE;
212                 else if(i%(s->max_b_frames+1))    rce.pict_type= B_TYPE;
213                 else                              rce.pict_type= P_TYPE;
214
215                 rce.new_pict_type= rce.pict_type;
216                 rce.mc_mb_var_sum= bits*s->mb_num/100000;
217                 rce.mb_var_sum   = s->mb_num;
218                 rce.qscale   = FF_QP2LAMBDA * 2;
219                 rce.f_code   = 2;
220                 rce.b_code   = 1;
221                 rce.misc_bits= 1;
222
223                 if(s->pict_type== I_TYPE){
224                     rce.i_count   = s->mb_num;
225                     rce.i_tex_bits= bits;
226                     rce.p_tex_bits= 0;
227                     rce.mv_bits= 0;
228                 }else{
229                     rce.i_count   = 0; //FIXME we do know this approx
230                     rce.i_tex_bits= 0;
231                     rce.p_tex_bits= bits*0.9;
232                     rce.mv_bits= bits*0.1;
233                 }
234                 rcc->i_cplx_sum [rce.pict_type] += rce.i_tex_bits*rce.qscale;
235                 rcc->p_cplx_sum [rce.pict_type] += rce.p_tex_bits*rce.qscale;
236                 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
237                 rcc->frame_count[rce.pict_type] ++;
238
239                 bits= rce.i_tex_bits + rce.p_tex_bits;
240
241                 q= get_qscale(s, &rce, rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum, i);
242                 rcc->pass1_wanted_bits+= s->bit_rate/(1/av_q2d(s->avctx->time_base)); //FIXME missbehaves a little for variable fps
243             }
244         }
245
246     }
247
248     return 0;
249 }
250
251 void ff_rate_control_uninit(MpegEncContext *s)
252 {
253     RateControlContext *rcc= &s->rc_context;
254     emms_c();
255
256     ff_eval_free(rcc->rc_eq_eval);
257     av_freep(&rcc->entry);
258
259 #ifdef CONFIG_XVID
260     if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
261         ff_xvid_rate_control_uninit(s);
262 #endif
263 }
264
265 int ff_vbv_update(MpegEncContext *s, int frame_size){
266     RateControlContext *rcc= &s->rc_context;
267     const double fps= 1/av_q2d(s->avctx->time_base);
268     const int buffer_size= s->avctx->rc_buffer_size;
269     const double min_rate= s->avctx->rc_min_rate/fps;
270     const double max_rate= s->avctx->rc_max_rate/fps;
271
272 //printf("%d %f %d %f %f\n", buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
273     if(buffer_size){
274         int left;
275
276         rcc->buffer_index-= frame_size;
277         if(rcc->buffer_index < 0){
278             av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
279             rcc->buffer_index= 0;
280         }
281
282         left= buffer_size - rcc->buffer_index - 1;
283         rcc->buffer_index += av_clip(left, min_rate, max_rate);
284
285         if(rcc->buffer_index > buffer_size){
286             int stuffing= ceil((rcc->buffer_index - buffer_size)/8);
287
288             if(stuffing < 4 && s->codec_id == CODEC_ID_MPEG4)
289                 stuffing=4;
290             rcc->buffer_index -= 8*stuffing;
291
292             if(s->avctx->debug & FF_DEBUG_RC)
293                 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
294
295             return stuffing;
296         }
297     }
298     return 0;
299 }
300
301 /**
302  * modifies the bitrate curve from pass1 for one frame
303  */
304 static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num){
305     RateControlContext *rcc= &s->rc_context;
306     AVCodecContext *a= s->avctx;
307     double q, bits;
308     const int pict_type= rce->new_pict_type;
309     const double mb_num= s->mb_num;
310     int i;
311
312     double const_values[]={
313         M_PI,
314         M_E,
315         rce->i_tex_bits*rce->qscale,
316         rce->p_tex_bits*rce->qscale,
317         (rce->i_tex_bits + rce->p_tex_bits)*(double)rce->qscale,
318         rce->mv_bits/mb_num,
319         rce->pict_type == B_TYPE ? (rce->f_code + rce->b_code)*0.5 : rce->f_code,
320         rce->i_count/mb_num,
321         rce->mc_mb_var_sum/mb_num,
322         rce->mb_var_sum/mb_num,
323         rce->pict_type == I_TYPE,
324         rce->pict_type == P_TYPE,
325         rce->pict_type == B_TYPE,
326         rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
327         a->qcompress,
328 /*        rcc->last_qscale_for[I_TYPE],
329         rcc->last_qscale_for[P_TYPE],
330         rcc->last_qscale_for[B_TYPE],
331         rcc->next_non_b_qscale,*/
332         rcc->i_cplx_sum[I_TYPE] / (double)rcc->frame_count[I_TYPE],
333         rcc->i_cplx_sum[P_TYPE] / (double)rcc->frame_count[P_TYPE],
334         rcc->p_cplx_sum[P_TYPE] / (double)rcc->frame_count[P_TYPE],
335         rcc->p_cplx_sum[B_TYPE] / (double)rcc->frame_count[B_TYPE],
336         (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
337         0
338     };
339
340     bits= ff_parse_eval(rcc->rc_eq_eval, const_values, rce);
341     if (isnan(bits)) {
342         av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq);
343         return -1;
344     }
345
346     rcc->pass1_rc_eq_output_sum+= bits;
347     bits*=rate_factor;
348     if(bits<0.0) bits=0.0;
349     bits+= 1.0; //avoid 1/0 issues
350
351     /* user override */
352     for(i=0; i<s->avctx->rc_override_count; i++){
353         RcOverride *rco= s->avctx->rc_override;
354         if(rco[i].start_frame > frame_num) continue;
355         if(rco[i].end_frame   < frame_num) continue;
356
357         if(rco[i].qscale)
358             bits= qp2bits(rce, rco[i].qscale); //FIXME move at end to really force it?
359         else
360             bits*= rco[i].quality_factor;
361     }
362
363     q= bits2qp(rce, bits);
364
365     /* I/B difference */
366     if     (pict_type==I_TYPE && s->avctx->i_quant_factor<0.0)
367         q= -q*s->avctx->i_quant_factor + s->avctx->i_quant_offset;
368     else if(pict_type==B_TYPE && s->avctx->b_quant_factor<0.0)
369         q= -q*s->avctx->b_quant_factor + s->avctx->b_quant_offset;
370
371     return q;
372 }
373
374 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q){
375     RateControlContext *rcc= &s->rc_context;
376     AVCodecContext *a= s->avctx;
377     const int pict_type= rce->new_pict_type;
378     const double last_p_q    = rcc->last_qscale_for[P_TYPE];
379     const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
380
381     if     (pict_type==I_TYPE && (a->i_quant_factor>0.0 || rcc->last_non_b_pict_type==P_TYPE))
382         q= last_p_q    *FFABS(a->i_quant_factor) + a->i_quant_offset;
383     else if(pict_type==B_TYPE && a->b_quant_factor>0.0)
384         q= last_non_b_q*    a->b_quant_factor  + a->b_quant_offset;
385
386     /* last qscale / qdiff stuff */
387     if(rcc->last_non_b_pict_type==pict_type || pict_type!=I_TYPE){
388         double last_q= rcc->last_qscale_for[pict_type];
389         const int maxdiff= FF_QP2LAMBDA * a->max_qdiff;
390
391         if     (q > last_q + maxdiff) q= last_q + maxdiff;
392         else if(q < last_q - maxdiff) q= last_q - maxdiff;
393     }
394
395     rcc->last_qscale_for[pict_type]= q; //Note we cant do that after blurring
396
397     if(pict_type!=B_TYPE)
398         rcc->last_non_b_pict_type= pict_type;
399
400     return q;
401 }
402
403 /**
404  * gets the qmin & qmax for pict_type
405  */
406 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type){
407     int qmin= s->avctx->lmin;
408     int qmax= s->avctx->lmax;
409
410     assert(qmin <= qmax);
411
412     if(pict_type==B_TYPE){
413         qmin= (int)(qmin*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
414         qmax= (int)(qmax*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
415     }else if(pict_type==I_TYPE){
416         qmin= (int)(qmin*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
417         qmax= (int)(qmax*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
418     }
419
420     qmin= av_clip(qmin, 1, FF_LAMBDA_MAX);
421     qmax= av_clip(qmax, 1, FF_LAMBDA_MAX);
422
423     if(qmax<qmin) qmax= qmin;
424
425     *qmin_ret= qmin;
426     *qmax_ret= qmax;
427 }
428
429 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce, double q, int frame_num){
430     RateControlContext *rcc= &s->rc_context;
431     int qmin, qmax;
432     double bits;
433     const int pict_type= rce->new_pict_type;
434     const double buffer_size= s->avctx->rc_buffer_size;
435     const double fps= 1/av_q2d(s->avctx->time_base);
436     const double min_rate= s->avctx->rc_min_rate / fps;
437     const double max_rate= s->avctx->rc_max_rate / fps;
438
439     get_qminmax(&qmin, &qmax, s, pict_type);
440
441     /* modulation */
442     if(s->avctx->rc_qmod_freq && frame_num%s->avctx->rc_qmod_freq==0 && pict_type==P_TYPE)
443         q*= s->avctx->rc_qmod_amp;
444
445     bits= qp2bits(rce, q);
446 //printf("q:%f\n", q);
447     /* buffer overflow/underflow protection */
448     if(buffer_size){
449         double expected_size= rcc->buffer_index;
450         double q_limit;
451
452         if(min_rate){
453             double d= 2*(buffer_size - expected_size)/buffer_size;
454             if(d>1.0) d=1.0;
455             else if(d<0.0001) d=0.0001;
456             q*= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
457
458             q_limit= bits2qp(rce, FFMAX((min_rate - buffer_size + rcc->buffer_index)*3, 1));
459             if(q > q_limit){
460                 if(s->avctx->debug&FF_DEBUG_RC){
461                     av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
462                 }
463                 q= q_limit;
464             }
465         }
466
467         if(max_rate){
468             double d= 2*expected_size/buffer_size;
469             if(d>1.0) d=1.0;
470             else if(d<0.0001) d=0.0001;
471             q/= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
472
473             q_limit= bits2qp(rce, FFMAX(rcc->buffer_index/3, 1));
474             if(q < q_limit){
475                 if(s->avctx->debug&FF_DEBUG_RC){
476                     av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
477                 }
478                 q= q_limit;
479             }
480         }
481     }
482 //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);
483     if(s->avctx->rc_qsquish==0.0 || qmin==qmax){
484         if     (q<qmin) q=qmin;
485         else if(q>qmax) q=qmax;
486     }else{
487         double min2= log(qmin);
488         double max2= log(qmax);
489
490         q= log(q);
491         q= (q - min2)/(max2-min2) - 0.5;
492         q*= -4.0;
493         q= 1.0/(1.0 + exp(q));
494         q= q*(max2-min2) + min2;
495
496         q= exp(q);
497     }
498
499     return q;
500 }
501
502 //----------------------------------
503 // 1 Pass Code
504
505 static double predict_size(Predictor *p, double q, double var)
506 {
507      return p->coeff*var / (q*p->count);
508 }
509
510 /*
511 static double predict_qp(Predictor *p, double size, double var)
512 {
513 //printf("coeff:%f, count:%f, var:%f, size:%f//\n", p->coeff, p->count, var, size);
514      return p->coeff*var / (size*p->count);
515 }
516 */
517
518 static void update_predictor(Predictor *p, double q, double var, double size)
519 {
520     double new_coeff= size*q / (var + 1);
521     if(var<10) return;
522
523     p->count*= p->decay;
524     p->coeff*= p->decay;
525     p->count++;
526     p->coeff+= new_coeff;
527 }
528
529 static void adaptive_quantization(MpegEncContext *s, double q){
530     int i;
531     const float lumi_masking= s->avctx->lumi_masking / (128.0*128.0);
532     const float dark_masking= s->avctx->dark_masking / (128.0*128.0);
533     const float temp_cplx_masking= s->avctx->temporal_cplx_masking;
534     const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
535     const float p_masking = s->avctx->p_masking;
536     const float border_masking = s->avctx->border_masking;
537     float bits_sum= 0.0;
538     float cplx_sum= 0.0;
539     float cplx_tab[s->mb_num];
540     float bits_tab[s->mb_num];
541     const int qmin= s->avctx->mb_lmin;
542     const int qmax= s->avctx->mb_lmax;
543     Picture * const pic= &s->current_picture;
544     const int mb_width = s->mb_width;
545     const int mb_height = s->mb_height;
546
547     for(i=0; i<s->mb_num; i++){
548         const int mb_xy= s->mb_index2xy[i];
549         float temp_cplx= sqrt(pic->mc_mb_var[mb_xy]); //FIXME merge in pow()
550         float spat_cplx= sqrt(pic->mb_var[mb_xy]);
551         const int lumi= pic->mb_mean[mb_xy];
552         float bits, cplx, factor;
553         int mb_x = mb_xy % s->mb_stride;
554         int mb_y = mb_xy / s->mb_stride;
555         int mb_distance;
556         float mb_factor = 0.0;
557 #if 0
558         if(spat_cplx < q/3) spat_cplx= q/3; //FIXME finetune
559         if(temp_cplx < q/3) temp_cplx= q/3; //FIXME finetune
560 #endif
561         if(spat_cplx < 4) spat_cplx= 4; //FIXME finetune
562         if(temp_cplx < 4) temp_cplx= 4; //FIXME finetune
563
564         if((s->mb_type[mb_xy]&CANDIDATE_MB_TYPE_INTRA)){//FIXME hq mode
565             cplx= spat_cplx;
566             factor= 1.0 + p_masking;
567         }else{
568             cplx= temp_cplx;
569             factor= pow(temp_cplx, - temp_cplx_masking);
570         }
571         factor*=pow(spat_cplx, - spatial_cplx_masking);
572
573         if(lumi>127)
574             factor*= (1.0 - (lumi-128)*(lumi-128)*lumi_masking);
575         else
576             factor*= (1.0 - (lumi-128)*(lumi-128)*dark_masking);
577
578         if(mb_x < mb_width/5){
579             mb_distance = mb_width/5 - mb_x;
580             mb_factor = (float)mb_distance / (float)(mb_width/5);
581         }else if(mb_x > 4*mb_width/5){
582             mb_distance = mb_x - 4*mb_width/5;
583             mb_factor = (float)mb_distance / (float)(mb_width/5);
584         }
585         if(mb_y < mb_height/5){
586             mb_distance = mb_height/5 - mb_y;
587             mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
588         }else if(mb_y > 4*mb_height/5){
589             mb_distance = mb_y - 4*mb_height/5;
590             mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
591         }
592
593         factor*= 1.0 - border_masking*mb_factor;
594
595         if(factor<0.00001) factor= 0.00001;
596
597         bits= cplx*factor;
598         cplx_sum+= cplx;
599         bits_sum+= bits;
600         cplx_tab[i]= cplx;
601         bits_tab[i]= bits;
602     }
603
604     /* handle qmin/qmax cliping */
605     if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
606         float factor= bits_sum/cplx_sum;
607         for(i=0; i<s->mb_num; i++){
608             float newq= q*cplx_tab[i]/bits_tab[i];
609             newq*= factor;
610
611             if     (newq > qmax){
612                 bits_sum -= bits_tab[i];
613                 cplx_sum -= cplx_tab[i]*q/qmax;
614             }
615             else if(newq < qmin){
616                 bits_sum -= bits_tab[i];
617                 cplx_sum -= cplx_tab[i]*q/qmin;
618             }
619         }
620         if(bits_sum < 0.001) bits_sum= 0.001;
621         if(cplx_sum < 0.001) cplx_sum= 0.001;
622     }
623
624     for(i=0; i<s->mb_num; i++){
625         const int mb_xy= s->mb_index2xy[i];
626         float newq= q*cplx_tab[i]/bits_tab[i];
627         int intq;
628
629         if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
630             newq*= bits_sum/cplx_sum;
631         }
632
633         intq= (int)(newq + 0.5);
634
635         if     (intq > qmax) intq= qmax;
636         else if(intq < qmin) intq= qmin;
637 //if(i%s->mb_width==0) printf("\n");
638 //printf("%2d%3d ", intq, ff_sqrt(s->mc_mb_var[i]));
639         s->lambda_table[mb_xy]= intq;
640     }
641 }
642
643 void ff_get_2pass_fcode(MpegEncContext *s){
644     RateControlContext *rcc= &s->rc_context;
645     int picture_number= s->picture_number;
646     RateControlEntry *rce;
647
648     rce= &rcc->entry[picture_number];
649     s->f_code= rce->f_code;
650     s->b_code= rce->b_code;
651 }
652
653 //FIXME rd or at least approx for dquant
654
655 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
656 {
657     float q;
658     int qmin, qmax;
659     float br_compensation;
660     double diff;
661     double short_term_q;
662     double fps;
663     int picture_number= s->picture_number;
664     int64_t wanted_bits;
665     RateControlContext *rcc= &s->rc_context;
666     AVCodecContext *a= s->avctx;
667     RateControlEntry local_rce, *rce;
668     double bits;
669     double rate_factor;
670     int var;
671     const int pict_type= s->pict_type;
672     Picture * const pic= &s->current_picture;
673     emms_c();
674
675 #ifdef CONFIG_XVID
676     if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
677         return ff_xvid_rate_estimate_qscale(s, dry_run);
678 #endif
679
680     get_qminmax(&qmin, &qmax, s, pict_type);
681
682     fps= 1/av_q2d(s->avctx->time_base);
683 //printf("input_pic_num:%d pic_num:%d frame_rate:%d\n", s->input_picture_number, s->picture_number, s->frame_rate);
684         /* update predictors */
685     if(picture_number>2 && !dry_run){
686         const int last_var= s->last_pict_type == I_TYPE ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum;
687         update_predictor(&rcc->pred[s->last_pict_type], rcc->last_qscale, sqrt(last_var), s->frame_bits);
688     }
689
690     if(s->flags&CODEC_FLAG_PASS2){
691         assert(picture_number>=0);
692         assert(picture_number<rcc->num_entries);
693         rce= &rcc->entry[picture_number];
694         wanted_bits= rce->expected_bits;
695     }else{
696         rce= &local_rce;
697         wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps);
698     }
699
700     diff= s->total_bits - wanted_bits;
701     br_compensation= (a->bit_rate_tolerance - diff)/a->bit_rate_tolerance;
702     if(br_compensation<=0.0) br_compensation=0.001;
703
704     var= pict_type == I_TYPE ? pic->mb_var_sum : pic->mc_mb_var_sum;
705
706     short_term_q = 0; /* avoid warning */
707     if(s->flags&CODEC_FLAG_PASS2){
708         if(pict_type!=I_TYPE)
709             assert(pict_type == rce->new_pict_type);
710
711         q= rce->new_qscale / br_compensation;
712 //printf("%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale, br_compensation, s->frame_bits, var, pict_type);
713     }else{
714         rce->pict_type=
715         rce->new_pict_type= pict_type;
716         rce->mc_mb_var_sum= pic->mc_mb_var_sum;
717         rce->mb_var_sum   = pic->   mb_var_sum;
718         rce->qscale   = FF_QP2LAMBDA * 2;
719         rce->f_code   = s->f_code;
720         rce->b_code   = s->b_code;
721         rce->misc_bits= 1;
722
723         bits= predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
724         if(pict_type== I_TYPE){
725             rce->i_count   = s->mb_num;
726             rce->i_tex_bits= bits;
727             rce->p_tex_bits= 0;
728             rce->mv_bits= 0;
729         }else{
730             rce->i_count   = 0; //FIXME we do know this approx
731             rce->i_tex_bits= 0;
732             rce->p_tex_bits= bits*0.9;
733
734             rce->mv_bits= bits*0.1;
735         }
736         rcc->i_cplx_sum [pict_type] += rce->i_tex_bits*rce->qscale;
737         rcc->p_cplx_sum [pict_type] += rce->p_tex_bits*rce->qscale;
738         rcc->mv_bits_sum[pict_type] += rce->mv_bits;
739         rcc->frame_count[pict_type] ++;
740
741         bits= rce->i_tex_bits + rce->p_tex_bits;
742         rate_factor= rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum * br_compensation;
743
744         q= get_qscale(s, rce, rate_factor, picture_number);
745         if (q < 0)
746             return -1;
747
748         assert(q>0.0);
749 //printf("%f ", q);
750         q= get_diff_limited_q(s, rce, q);
751 //printf("%f ", q);
752         assert(q>0.0);
753
754         if(pict_type==P_TYPE || s->intra_only){ //FIXME type dependent blur like in 2-pass
755             rcc->short_term_qsum*=a->qblur;
756             rcc->short_term_qcount*=a->qblur;
757
758             rcc->short_term_qsum+= q;
759             rcc->short_term_qcount++;
760 //printf("%f ", q);
761             q= short_term_q= rcc->short_term_qsum/rcc->short_term_qcount;
762 //printf("%f ", q);
763         }
764         assert(q>0.0);
765
766         q= modify_qscale(s, rce, q, picture_number);
767
768         rcc->pass1_wanted_bits+= s->bit_rate/fps;
769
770         assert(q>0.0);
771     }
772
773     if(s->avctx->debug&FF_DEBUG_RC){
774         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",
775         av_get_pict_type_char(pict_type), qmin, q, qmax, picture_number, (int)wanted_bits/1000, (int)s->total_bits/1000,
776         br_compensation, short_term_q, s->frame_bits, pic->mb_var_sum, pic->mc_mb_var_sum, s->bit_rate/1000, (int)fps
777         );
778     }
779
780     if     (q<qmin) q=qmin;
781     else if(q>qmax) q=qmax;
782
783     if(s->adaptive_quant)
784         adaptive_quantization(s, q);
785     else
786         q= (int)(q + 0.5);
787
788     if(!dry_run){
789         rcc->last_qscale= q;
790         rcc->last_mc_mb_var_sum= pic->mc_mb_var_sum;
791         rcc->last_mb_var_sum= pic->mb_var_sum;
792     }
793 #if 0
794 {
795     static int mvsum=0, texsum=0;
796     mvsum += s->mv_bits;
797     texsum += s->i_tex_bits + s->p_tex_bits;
798     printf("%d %d//\n\n", mvsum, texsum);
799 }
800 #endif
801     return q;
802 }
803
804 //----------------------------------------------
805 // 2-Pass code
806
807 static int init_pass2(MpegEncContext *s)
808 {
809     RateControlContext *rcc= &s->rc_context;
810     AVCodecContext *a= s->avctx;
811     int i, toobig;
812     double fps= 1/av_q2d(s->avctx->time_base);
813     double complexity[5]={0,0,0,0,0};   // aproximate bits at quant=1
814     uint64_t const_bits[5]={0,0,0,0,0}; // quantizer independent bits
815     uint64_t all_const_bits;
816     uint64_t all_available_bits= (uint64_t)(s->bit_rate*(double)rcc->num_entries/fps);
817     double rate_factor=0;
818     double step;
819     //int last_i_frame=-10000000;
820     const int filter_size= (int)(a->qblur*4) | 1;
821     double expected_bits;
822     double *qscale, *blured_qscale, qscale_sum;
823
824     /* find complexity & const_bits & decide the pict_types */
825     for(i=0; i<rcc->num_entries; i++){
826         RateControlEntry *rce= &rcc->entry[i];
827
828         rce->new_pict_type= rce->pict_type;
829         rcc->i_cplx_sum [rce->pict_type] += rce->i_tex_bits*rce->qscale;
830         rcc->p_cplx_sum [rce->pict_type] += rce->p_tex_bits*rce->qscale;
831         rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
832         rcc->frame_count[rce->pict_type] ++;
833
834         complexity[rce->new_pict_type]+= (rce->i_tex_bits+ rce->p_tex_bits)*(double)rce->qscale;
835         const_bits[rce->new_pict_type]+= rce->mv_bits + rce->misc_bits;
836     }
837     all_const_bits= const_bits[I_TYPE] + const_bits[P_TYPE] + const_bits[B_TYPE];
838
839     if(all_available_bits < all_const_bits){
840         av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
841         return -1;
842     }
843
844     qscale= av_malloc(sizeof(double)*rcc->num_entries);
845     blured_qscale= av_malloc(sizeof(double)*rcc->num_entries);
846     toobig = 0;
847
848     for(step=256*256; step>0.0000001; step*=0.5){
849         expected_bits=0;
850         rate_factor+= step;
851
852         rcc->buffer_index= s->avctx->rc_buffer_size/2;
853
854         /* find qscale */
855         for(i=0; i<rcc->num_entries; i++){
856             qscale[i]= get_qscale(s, &rcc->entry[i], rate_factor, i);
857         }
858         assert(filter_size%2==1);
859
860         /* fixed I/B QP relative to P mode */
861         for(i=rcc->num_entries-1; i>=0; i--){
862             RateControlEntry *rce= &rcc->entry[i];
863
864             qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
865         }
866
867         /* smooth curve */
868         for(i=0; i<rcc->num_entries; i++){
869             RateControlEntry *rce= &rcc->entry[i];
870             const int pict_type= rce->new_pict_type;
871             int j;
872             double q=0.0, sum=0.0;
873
874             for(j=0; j<filter_size; j++){
875                 int index= i+j-filter_size/2;
876                 double d= index-i;
877                 double coeff= a->qblur==0 ? 1.0 : exp(-d*d/(a->qblur * a->qblur));
878
879                 if(index < 0 || index >= rcc->num_entries) continue;
880                 if(pict_type != rcc->entry[index].new_pict_type) continue;
881                 q+= qscale[index] * coeff;
882                 sum+= coeff;
883             }
884             blured_qscale[i]= q/sum;
885         }
886
887         /* find expected bits */
888         for(i=0; i<rcc->num_entries; i++){
889             RateControlEntry *rce= &rcc->entry[i];
890             double bits;
891             rce->new_qscale= modify_qscale(s, rce, blured_qscale[i], i);
892             bits= qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
893 //printf("%d %f\n", rce->new_bits, blured_qscale[i]);
894             bits += 8*ff_vbv_update(s, bits);
895
896             rce->expected_bits= expected_bits;
897             expected_bits += bits;
898         }
899
900         /*
901         av_log(s->avctx, AV_LOG_INFO,
902             "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
903             expected_bits, (int)all_available_bits, rate_factor);
904         */
905         if(expected_bits > all_available_bits) {
906             rate_factor-= step;
907             ++toobig;
908         }
909     }
910     av_free(qscale);
911     av_free(blured_qscale);
912
913     /* check bitrate calculations and print info */
914     qscale_sum = 0.0;
915     for(i=0; i<rcc->num_entries; i++){
916         /* av_log(s->avctx, AV_LOG_DEBUG, "[lavc rc] entry[%d].new_qscale = %.3f  qp = %.3f\n",
917             i, rcc->entry[i].new_qscale, rcc->entry[i].new_qscale / FF_QP2LAMBDA); */
918         qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA, s->avctx->qmin, s->avctx->qmax);
919     }
920     assert(toobig <= 40);
921     av_log(s->avctx, AV_LOG_DEBUG,
922         "[lavc rc] requested bitrate: %d bps  expected bitrate: %d bps\n",
923         s->bit_rate,
924         (int)(expected_bits / ((double)all_available_bits/s->bit_rate)));
925     av_log(s->avctx, AV_LOG_DEBUG,
926         "[lavc rc] estimated target average qp: %.3f\n",
927         (float)qscale_sum / rcc->num_entries);
928     if (toobig == 0) {
929         av_log(s->avctx, AV_LOG_INFO,
930             "[lavc rc] Using all of requested bitrate is not "
931             "necessary for this video with these parameters.\n");
932     } else if (toobig == 40) {
933         av_log(s->avctx, AV_LOG_ERROR,
934             "[lavc rc] Error: bitrate too low for this video "
935             "with these parameters.\n");
936         return -1;
937     } else if (fabs(expected_bits/all_available_bits - 1.0) > 0.01) {
938         av_log(s->avctx, AV_LOG_ERROR,
939             "[lavc rc] Error: 2pass curve failed to converge\n");
940         return -1;
941     }
942
943     return 0;
944 }