alac: make block_size signed
[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 Libav.
7  *
8  * Libav 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  * Libav 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 Libav; 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
25  * Rate control for video encoders.
26  */
27
28 #include "libavutil/intmath.h"
29 #include "avcodec.h"
30 #include "dsputil.h"
31 #include "ratecontrol.h"
32 #include "mpegvideo.h"
33 #include "libavutil/eval.h"
34
35 #undef NDEBUG // Always check asserts, the speed effect is far too small to disable them.
36 #include <assert.h>
37
38 #ifndef M_E
39 #define M_E 2.718281828
40 #endif
41
42 static int init_pass2(MpegEncContext *s);
43 static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num);
44
45 void ff_write_pass1_stats(MpegEncContext *s){
46     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",
47              s->current_picture_ptr->f.display_picture_number, s->current_picture_ptr->f.coded_picture_number, s->pict_type,
48              s->current_picture.f.quality, s->i_tex_bits, s->p_tex_bits, s->mv_bits, s->misc_bits,
49              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);
50 }
51
52 static inline double qp2bits(RateControlEntry *rce, double qp){
53     if(qp<=0.0){
54         av_log(NULL, AV_LOG_ERROR, "qp<=0.0\n");
55     }
56     return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ qp;
57 }
58
59 static inline double bits2qp(RateControlEntry *rce, double bits){
60     if(bits<0.9){
61         av_log(NULL, AV_LOG_ERROR, "bits<0.9\n");
62     }
63     return rce->qscale * (double)(rce->i_tex_bits + rce->p_tex_bits+1)/ bits;
64 }
65
66 int ff_rate_control_init(MpegEncContext *s)
67 {
68     RateControlContext *rcc= &s->rc_context;
69     int i, res;
70     static const char * const 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 (* const func1[])(void *, double)={
98         (void *)bits2qp,
99         (void *)qp2bits,
100         NULL
101     };
102     static const char * const func1_names[]={
103         "bits2qp",
104         "qp2bits",
105         NULL
106     };
107     emms_c();
108
109     res = av_expr_parse(&rcc->rc_eq_eval, s->avctx->rc_eq ? s->avctx->rc_eq : "tex^qComp", const_names, func1_names, func1, NULL, NULL, 0, s->avctx);
110     if (res < 0) {
111         av_log(s->avctx, AV_LOG_ERROR, "Error parsing rc_eq \"%s\"\n", s->avctx->rc_eq);
112         return res;
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 because 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 = 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=AV_PICTURE_TYPE_P;
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 #if 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
214                 if     (i%((s->gop_size+3)/4)==0) rce.pict_type= AV_PICTURE_TYPE_I;
215                 else if(i%(s->max_b_frames+1))    rce.pict_type= AV_PICTURE_TYPE_B;
216                 else                              rce.pict_type= AV_PICTURE_TYPE_P;
217
218                 rce.new_pict_type= rce.pict_type;
219                 rce.mc_mb_var_sum= bits*s->mb_num/100000;
220                 rce.mb_var_sum   = s->mb_num;
221                 rce.qscale   = FF_QP2LAMBDA * 2;
222                 rce.f_code   = 2;
223                 rce.b_code   = 1;
224                 rce.misc_bits= 1;
225
226                 if(s->pict_type== AV_PICTURE_TYPE_I){
227                     rce.i_count   = s->mb_num;
228                     rce.i_tex_bits= bits;
229                     rce.p_tex_bits= 0;
230                     rce.mv_bits= 0;
231                 }else{
232                     rce.i_count   = 0; //FIXME we do know this approx
233                     rce.i_tex_bits= 0;
234                     rce.p_tex_bits= bits*0.9;
235                     rce.mv_bits= bits*0.1;
236                 }
237                 rcc->i_cplx_sum [rce.pict_type] += rce.i_tex_bits*rce.qscale;
238                 rcc->p_cplx_sum [rce.pict_type] += rce.p_tex_bits*rce.qscale;
239                 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
240                 rcc->frame_count[rce.pict_type] ++;
241
242                 get_qscale(s, &rce, rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum, i);
243                 rcc->pass1_wanted_bits+= s->bit_rate/(1/av_q2d(s->avctx->time_base)); //FIXME misbehaves a little for variable fps
244             }
245         }
246
247     }
248
249     return 0;
250 }
251
252 void ff_rate_control_uninit(MpegEncContext *s)
253 {
254     RateControlContext *rcc= &s->rc_context;
255     emms_c();
256
257     av_expr_free(rcc->rc_eq_eval);
258     av_freep(&rcc->entry);
259
260 #if CONFIG_LIBXVID
261     if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
262         ff_xvid_rate_control_uninit(s);
263 #endif
264 }
265
266 int ff_vbv_update(MpegEncContext *s, int frame_size){
267     RateControlContext *rcc= &s->rc_context;
268     const double fps= 1/av_q2d(s->avctx->time_base);
269     const int buffer_size= s->avctx->rc_buffer_size;
270     const double min_rate= s->avctx->rc_min_rate/fps;
271     const double max_rate= s->avctx->rc_max_rate/fps;
272
273 //printf("%d %f %d %f %f\n", buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
274     if(buffer_size){
275         int left;
276
277         rcc->buffer_index-= frame_size;
278         if(rcc->buffer_index < 0){
279             av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
280             rcc->buffer_index= 0;
281         }
282
283         left= buffer_size - rcc->buffer_index - 1;
284         rcc->buffer_index += av_clip(left, min_rate, max_rate);
285
286         if(rcc->buffer_index > buffer_size){
287             int stuffing= ceil((rcc->buffer_index - buffer_size)/8);
288
289             if(stuffing < 4 && s->codec_id == CODEC_ID_MPEG4)
290                 stuffing=4;
291             rcc->buffer_index -= 8*stuffing;
292
293             if(s->avctx->debug & FF_DEBUG_RC)
294                 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
295
296             return stuffing;
297         }
298     }
299     return 0;
300 }
301
302 /**
303  * Modify the bitrate curve from pass1 for one frame.
304  */
305 static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num){
306     RateControlContext *rcc= &s->rc_context;
307     AVCodecContext *a= s->avctx;
308     double q, bits;
309     const int pict_type= rce->new_pict_type;
310     const double mb_num= s->mb_num;
311     int i;
312
313     double const_values[]={
314         M_PI,
315         M_E,
316         rce->i_tex_bits*rce->qscale,
317         rce->p_tex_bits*rce->qscale,
318         (rce->i_tex_bits + rce->p_tex_bits)*(double)rce->qscale,
319         rce->mv_bits/mb_num,
320         rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code)*0.5 : rce->f_code,
321         rce->i_count/mb_num,
322         rce->mc_mb_var_sum/mb_num,
323         rce->mb_var_sum/mb_num,
324         rce->pict_type == AV_PICTURE_TYPE_I,
325         rce->pict_type == AV_PICTURE_TYPE_P,
326         rce->pict_type == AV_PICTURE_TYPE_B,
327         rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
328         a->qcompress,
329 /*        rcc->last_qscale_for[AV_PICTURE_TYPE_I],
330         rcc->last_qscale_for[AV_PICTURE_TYPE_P],
331         rcc->last_qscale_for[AV_PICTURE_TYPE_B],
332         rcc->next_non_b_qscale,*/
333         rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
334         rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
335         rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
336         rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
337         (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
338         0
339     };
340
341     bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
342     if (isnan(bits)) {
343         av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq);
344         return -1;
345     }
346
347     rcc->pass1_rc_eq_output_sum+= bits;
348     bits*=rate_factor;
349     if(bits<0.0) bits=0.0;
350     bits+= 1.0; //avoid 1/0 issues
351
352     /* user override */
353     for(i=0; i<s->avctx->rc_override_count; i++){
354         RcOverride *rco= s->avctx->rc_override;
355         if(rco[i].start_frame > frame_num) continue;
356         if(rco[i].end_frame   < frame_num) continue;
357
358         if(rco[i].qscale)
359             bits= qp2bits(rce, rco[i].qscale); //FIXME move at end to really force it?
360         else
361             bits*= rco[i].quality_factor;
362     }
363
364     q= bits2qp(rce, bits);
365
366     /* I/B difference */
367     if     (pict_type==AV_PICTURE_TYPE_I && s->avctx->i_quant_factor<0.0)
368         q= -q*s->avctx->i_quant_factor + s->avctx->i_quant_offset;
369     else if(pict_type==AV_PICTURE_TYPE_B && s->avctx->b_quant_factor<0.0)
370         q= -q*s->avctx->b_quant_factor + s->avctx->b_quant_offset;
371     if(q<1) q=1;
372
373     return q;
374 }
375
376 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q){
377     RateControlContext *rcc= &s->rc_context;
378     AVCodecContext *a= s->avctx;
379     const int pict_type= rce->new_pict_type;
380     const double last_p_q    = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
381     const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
382
383     if     (pict_type==AV_PICTURE_TYPE_I && (a->i_quant_factor>0.0 || rcc->last_non_b_pict_type==AV_PICTURE_TYPE_P))
384         q= last_p_q    *FFABS(a->i_quant_factor) + a->i_quant_offset;
385     else if(pict_type==AV_PICTURE_TYPE_B && a->b_quant_factor>0.0)
386         q= last_non_b_q*    a->b_quant_factor  + a->b_quant_offset;
387     if(q<1) q=1;
388
389     /* last qscale / qdiff stuff */
390     if(rcc->last_non_b_pict_type==pict_type || pict_type!=AV_PICTURE_TYPE_I){
391         double last_q= rcc->last_qscale_for[pict_type];
392         const int maxdiff= FF_QP2LAMBDA * a->max_qdiff;
393
394         if     (q > last_q + maxdiff) q= last_q + maxdiff;
395         else if(q < last_q - maxdiff) q= last_q - maxdiff;
396     }
397
398     rcc->last_qscale_for[pict_type]= q; //Note we cannot do that after blurring
399
400     if(pict_type!=AV_PICTURE_TYPE_B)
401         rcc->last_non_b_pict_type= pict_type;
402
403     return q;
404 }
405
406 /**
407  * Get the qmin & qmax for pict_type.
408  */
409 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type){
410     int qmin= s->avctx->lmin;
411     int qmax= s->avctx->lmax;
412
413     assert(qmin <= qmax);
414
415     if(pict_type==AV_PICTURE_TYPE_B){
416         qmin= (int)(qmin*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
417         qmax= (int)(qmax*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
418     }else if(pict_type==AV_PICTURE_TYPE_I){
419         qmin= (int)(qmin*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
420         qmax= (int)(qmax*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
421     }
422
423     qmin= av_clip(qmin, 1, FF_LAMBDA_MAX);
424     qmax= av_clip(qmax, 1, FF_LAMBDA_MAX);
425
426     if(qmax<qmin) qmax= qmin;
427
428     *qmin_ret= qmin;
429     *qmax_ret= qmax;
430 }
431
432 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce, double q, int frame_num){
433     RateControlContext *rcc= &s->rc_context;
434     int qmin, qmax;
435     const int pict_type= rce->new_pict_type;
436     const double buffer_size= s->avctx->rc_buffer_size;
437     const double fps= 1/av_q2d(s->avctx->time_base);
438     const double min_rate= s->avctx->rc_min_rate / fps;
439     const double max_rate= s->avctx->rc_max_rate / fps;
440
441     get_qminmax(&qmin, &qmax, s, pict_type);
442
443     /* modulation */
444     if(s->avctx->rc_qmod_freq && frame_num%s->avctx->rc_qmod_freq==0 && pict_type==AV_PICTURE_TYPE_P)
445         q*= s->avctx->rc_qmod_amp;
446
447 //printf("q:%f\n", q);
448     /* buffer overflow/underflow protection */
449     if(buffer_size){
450         double expected_size= rcc->buffer_index;
451         double q_limit;
452
453         if(min_rate){
454             double d= 2*(buffer_size - expected_size)/buffer_size;
455             if(d>1.0) d=1.0;
456             else if(d<0.0001) d=0.0001;
457             q*= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
458
459             q_limit= bits2qp(rce, FFMAX((min_rate - buffer_size + rcc->buffer_index) * s->avctx->rc_min_vbv_overflow_use, 1));
460             if(q > q_limit){
461                 if(s->avctx->debug&FF_DEBUG_RC){
462                     av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
463                 }
464                 q= q_limit;
465             }
466         }
467
468         if(max_rate){
469             double d= 2*expected_size/buffer_size;
470             if(d>1.0) d=1.0;
471             else if(d<0.0001) d=0.0001;
472             q/= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
473
474             q_limit= bits2qp(rce, FFMAX(rcc->buffer_index * s->avctx->rc_max_available_vbv_use, 1));
475             if(q < q_limit){
476                 if(s->avctx->debug&FF_DEBUG_RC){
477                     av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
478                 }
479                 q= q_limit;
480             }
481         }
482     }
483 //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);
484     if(s->avctx->rc_qsquish==0.0 || qmin==qmax){
485         if     (q<qmin) q=qmin;
486         else if(q>qmax) q=qmax;
487     }else{
488         double min2= log(qmin);
489         double max2= log(qmax);
490
491         q= log(q);
492         q= (q - min2)/(max2-min2) - 0.5;
493         q*= -4.0;
494         q= 1.0/(1.0 + exp(q));
495         q= q*(max2-min2) + min2;
496
497         q= exp(q);
498     }
499
500     return q;
501 }
502
503 //----------------------------------
504 // 1 Pass Code
505
506 static double predict_size(Predictor *p, double q, double var)
507 {
508      return p->coeff*var / (q*p->count);
509 }
510
511 static void update_predictor(Predictor *p, double q, double var, double size)
512 {
513     double new_coeff= size*q / (var + 1);
514     if(var<10) return;
515
516     p->count*= p->decay;
517     p->coeff*= p->decay;
518     p->count++;
519     p->coeff+= new_coeff;
520 }
521
522 static void adaptive_quantization(MpegEncContext *s, double q){
523     int i;
524     const float lumi_masking= s->avctx->lumi_masking / (128.0*128.0);
525     const float dark_masking= s->avctx->dark_masking / (128.0*128.0);
526     const float temp_cplx_masking= s->avctx->temporal_cplx_masking;
527     const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
528     const float p_masking = s->avctx->p_masking;
529     const float border_masking = s->avctx->border_masking;
530     float bits_sum= 0.0;
531     float cplx_sum= 0.0;
532     float *cplx_tab = s->cplx_tab;
533     float *bits_tab = s->bits_tab;
534     const int qmin= s->avctx->mb_lmin;
535     const int qmax= s->avctx->mb_lmax;
536     Picture * const pic= &s->current_picture;
537     const int mb_width = s->mb_width;
538     const int mb_height = s->mb_height;
539
540     for(i=0; i<s->mb_num; i++){
541         const int mb_xy= s->mb_index2xy[i];
542         float temp_cplx= sqrt(pic->mc_mb_var[mb_xy]); //FIXME merge in pow()
543         float spat_cplx= sqrt(pic->mb_var[mb_xy]);
544         const int lumi= pic->mb_mean[mb_xy];
545         float bits, cplx, factor;
546         int mb_x = mb_xy % s->mb_stride;
547         int mb_y = mb_xy / s->mb_stride;
548         int mb_distance;
549         float mb_factor = 0.0;
550         if(spat_cplx < 4) spat_cplx= 4; //FIXME finetune
551         if(temp_cplx < 4) temp_cplx= 4; //FIXME finetune
552
553         if((s->mb_type[mb_xy]&CANDIDATE_MB_TYPE_INTRA)){//FIXME hq mode
554             cplx= spat_cplx;
555             factor= 1.0 + p_masking;
556         }else{
557             cplx= temp_cplx;
558             factor= pow(temp_cplx, - temp_cplx_masking);
559         }
560         factor*=pow(spat_cplx, - spatial_cplx_masking);
561
562         if(lumi>127)
563             factor*= (1.0 - (lumi-128)*(lumi-128)*lumi_masking);
564         else
565             factor*= (1.0 - (lumi-128)*(lumi-128)*dark_masking);
566
567         if(mb_x < mb_width/5){
568             mb_distance = mb_width/5 - mb_x;
569             mb_factor = (float)mb_distance / (float)(mb_width/5);
570         }else if(mb_x > 4*mb_width/5){
571             mb_distance = mb_x - 4*mb_width/5;
572             mb_factor = (float)mb_distance / (float)(mb_width/5);
573         }
574         if(mb_y < mb_height/5){
575             mb_distance = mb_height/5 - mb_y;
576             mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
577         }else if(mb_y > 4*mb_height/5){
578             mb_distance = mb_y - 4*mb_height/5;
579             mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
580         }
581
582         factor*= 1.0 - border_masking*mb_factor;
583
584         if(factor<0.00001) factor= 0.00001;
585
586         bits= cplx*factor;
587         cplx_sum+= cplx;
588         bits_sum+= bits;
589         cplx_tab[i]= cplx;
590         bits_tab[i]= bits;
591     }
592
593     /* handle qmin/qmax clipping */
594     if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
595         float factor= bits_sum/cplx_sum;
596         for(i=0; i<s->mb_num; i++){
597             float newq= q*cplx_tab[i]/bits_tab[i];
598             newq*= factor;
599
600             if     (newq > qmax){
601                 bits_sum -= bits_tab[i];
602                 cplx_sum -= cplx_tab[i]*q/qmax;
603             }
604             else if(newq < qmin){
605                 bits_sum -= bits_tab[i];
606                 cplx_sum -= cplx_tab[i]*q/qmin;
607             }
608         }
609         if(bits_sum < 0.001) bits_sum= 0.001;
610         if(cplx_sum < 0.001) cplx_sum= 0.001;
611     }
612
613     for(i=0; i<s->mb_num; i++){
614         const int mb_xy= s->mb_index2xy[i];
615         float newq= q*cplx_tab[i]/bits_tab[i];
616         int intq;
617
618         if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
619             newq*= bits_sum/cplx_sum;
620         }
621
622         intq= (int)(newq + 0.5);
623
624         if     (intq > qmax) intq= qmax;
625         else if(intq < qmin) intq= qmin;
626 //if(i%s->mb_width==0) printf("\n");
627 //printf("%2d%3d ", intq, ff_sqrt(s->mc_mb_var[i]));
628         s->lambda_table[mb_xy]= intq;
629     }
630 }
631
632 void ff_get_2pass_fcode(MpegEncContext *s){
633     RateControlContext *rcc= &s->rc_context;
634     int picture_number= s->picture_number;
635     RateControlEntry *rce;
636
637     rce= &rcc->entry[picture_number];
638     s->f_code= rce->f_code;
639     s->b_code= rce->b_code;
640 }
641
642 //FIXME rd or at least approx for dquant
643
644 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
645 {
646     float q;
647     int qmin, qmax;
648     float br_compensation;
649     double diff;
650     double short_term_q;
651     double fps;
652     int picture_number= s->picture_number;
653     int64_t wanted_bits;
654     RateControlContext *rcc= &s->rc_context;
655     AVCodecContext *a= s->avctx;
656     RateControlEntry local_rce, *rce;
657     double bits;
658     double rate_factor;
659     int var;
660     const int pict_type= s->pict_type;
661     Picture * const pic= &s->current_picture;
662     emms_c();
663
664 #if CONFIG_LIBXVID
665     if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
666         return ff_xvid_rate_estimate_qscale(s, dry_run);
667 #endif
668
669     get_qminmax(&qmin, &qmax, s, pict_type);
670
671     fps= 1/av_q2d(s->avctx->time_base);
672 //printf("input_pic_num:%d pic_num:%d frame_rate:%d\n", s->input_picture_number, s->picture_number, s->frame_rate);
673         /* update predictors */
674     if(picture_number>2 && !dry_run){
675         const int last_var= s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum;
676         update_predictor(&rcc->pred[s->last_pict_type], rcc->last_qscale, sqrt(last_var), s->frame_bits);
677     }
678
679     if(s->flags&CODEC_FLAG_PASS2){
680         assert(picture_number>=0);
681         assert(picture_number<rcc->num_entries);
682         rce= &rcc->entry[picture_number];
683         wanted_bits= rce->expected_bits;
684     }else{
685         Picture *dts_pic;
686         rce= &local_rce;
687
688         //FIXME add a dts field to AVFrame and ensure its set and use it here instead of reordering
689         //but the reordering is simpler for now until h.264 b pyramid must be handeld
690         if(s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
691             dts_pic= s->current_picture_ptr;
692         else
693             dts_pic= s->last_picture_ptr;
694
695 //if(dts_pic)
696 //            av_log(NULL, AV_LOG_ERROR, "%Ld %Ld %Ld %d\n", s->current_picture_ptr->pts, s->user_specified_pts, dts_pic->pts, picture_number);
697
698         if (!dts_pic || dts_pic->f.pts == AV_NOPTS_VALUE)
699             wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps);
700         else
701             wanted_bits = (uint64_t)(s->bit_rate*(double)dts_pic->f.pts / 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 == AV_PICTURE_TYPE_I ? 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!=AV_PICTURE_TYPE_I)
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== AV_PICTURE_TYPE_I){
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==AV_PICTURE_TYPE_P || 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_picture_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     return q;
798 }
799
800 //----------------------------------------------
801 // 2-Pass code
802
803 static int init_pass2(MpegEncContext *s)
804 {
805     RateControlContext *rcc= &s->rc_context;
806     AVCodecContext *a= s->avctx;
807     int i, toobig;
808     double fps= 1/av_q2d(s->avctx->time_base);
809     double complexity[5]={0,0,0,0,0};   // aproximate bits at quant=1
810     uint64_t const_bits[5]={0,0,0,0,0}; // quantizer independent bits
811     uint64_t all_const_bits;
812     uint64_t all_available_bits= (uint64_t)(s->bit_rate*(double)rcc->num_entries/fps);
813     double rate_factor=0;
814     double step;
815     //int last_i_frame=-10000000;
816     const int filter_size= (int)(a->qblur*4) | 1;
817     double expected_bits;
818     double *qscale, *blurred_qscale, qscale_sum;
819
820     /* find complexity & const_bits & decide the pict_types */
821     for(i=0; i<rcc->num_entries; i++){
822         RateControlEntry *rce= &rcc->entry[i];
823
824         rce->new_pict_type= rce->pict_type;
825         rcc->i_cplx_sum [rce->pict_type] += rce->i_tex_bits*rce->qscale;
826         rcc->p_cplx_sum [rce->pict_type] += rce->p_tex_bits*rce->qscale;
827         rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
828         rcc->frame_count[rce->pict_type] ++;
829
830         complexity[rce->new_pict_type]+= (rce->i_tex_bits+ rce->p_tex_bits)*(double)rce->qscale;
831         const_bits[rce->new_pict_type]+= rce->mv_bits + rce->misc_bits;
832     }
833     all_const_bits= const_bits[AV_PICTURE_TYPE_I] + const_bits[AV_PICTURE_TYPE_P] + const_bits[AV_PICTURE_TYPE_B];
834
835     if(all_available_bits < all_const_bits){
836         av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
837         return -1;
838     }
839
840     qscale= av_malloc(sizeof(double)*rcc->num_entries);
841     blurred_qscale= av_malloc(sizeof(double)*rcc->num_entries);
842     toobig = 0;
843
844     for(step=256*256; step>0.0000001; step*=0.5){
845         expected_bits=0;
846         rate_factor+= step;
847
848         rcc->buffer_index= s->avctx->rc_buffer_size/2;
849
850         /* find qscale */
851         for(i=0; i<rcc->num_entries; i++){
852             RateControlEntry *rce= &rcc->entry[i];
853             qscale[i]= get_qscale(s, &rcc->entry[i], rate_factor, i);
854             rcc->last_qscale_for[rce->pict_type] = qscale[i];
855         }
856         assert(filter_size%2==1);
857
858         /* fixed I/B QP relative to P mode */
859         for(i=rcc->num_entries-1; i>=0; i--){
860             RateControlEntry *rce= &rcc->entry[i];
861
862             qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
863         }
864
865         /* smooth curve */
866         for(i=0; i<rcc->num_entries; i++){
867             RateControlEntry *rce= &rcc->entry[i];
868             const int pict_type= rce->new_pict_type;
869             int j;
870             double q=0.0, sum=0.0;
871
872             for(j=0; j<filter_size; j++){
873                 int index= i+j-filter_size/2;
874                 double d= index-i;
875                 double coeff= a->qblur==0 ? 1.0 : exp(-d*d/(a->qblur * a->qblur));
876
877                 if(index < 0 || index >= rcc->num_entries) continue;
878                 if(pict_type != rcc->entry[index].new_pict_type) continue;
879                 q+= qscale[index] * coeff;
880                 sum+= coeff;
881             }
882             blurred_qscale[i]= q/sum;
883         }
884
885         /* find expected bits */
886         for(i=0; i<rcc->num_entries; i++){
887             RateControlEntry *rce= &rcc->entry[i];
888             double bits;
889             rce->new_qscale= modify_qscale(s, rce, blurred_qscale[i], i);
890             bits= qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
891 //printf("%d %f\n", rce->new_bits, blurred_qscale[i]);
892             bits += 8*ff_vbv_update(s, bits);
893
894             rce->expected_bits= expected_bits;
895             expected_bits += bits;
896         }
897
898         /*
899         av_log(s->avctx, AV_LOG_INFO,
900             "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
901             expected_bits, (int)all_available_bits, rate_factor);
902         */
903         if(expected_bits > all_available_bits) {
904             rate_factor-= step;
905             ++toobig;
906         }
907     }
908     av_free(qscale);
909     av_free(blurred_qscale);
910
911     /* check bitrate calculations and print info */
912     qscale_sum = 0.0;
913     for(i=0; i<rcc->num_entries; i++){
914         /* av_log(s->avctx, AV_LOG_DEBUG, "[lavc rc] entry[%d].new_qscale = %.3f  qp = %.3f\n",
915             i, rcc->entry[i].new_qscale, rcc->entry[i].new_qscale / FF_QP2LAMBDA); */
916         qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA, s->avctx->qmin, s->avctx->qmax);
917     }
918     assert(toobig <= 40);
919     av_log(s->avctx, AV_LOG_DEBUG,
920         "[lavc rc] requested bitrate: %d bps  expected bitrate: %d bps\n",
921         s->bit_rate,
922         (int)(expected_bits / ((double)all_available_bits/s->bit_rate)));
923     av_log(s->avctx, AV_LOG_DEBUG,
924         "[lavc rc] estimated target average qp: %.3f\n",
925         (float)qscale_sum / rcc->num_entries);
926     if (toobig == 0) {
927         av_log(s->avctx, AV_LOG_INFO,
928             "[lavc rc] Using all of requested bitrate is not "
929             "necessary for this video with these parameters.\n");
930     } else if (toobig == 40) {
931         av_log(s->avctx, AV_LOG_ERROR,
932             "[lavc rc] Error: bitrate too low for this video "
933             "with these parameters.\n");
934         return -1;
935     } else if (fabs(expected_bits/all_available_bits - 1.0) > 0.01) {
936         av_log(s->avctx, AV_LOG_ERROR,
937             "[lavc rc] Error: 2pass curve failed to converge\n");
938         return -1;
939     }
940
941     return 0;
942 }