RC: Fix initial buffer occupancy.
[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
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     if (!rcc->buffer_index)
129         rcc->buffer_index = s->avctx->rc_buffer_size * 3 / 4;
130
131     if(s->flags&CODEC_FLAG_PASS2){
132         int i;
133         char *p;
134
135         /* find number of pics */
136         p= s->avctx->stats_in;
137         for(i=-1; p; i++){
138             p= strchr(p+1, ';');
139         }
140         i+= s->max_b_frames;
141         if(i<=0 || i>=INT_MAX / sizeof(RateControlEntry))
142             return -1;
143         rcc->entry = av_mallocz(i*sizeof(RateControlEntry));
144         rcc->num_entries= i;
145
146         /* init all to skipped p frames (with b frames we might have a not encoded frame at the end FIXME) */
147         for(i=0; i<rcc->num_entries; i++){
148             RateControlEntry *rce= &rcc->entry[i];
149             rce->pict_type= rce->new_pict_type=AV_PICTURE_TYPE_P;
150             rce->qscale= rce->new_qscale=FF_QP2LAMBDA * 2;
151             rce->misc_bits= s->mb_num + 10;
152             rce->mb_var_sum= s->mb_num*100;
153         }
154
155         /* read stats */
156         p= s->avctx->stats_in;
157         for(i=0; i<rcc->num_entries - s->max_b_frames; i++){
158             RateControlEntry *rce;
159             int picture_number;
160             int e;
161             char *next;
162
163             next= strchr(p, ';');
164             if(next){
165                 (*next)=0; //sscanf in unbelievably slow on looong strings //FIXME copy / do not write
166                 next++;
167             }
168             e= sscanf(p, " in:%d ", &picture_number);
169
170             assert(picture_number >= 0);
171             assert(picture_number < rcc->num_entries);
172             rce= &rcc->entry[picture_number];
173
174             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",
175                    &rce->pict_type, &rce->qscale, &rce->i_tex_bits, &rce->p_tex_bits, &rce->mv_bits, &rce->misc_bits,
176                    &rce->f_code, &rce->b_code, &rce->mc_mb_var_sum, &rce->mb_var_sum, &rce->i_count, &rce->skip_count, &rce->header_bits);
177             if(e!=14){
178                 av_log(s->avctx, AV_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
179                 return -1;
180             }
181
182             p= next;
183         }
184
185         if(init_pass2(s) < 0) return -1;
186
187         //FIXME maybe move to end
188         if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID) {
189 #if CONFIG_LIBXVID
190             return ff_xvid_rate_control_init(s);
191 #else
192             av_log(s->avctx, AV_LOG_ERROR, "Xvid ratecontrol requires libavcodec compiled with Xvid support.\n");
193             return -1;
194 #endif
195         }
196     }
197
198     if(!(s->flags&CODEC_FLAG_PASS2)){
199
200         rcc->short_term_qsum=0.001;
201         rcc->short_term_qcount=0.001;
202
203         rcc->pass1_rc_eq_output_sum= 0.001;
204         rcc->pass1_wanted_bits=0.001;
205
206         if(s->avctx->qblur > 1.0){
207             av_log(s->avctx, AV_LOG_ERROR, "qblur too large\n");
208             return -1;
209         }
210         /* init stuff with the user specified complexity */
211         if(s->avctx->rc_initial_cplx){
212             for(i=0; i<60*30; i++){
213                 double bits= s->avctx->rc_initial_cplx * (i/10000.0 + 1.0)*s->mb_num;
214                 RateControlEntry rce;
215
216                 if     (i%((s->gop_size+3)/4)==0) rce.pict_type= AV_PICTURE_TYPE_I;
217                 else if(i%(s->max_b_frames+1))    rce.pict_type= AV_PICTURE_TYPE_B;
218                 else                              rce.pict_type= AV_PICTURE_TYPE_P;
219
220                 rce.new_pict_type= rce.pict_type;
221                 rce.mc_mb_var_sum= bits*s->mb_num/100000;
222                 rce.mb_var_sum   = s->mb_num;
223                 rce.qscale   = FF_QP2LAMBDA * 2;
224                 rce.f_code   = 2;
225                 rce.b_code   = 1;
226                 rce.misc_bits= 1;
227
228                 if(s->pict_type== AV_PICTURE_TYPE_I){
229                     rce.i_count   = s->mb_num;
230                     rce.i_tex_bits= bits;
231                     rce.p_tex_bits= 0;
232                     rce.mv_bits= 0;
233                 }else{
234                     rce.i_count   = 0; //FIXME we do know this approx
235                     rce.i_tex_bits= 0;
236                     rce.p_tex_bits= bits*0.9;
237                     rce.mv_bits= bits*0.1;
238                 }
239                 rcc->i_cplx_sum [rce.pict_type] += rce.i_tex_bits*rce.qscale;
240                 rcc->p_cplx_sum [rce.pict_type] += rce.p_tex_bits*rce.qscale;
241                 rcc->mv_bits_sum[rce.pict_type] += rce.mv_bits;
242                 rcc->frame_count[rce.pict_type] ++;
243
244                 get_qscale(s, &rce, rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum, i);
245                 rcc->pass1_wanted_bits+= s->bit_rate/(1/av_q2d(s->avctx->time_base)); //FIXME misbehaves a little for variable fps
246             }
247         }
248
249     }
250
251     return 0;
252 }
253
254 void ff_rate_control_uninit(MpegEncContext *s)
255 {
256     RateControlContext *rcc= &s->rc_context;
257     emms_c();
258
259     av_expr_free(rcc->rc_eq_eval);
260     av_freep(&rcc->entry);
261
262 #if CONFIG_LIBXVID
263     if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
264         ff_xvid_rate_control_uninit(s);
265 #endif
266 }
267
268 int ff_vbv_update(MpegEncContext *s, int frame_size){
269     RateControlContext *rcc= &s->rc_context;
270     const double fps= 1/av_q2d(s->avctx->time_base);
271     const int buffer_size= s->avctx->rc_buffer_size;
272     const double min_rate= s->avctx->rc_min_rate/fps;
273     const double max_rate= s->avctx->rc_max_rate/fps;
274
275 //printf("%d %f %d %f %f\n", buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
276     if(buffer_size){
277         int left;
278
279         rcc->buffer_index-= frame_size;
280         if(rcc->buffer_index < 0){
281             av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
282             rcc->buffer_index= 0;
283         }
284
285         left= buffer_size - rcc->buffer_index - 1;
286         rcc->buffer_index += av_clip(left, min_rate, max_rate);
287
288         if(rcc->buffer_index > buffer_size){
289             int stuffing= ceil((rcc->buffer_index - buffer_size)/8);
290
291             if(stuffing < 4 && s->codec_id == CODEC_ID_MPEG4)
292                 stuffing=4;
293             rcc->buffer_index -= 8*stuffing;
294
295             if(s->avctx->debug & FF_DEBUG_RC)
296                 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
297
298             return stuffing;
299         }
300     }
301     return 0;
302 }
303
304 /**
305  * Modify the bitrate curve from pass1 for one frame.
306  */
307 static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num){
308     RateControlContext *rcc= &s->rc_context;
309     AVCodecContext *a= s->avctx;
310     double q, bits;
311     const int pict_type= rce->new_pict_type;
312     const double mb_num= s->mb_num;
313     int i;
314
315     double const_values[]={
316         M_PI,
317         M_E,
318         rce->i_tex_bits*rce->qscale,
319         rce->p_tex_bits*rce->qscale,
320         (rce->i_tex_bits + rce->p_tex_bits)*(double)rce->qscale,
321         rce->mv_bits/mb_num,
322         rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code)*0.5 : rce->f_code,
323         rce->i_count/mb_num,
324         rce->mc_mb_var_sum/mb_num,
325         rce->mb_var_sum/mb_num,
326         rce->pict_type == AV_PICTURE_TYPE_I,
327         rce->pict_type == AV_PICTURE_TYPE_P,
328         rce->pict_type == AV_PICTURE_TYPE_B,
329         rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
330         a->qcompress,
331 /*        rcc->last_qscale_for[AV_PICTURE_TYPE_I],
332         rcc->last_qscale_for[AV_PICTURE_TYPE_P],
333         rcc->last_qscale_for[AV_PICTURE_TYPE_B],
334         rcc->next_non_b_qscale,*/
335         rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
336         rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
337         rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
338         rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
339         (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
340         0
341     };
342
343     bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
344     if (isnan(bits)) {
345         av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq);
346         return -1;
347     }
348
349     rcc->pass1_rc_eq_output_sum+= bits;
350     bits*=rate_factor;
351     if(bits<0.0) bits=0.0;
352     bits+= 1.0; //avoid 1/0 issues
353
354     /* user override */
355     for(i=0; i<s->avctx->rc_override_count; i++){
356         RcOverride *rco= s->avctx->rc_override;
357         if(rco[i].start_frame > frame_num) continue;
358         if(rco[i].end_frame   < frame_num) continue;
359
360         if(rco[i].qscale)
361             bits= qp2bits(rce, rco[i].qscale); //FIXME move at end to really force it?
362         else
363             bits*= rco[i].quality_factor;
364     }
365
366     q= bits2qp(rce, bits);
367
368     /* I/B difference */
369     if     (pict_type==AV_PICTURE_TYPE_I && s->avctx->i_quant_factor<0.0)
370         q= -q*s->avctx->i_quant_factor + s->avctx->i_quant_offset;
371     else if(pict_type==AV_PICTURE_TYPE_B && s->avctx->b_quant_factor<0.0)
372         q= -q*s->avctx->b_quant_factor + s->avctx->b_quant_offset;
373     if(q<1) q=1;
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[AV_PICTURE_TYPE_P];
383     const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
384
385     if     (pict_type==AV_PICTURE_TYPE_I && (a->i_quant_factor>0.0 || rcc->last_non_b_pict_type==AV_PICTURE_TYPE_P))
386         q= last_p_q    *FFABS(a->i_quant_factor) + a->i_quant_offset;
387     else if(pict_type==AV_PICTURE_TYPE_B && a->b_quant_factor>0.0)
388         q= last_non_b_q*    a->b_quant_factor  + a->b_quant_offset;
389     if(q<1) q=1;
390
391     /* last qscale / qdiff stuff */
392     if(rcc->last_non_b_pict_type==pict_type || pict_type!=AV_PICTURE_TYPE_I){
393         double last_q= rcc->last_qscale_for[pict_type];
394         const int maxdiff= FF_QP2LAMBDA * a->max_qdiff;
395
396         if     (q > last_q + maxdiff) q= last_q + maxdiff;
397         else if(q < last_q - maxdiff) q= last_q - maxdiff;
398     }
399
400     rcc->last_qscale_for[pict_type]= q; //Note we cannot do that after blurring
401
402     if(pict_type!=AV_PICTURE_TYPE_B)
403         rcc->last_non_b_pict_type= pict_type;
404
405     return q;
406 }
407
408 /**
409  * Get the qmin & qmax for pict_type.
410  */
411 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type){
412     int qmin= s->avctx->lmin;
413     int qmax= s->avctx->lmax;
414
415     assert(qmin <= qmax);
416
417     if(pict_type==AV_PICTURE_TYPE_B){
418         qmin= (int)(qmin*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
419         qmax= (int)(qmax*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
420     }else if(pict_type==AV_PICTURE_TYPE_I){
421         qmin= (int)(qmin*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
422         qmax= (int)(qmax*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
423     }
424
425     qmin= av_clip(qmin, 1, FF_LAMBDA_MAX);
426     qmax= av_clip(qmax, 1, FF_LAMBDA_MAX);
427
428     if(qmax<qmin) qmax= qmin;
429
430     *qmin_ret= qmin;
431     *qmax_ret= qmax;
432 }
433
434 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce, double q, int frame_num){
435     RateControlContext *rcc= &s->rc_context;
436     int qmin, qmax;
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==AV_PICTURE_TYPE_P)
447         q*= s->avctx->rc_qmod_amp;
448
449 //printf("q:%f\n", q);
450     /* buffer overflow/underflow protection */
451     if(buffer_size){
452         double expected_size= rcc->buffer_index;
453         double q_limit;
454
455         if(min_rate){
456             double d= 2*(buffer_size - expected_size)/buffer_size;
457             if(d>1.0) d=1.0;
458             else if(d<0.0001) d=0.0001;
459             q*= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
460
461             q_limit= bits2qp(rce, FFMAX((min_rate - buffer_size + rcc->buffer_index) * s->avctx->rc_min_vbv_overflow_use, 1));
462             if(q > q_limit){
463                 if(s->avctx->debug&FF_DEBUG_RC){
464                     av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
465                 }
466                 q= q_limit;
467             }
468         }
469
470         if(max_rate){
471             double d= 2*expected_size/buffer_size;
472             if(d>1.0) d=1.0;
473             else if(d<0.0001) d=0.0001;
474             q/= pow(d, 1.0/s->avctx->rc_buffer_aggressivity);
475
476             q_limit= bits2qp(rce, FFMAX(rcc->buffer_index * s->avctx->rc_max_available_vbv_use, 1));
477             if(q < q_limit){
478                 if(s->avctx->debug&FF_DEBUG_RC){
479                     av_log(s->avctx, AV_LOG_DEBUG, "limiting QP %f -> %f\n", q, q_limit);
480                 }
481                 q= q_limit;
482             }
483         }
484     }
485 //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);
486     if(s->avctx->rc_qsquish==0.0 || qmin==qmax){
487         if     (q<qmin) q=qmin;
488         else if(q>qmax) q=qmax;
489     }else{
490         double min2= log(qmin);
491         double max2= log(qmax);
492
493         q= log(q);
494         q= (q - min2)/(max2-min2) - 0.5;
495         q*= -4.0;
496         q= 1.0/(1.0 + exp(q));
497         q= q*(max2-min2) + min2;
498
499         q= exp(q);
500     }
501
502     return q;
503 }
504
505 //----------------------------------
506 // 1 Pass Code
507
508 static double predict_size(Predictor *p, double q, double var)
509 {
510      return p->coeff*var / (q*p->count);
511 }
512
513 /*
514 static double predict_qp(Predictor *p, double size, double var)
515 {
516 //printf("coeff:%f, count:%f, var:%f, size:%f//\n", p->coeff, p->count, var, size);
517      return p->coeff*var / (size*p->count);
518 }
519 */
520
521 static void update_predictor(Predictor *p, double q, double var, double size)
522 {
523     double new_coeff= size*q / (var + 1);
524     if(var<10) return;
525
526     p->count*= p->decay;
527     p->coeff*= p->decay;
528     p->count++;
529     p->coeff+= new_coeff;
530 }
531
532 static void adaptive_quantization(MpegEncContext *s, double q){
533     int i;
534     const float lumi_masking= s->avctx->lumi_masking / (128.0*128.0);
535     const float dark_masking= s->avctx->dark_masking / (128.0*128.0);
536     const float temp_cplx_masking= s->avctx->temporal_cplx_masking;
537     const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
538     const float p_masking = s->avctx->p_masking;
539     const float border_masking = s->avctx->border_masking;
540     float bits_sum= 0.0;
541     float cplx_sum= 0.0;
542     float *cplx_tab = av_malloc(s->mb_num * sizeof(*cplx_tab));
543     float *bits_tab = av_malloc(s->mb_num * sizeof(*bits_tab));
544     const int qmin= s->avctx->mb_lmin;
545     const int qmax= s->avctx->mb_lmax;
546     Picture * const pic= &s->current_picture;
547     const int mb_width = s->mb_width;
548     const int mb_height = s->mb_height;
549
550     for(i=0; i<s->mb_num; i++){
551         const int mb_xy= s->mb_index2xy[i];
552         float temp_cplx= sqrt(pic->mc_mb_var[mb_xy]); //FIXME merge in pow()
553         float spat_cplx= sqrt(pic->mb_var[mb_xy]);
554         const int lumi= pic->mb_mean[mb_xy];
555         float bits, cplx, factor;
556         int mb_x = mb_xy % s->mb_stride;
557         int mb_y = mb_xy / s->mb_stride;
558         int mb_distance;
559         float mb_factor = 0.0;
560 #if 0
561         if(spat_cplx < q/3) spat_cplx= q/3; //FIXME finetune
562         if(temp_cplx < q/3) temp_cplx= q/3; //FIXME finetune
563 #endif
564         if(spat_cplx < 4) spat_cplx= 4; //FIXME finetune
565         if(temp_cplx < 4) temp_cplx= 4; //FIXME finetune
566
567         if((s->mb_type[mb_xy]&CANDIDATE_MB_TYPE_INTRA)){//FIXME hq mode
568             cplx= spat_cplx;
569             factor= 1.0 + p_masking;
570         }else{
571             cplx= temp_cplx;
572             factor= pow(temp_cplx, - temp_cplx_masking);
573         }
574         factor*=pow(spat_cplx, - spatial_cplx_masking);
575
576         if(lumi>127)
577             factor*= (1.0 - (lumi-128)*(lumi-128)*lumi_masking);
578         else
579             factor*= (1.0 - (lumi-128)*(lumi-128)*dark_masking);
580
581         if(mb_x < mb_width/5){
582             mb_distance = mb_width/5 - mb_x;
583             mb_factor = (float)mb_distance / (float)(mb_width/5);
584         }else if(mb_x > 4*mb_width/5){
585             mb_distance = mb_x - 4*mb_width/5;
586             mb_factor = (float)mb_distance / (float)(mb_width/5);
587         }
588         if(mb_y < mb_height/5){
589             mb_distance = mb_height/5 - mb_y;
590             mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
591         }else if(mb_y > 4*mb_height/5){
592             mb_distance = mb_y - 4*mb_height/5;
593             mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
594         }
595
596         factor*= 1.0 - border_masking*mb_factor;
597
598         if(factor<0.00001) factor= 0.00001;
599
600         bits= cplx*factor;
601         cplx_sum+= cplx;
602         bits_sum+= bits;
603         cplx_tab[i]= cplx;
604         bits_tab[i]= bits;
605     }
606
607     /* handle qmin/qmax clipping */
608     if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
609         float factor= bits_sum/cplx_sum;
610         for(i=0; i<s->mb_num; i++){
611             float newq= q*cplx_tab[i]/bits_tab[i];
612             newq*= factor;
613
614             if     (newq > qmax){
615                 bits_sum -= bits_tab[i];
616                 cplx_sum -= cplx_tab[i]*q/qmax;
617             }
618             else if(newq < qmin){
619                 bits_sum -= bits_tab[i];
620                 cplx_sum -= cplx_tab[i]*q/qmin;
621             }
622         }
623         if(bits_sum < 0.001) bits_sum= 0.001;
624         if(cplx_sum < 0.001) cplx_sum= 0.001;
625     }
626
627     for(i=0; i<s->mb_num; i++){
628         const int mb_xy= s->mb_index2xy[i];
629         float newq= q*cplx_tab[i]/bits_tab[i];
630         int intq;
631
632         if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
633             newq*= bits_sum/cplx_sum;
634         }
635
636         intq= (int)(newq + 0.5);
637
638         if     (intq > qmax) intq= qmax;
639         else if(intq < qmin) intq= qmin;
640 //if(i%s->mb_width==0) printf("\n");
641 //printf("%2d%3d ", intq, ff_sqrt(s->mc_mb_var[i]));
642         s->lambda_table[mb_xy]= intq;
643     }
644
645     av_free(cplx_tab);
646     av_free(bits_tab);
647 }
648
649 void ff_get_2pass_fcode(MpegEncContext *s){
650     RateControlContext *rcc= &s->rc_context;
651     int picture_number= s->picture_number;
652     RateControlEntry *rce;
653
654     rce= &rcc->entry[picture_number];
655     s->f_code= rce->f_code;
656     s->b_code= rce->b_code;
657 }
658
659 //FIXME rd or at least approx for dquant
660
661 float ff_rate_estimate_qscale(MpegEncContext *s, int dry_run)
662 {
663     float q;
664     int qmin, qmax;
665     float br_compensation;
666     double diff;
667     double short_term_q;
668     double fps;
669     int picture_number= s->picture_number;
670     int64_t wanted_bits;
671     RateControlContext *rcc= &s->rc_context;
672     AVCodecContext *a= s->avctx;
673     RateControlEntry local_rce, *rce;
674     double bits;
675     double rate_factor;
676     int var;
677     const int pict_type= s->pict_type;
678     Picture * const pic= &s->current_picture;
679     emms_c();
680
681 #if CONFIG_LIBXVID
682     if((s->flags&CODEC_FLAG_PASS2) && s->avctx->rc_strategy == FF_RC_STRATEGY_XVID)
683         return ff_xvid_rate_estimate_qscale(s, dry_run);
684 #endif
685
686     get_qminmax(&qmin, &qmax, s, pict_type);
687
688     fps= 1/av_q2d(s->avctx->time_base);
689 //printf("input_pic_num:%d pic_num:%d frame_rate:%d\n", s->input_picture_number, s->picture_number, s->frame_rate);
690         /* update predictors */
691     if(picture_number>2 && !dry_run){
692         const int last_var= s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum;
693         update_predictor(&rcc->pred[s->last_pict_type], rcc->last_qscale, sqrt(last_var), s->frame_bits);
694     }
695
696     if(s->flags&CODEC_FLAG_PASS2){
697         assert(picture_number>=0);
698         assert(picture_number<rcc->num_entries);
699         rce= &rcc->entry[picture_number];
700         wanted_bits= rce->expected_bits;
701     }else{
702         Picture *dts_pic;
703         rce= &local_rce;
704
705         //FIXME add a dts field to AVFrame and ensure its set and use it here instead of reordering
706         //but the reordering is simpler for now until h.264 b pyramid must be handeld
707         if(s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
708             dts_pic= s->current_picture_ptr;
709         else
710             dts_pic= s->last_picture_ptr;
711
712 //if(dts_pic)
713 //            av_log(NULL, AV_LOG_ERROR, "%"PRId64" %"PRId64" %"PRId64" %d\n", s->current_picture_ptr->pts, s->user_specified_pts, dts_pic->pts, picture_number);
714
715         if (!dts_pic || dts_pic->f.pts == AV_NOPTS_VALUE)
716             wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps);
717         else
718             wanted_bits = (uint64_t)(s->bit_rate*(double)dts_pic->f.pts / fps);
719     }
720
721     diff= s->total_bits - wanted_bits;
722     br_compensation= (a->bit_rate_tolerance - diff)/a->bit_rate_tolerance;
723     if(br_compensation<=0.0) br_compensation=0.001;
724
725     var= pict_type == AV_PICTURE_TYPE_I ? pic->mb_var_sum : pic->mc_mb_var_sum;
726
727     short_term_q = 0; /* avoid warning */
728     if(s->flags&CODEC_FLAG_PASS2){
729         if(pict_type!=AV_PICTURE_TYPE_I)
730             assert(pict_type == rce->new_pict_type);
731
732         q= rce->new_qscale / br_compensation;
733 //printf("%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale, br_compensation, s->frame_bits, var, pict_type);
734     }else{
735         rce->pict_type=
736         rce->new_pict_type= pict_type;
737         rce->mc_mb_var_sum= pic->mc_mb_var_sum;
738         rce->mb_var_sum   = pic->   mb_var_sum;
739         rce->qscale   = FF_QP2LAMBDA * 2;
740         rce->f_code   = s->f_code;
741         rce->b_code   = s->b_code;
742         rce->misc_bits= 1;
743
744         bits= predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
745         if(pict_type== AV_PICTURE_TYPE_I){
746             rce->i_count   = s->mb_num;
747             rce->i_tex_bits= bits;
748             rce->p_tex_bits= 0;
749             rce->mv_bits= 0;
750         }else{
751             rce->i_count   = 0; //FIXME we do know this approx
752             rce->i_tex_bits= 0;
753             rce->p_tex_bits= bits*0.9;
754
755             rce->mv_bits= bits*0.1;
756         }
757         rcc->i_cplx_sum [pict_type] += rce->i_tex_bits*rce->qscale;
758         rcc->p_cplx_sum [pict_type] += rce->p_tex_bits*rce->qscale;
759         rcc->mv_bits_sum[pict_type] += rce->mv_bits;
760         rcc->frame_count[pict_type] ++;
761
762         bits= rce->i_tex_bits + rce->p_tex_bits;
763         rate_factor= rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum * br_compensation;
764
765         q= get_qscale(s, rce, rate_factor, picture_number);
766         if (q < 0)
767             return -1;
768
769         assert(q>0.0);
770 //printf("%f ", q);
771         q= get_diff_limited_q(s, rce, q);
772 //printf("%f ", q);
773         assert(q>0.0);
774
775         if(pict_type==AV_PICTURE_TYPE_P || s->intra_only){ //FIXME type dependent blur like in 2-pass
776             rcc->short_term_qsum*=a->qblur;
777             rcc->short_term_qcount*=a->qblur;
778
779             rcc->short_term_qsum+= q;
780             rcc->short_term_qcount++;
781 //printf("%f ", q);
782             q= short_term_q= rcc->short_term_qsum/rcc->short_term_qcount;
783 //printf("%f ", q);
784         }
785         assert(q>0.0);
786
787         q= modify_qscale(s, rce, q, picture_number);
788
789         rcc->pass1_wanted_bits+= s->bit_rate/fps;
790
791         assert(q>0.0);
792     }
793
794     if(s->avctx->debug&FF_DEBUG_RC){
795         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",
796         av_get_picture_type_char(pict_type), qmin, q, qmax, picture_number, (int)wanted_bits/1000, (int)s->total_bits/1000,
797         br_compensation, short_term_q, s->frame_bits, pic->mb_var_sum, pic->mc_mb_var_sum, s->bit_rate/1000, (int)fps
798         );
799     }
800
801     if     (q<qmin) q=qmin;
802     else if(q>qmax) q=qmax;
803
804     if(s->adaptive_quant)
805         adaptive_quantization(s, q);
806     else
807         q= (int)(q + 0.5);
808
809     if(!dry_run){
810         rcc->last_qscale= q;
811         rcc->last_mc_mb_var_sum= pic->mc_mb_var_sum;
812         rcc->last_mb_var_sum= pic->mb_var_sum;
813     }
814     return q;
815 }
816
817 //----------------------------------------------
818 // 2-Pass code
819
820 static int init_pass2(MpegEncContext *s)
821 {
822     RateControlContext *rcc= &s->rc_context;
823     AVCodecContext *a= s->avctx;
824     int i, toobig;
825     double fps= 1/av_q2d(s->avctx->time_base);
826     double complexity[5]={0,0,0,0,0};   // aproximate bits at quant=1
827     uint64_t const_bits[5]={0,0,0,0,0}; // quantizer independent bits
828     uint64_t all_const_bits;
829     uint64_t all_available_bits= (uint64_t)(s->bit_rate*(double)rcc->num_entries/fps);
830     double rate_factor=0;
831     double step;
832     //int last_i_frame=-10000000;
833     const int filter_size= (int)(a->qblur*4) | 1;
834     double expected_bits;
835     double *qscale, *blurred_qscale, qscale_sum;
836
837     /* find complexity & const_bits & decide the pict_types */
838     for(i=0; i<rcc->num_entries; i++){
839         RateControlEntry *rce= &rcc->entry[i];
840
841         rce->new_pict_type= rce->pict_type;
842         rcc->i_cplx_sum [rce->pict_type] += rce->i_tex_bits*rce->qscale;
843         rcc->p_cplx_sum [rce->pict_type] += rce->p_tex_bits*rce->qscale;
844         rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
845         rcc->frame_count[rce->pict_type] ++;
846
847         complexity[rce->new_pict_type]+= (rce->i_tex_bits+ rce->p_tex_bits)*(double)rce->qscale;
848         const_bits[rce->new_pict_type]+= rce->mv_bits + rce->misc_bits;
849     }
850     all_const_bits= const_bits[AV_PICTURE_TYPE_I] + const_bits[AV_PICTURE_TYPE_P] + const_bits[AV_PICTURE_TYPE_B];
851
852     if(all_available_bits < all_const_bits){
853         av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
854         return -1;
855     }
856
857     qscale= av_malloc(sizeof(double)*rcc->num_entries);
858     blurred_qscale= av_malloc(sizeof(double)*rcc->num_entries);
859     toobig = 0;
860
861     for(step=256*256; step>0.0000001; step*=0.5){
862         expected_bits=0;
863         rate_factor+= step;
864
865         rcc->buffer_index= s->avctx->rc_buffer_size/2;
866
867         /* find qscale */
868         for(i=0; i<rcc->num_entries; i++){
869             RateControlEntry *rce= &rcc->entry[i];
870             qscale[i]= get_qscale(s, &rcc->entry[i], rate_factor, i);
871             rcc->last_qscale_for[rce->pict_type] = qscale[i];
872         }
873         assert(filter_size%2==1);
874
875         /* fixed I/B QP relative to P mode */
876         for(i=FFMAX(0, rcc->num_entries-300); i<rcc->num_entries; i++){
877             RateControlEntry *rce= &rcc->entry[i];
878
879             qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
880         }
881
882         for(i=rcc->num_entries-1; i>=0; i--){
883             RateControlEntry *rce= &rcc->entry[i];
884
885             qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
886         }
887
888         /* smooth curve */
889         for(i=0; i<rcc->num_entries; i++){
890             RateControlEntry *rce= &rcc->entry[i];
891             const int pict_type= rce->new_pict_type;
892             int j;
893             double q=0.0, sum=0.0;
894
895             for(j=0; j<filter_size; j++){
896                 int index= i+j-filter_size/2;
897                 double d= index-i;
898                 double coeff= a->qblur==0 ? 1.0 : exp(-d*d/(a->qblur * a->qblur));
899
900                 if(index < 0 || index >= rcc->num_entries) continue;
901                 if(pict_type != rcc->entry[index].new_pict_type) continue;
902                 q+= qscale[index] * coeff;
903                 sum+= coeff;
904             }
905             blurred_qscale[i]= q/sum;
906         }
907
908         /* find expected bits */
909         for(i=0; i<rcc->num_entries; i++){
910             RateControlEntry *rce= &rcc->entry[i];
911             double bits;
912             rce->new_qscale= modify_qscale(s, rce, blurred_qscale[i], i);
913             bits= qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
914 //printf("%d %f\n", rce->new_bits, blurred_qscale[i]);
915             bits += 8*ff_vbv_update(s, bits);
916
917             rce->expected_bits= expected_bits;
918             expected_bits += bits;
919         }
920
921         /*
922         av_log(s->avctx, AV_LOG_INFO,
923             "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
924             expected_bits, (int)all_available_bits, rate_factor);
925         */
926         if(expected_bits > all_available_bits) {
927             rate_factor-= step;
928             ++toobig;
929         }
930     }
931     av_free(qscale);
932     av_free(blurred_qscale);
933
934     /* check bitrate calculations and print info */
935     qscale_sum = 0.0;
936     for(i=0; i<rcc->num_entries; i++){
937         /* av_log(s->avctx, AV_LOG_DEBUG, "[lavc rc] entry[%d].new_qscale = %.3f  qp = %.3f\n",
938             i, rcc->entry[i].new_qscale, rcc->entry[i].new_qscale / FF_QP2LAMBDA); */
939         qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA, s->avctx->qmin, s->avctx->qmax);
940     }
941     assert(toobig <= 40);
942     av_log(s->avctx, AV_LOG_DEBUG,
943         "[lavc rc] requested bitrate: %d bps  expected bitrate: %d bps\n",
944         s->bit_rate,
945         (int)(expected_bits / ((double)all_available_bits/s->bit_rate)));
946     av_log(s->avctx, AV_LOG_DEBUG,
947         "[lavc rc] estimated target average qp: %.3f\n",
948         (float)qscale_sum / rcc->num_entries);
949     if (toobig == 0) {
950         av_log(s->avctx, AV_LOG_INFO,
951             "[lavc rc] Using all of requested bitrate is not "
952             "necessary for this video with these parameters.\n");
953     } else if (toobig == 40) {
954         av_log(s->avctx, AV_LOG_ERROR,
955             "[lavc rc] Error: bitrate too low for this video "
956             "with these parameters.\n");
957         return -1;
958     } else if (fabs(expected_bits/all_available_bits - 1.0) > 0.01) {
959         av_log(s->avctx, AV_LOG_ERROR,
960             "[lavc rc] Error: 2pass curve failed to converge\n");
961         return -1;
962     }
963
964     return 0;
965 }