git.sesse.net Git - ffmpeg/blob - 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     av_dlog(s, "%d %f %d %f %f\n",
 274             buffer_size, rcc->buffer_index, frame_size, min_rate, max_rate);
 275     if(buffer_size){
 276         int left;
 277
 278         rcc->buffer_index-= frame_size;
 279         if(rcc->buffer_index < 0){
 280             av_log(s->avctx, AV_LOG_ERROR, "rc buffer underflow\n");
 281             rcc->buffer_index= 0;
 282         }
 283
 284         left= buffer_size - rcc->buffer_index - 1;
 285         rcc->buffer_index += av_clip(left, min_rate, max_rate);
 286
 287         if(rcc->buffer_index > buffer_size){
 288             int stuffing= ceil((rcc->buffer_index - buffer_size)/8);
 289
 290             if(stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
 291                 stuffing=4;
 292             rcc->buffer_index -= 8*stuffing;
 293
 294             if(s->avctx->debug & FF_DEBUG_RC)
 295                 av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n", stuffing);
 296
 297             return stuffing;
 298         }
 299     }
 300     return 0;
 301 }
 302
 303 /**
 304  * Modify the bitrate curve from pass1 for one frame.
 305  */
 306 static double get_qscale(MpegEncContext *s, RateControlEntry *rce, double rate_factor, int frame_num){
 307     RateControlContext *rcc= &s->rc_context;
 308     AVCodecContext *a= s->avctx;
 309     double q, bits;
 310     const int pict_type= rce->new_pict_type;
 311     const double mb_num= s->mb_num;
 312     int i;
 313
 314     double const_values[]={
 315         M_PI,
 316         M_E,
 317         rce->i_tex_bits*rce->qscale,
 318         rce->p_tex_bits*rce->qscale,
 319         (rce->i_tex_bits + rce->p_tex_bits)*(double)rce->qscale,
 320         rce->mv_bits/mb_num,
 321         rce->pict_type == AV_PICTURE_TYPE_B ? (rce->f_code + rce->b_code)*0.5 : rce->f_code,
 322         rce->i_count/mb_num,
 323         rce->mc_mb_var_sum/mb_num,
 324         rce->mb_var_sum/mb_num,
 325         rce->pict_type == AV_PICTURE_TYPE_I,
 326         rce->pict_type == AV_PICTURE_TYPE_P,
 327         rce->pict_type == AV_PICTURE_TYPE_B,
 328         rcc->qscale_sum[pict_type] / (double)rcc->frame_count[pict_type],
 329         a->qcompress,
 330 /*        rcc->last_qscale_for[AV_PICTURE_TYPE_I],
 331         rcc->last_qscale_for[AV_PICTURE_TYPE_P],
 332         rcc->last_qscale_for[AV_PICTURE_TYPE_B],
 333         rcc->next_non_b_qscale,*/
 334         rcc->i_cplx_sum[AV_PICTURE_TYPE_I] / (double)rcc->frame_count[AV_PICTURE_TYPE_I],
 335         rcc->i_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
 336         rcc->p_cplx_sum[AV_PICTURE_TYPE_P] / (double)rcc->frame_count[AV_PICTURE_TYPE_P],
 337         rcc->p_cplx_sum[AV_PICTURE_TYPE_B] / (double)rcc->frame_count[AV_PICTURE_TYPE_B],
 338         (rcc->i_cplx_sum[pict_type] + rcc->p_cplx_sum[pict_type]) / (double)rcc->frame_count[pict_type],
 339         0
 340     };
 341
 342     bits = av_expr_eval(rcc->rc_eq_eval, const_values, rce);
 343     if (isnan(bits)) {
 344         av_log(s->avctx, AV_LOG_ERROR, "Error evaluating rc_eq \"%s\"\n", s->avctx->rc_eq);
 345         return -1;
 346     }
 347
 348     rcc->pass1_rc_eq_output_sum+= bits;
 349     bits*=rate_factor;
 350     if(bits<0.0) bits=0.0;
 351     bits+= 1.0; //avoid 1/0 issues
 352
 353     /* user override */
 354     for(i=0; i<s->avctx->rc_override_count; i++){
 355         RcOverride *rco= s->avctx->rc_override;
 356         if(rco[i].start_frame > frame_num) continue;
 357         if(rco[i].end_frame   < frame_num) continue;
 358
 359         if(rco[i].qscale)
 360             bits= qp2bits(rce, rco[i].qscale); //FIXME move at end to really force it?
 361         else
 362             bits*= rco[i].quality_factor;
 363     }
 364
 365     q= bits2qp(rce, bits);
 366
 367     /* I/B difference */
 368     if     (pict_type==AV_PICTURE_TYPE_I && s->avctx->i_quant_factor<0.0)
 369         q= -q*s->avctx->i_quant_factor + s->avctx->i_quant_offset;
 370     else if(pict_type==AV_PICTURE_TYPE_B && s->avctx->b_quant_factor<0.0)
 371         q= -q*s->avctx->b_quant_factor + s->avctx->b_quant_offset;
 372     if(q<1) q=1;
 373
 374     return q;
 375 }
 376
 377 static double get_diff_limited_q(MpegEncContext *s, RateControlEntry *rce, double q){
 378     RateControlContext *rcc= &s->rc_context;
 379     AVCodecContext *a= s->avctx;
 380     const int pict_type= rce->new_pict_type;
 381     const double last_p_q    = rcc->last_qscale_for[AV_PICTURE_TYPE_P];
 382     const double last_non_b_q= rcc->last_qscale_for[rcc->last_non_b_pict_type];
 383
 384     if     (pict_type==AV_PICTURE_TYPE_I && (a->i_quant_factor>0.0 || rcc->last_non_b_pict_type==AV_PICTURE_TYPE_P))
 385         q= last_p_q    *FFABS(a->i_quant_factor) + a->i_quant_offset;
 386     else if(pict_type==AV_PICTURE_TYPE_B && a->b_quant_factor>0.0)
 387         q= last_non_b_q*    a->b_quant_factor  + a->b_quant_offset;
 388     if(q<1) q=1;
 389
 390     /* last qscale / qdiff stuff */
 391     if(rcc->last_non_b_pict_type==pict_type || pict_type!=AV_PICTURE_TYPE_I){
 392         double last_q= rcc->last_qscale_for[pict_type];
 393         const int maxdiff= FF_QP2LAMBDA * a->max_qdiff;
 394
 395         if     (q > last_q + maxdiff) q= last_q + maxdiff;
 396         else if(q < last_q - maxdiff) q= last_q - maxdiff;
 397     }
 398
 399     rcc->last_qscale_for[pict_type]= q; //Note we cannot do that after blurring
 400
 401     if(pict_type!=AV_PICTURE_TYPE_B)
 402         rcc->last_non_b_pict_type= pict_type;
 403
 404     return q;
 405 }
 406
 407 /**
 408  * Get the qmin & qmax for pict_type.
 409  */
 410 static void get_qminmax(int *qmin_ret, int *qmax_ret, MpegEncContext *s, int pict_type){
 411     int qmin= s->avctx->lmin;
 412     int qmax= s->avctx->lmax;
 413
 414     assert(qmin <= qmax);
 415
 416     if(pict_type==AV_PICTURE_TYPE_B){
 417         qmin= (int)(qmin*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
 418         qmax= (int)(qmax*FFABS(s->avctx->b_quant_factor)+s->avctx->b_quant_offset + 0.5);
 419     }else if(pict_type==AV_PICTURE_TYPE_I){
 420         qmin= (int)(qmin*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
 421         qmax= (int)(qmax*FFABS(s->avctx->i_quant_factor)+s->avctx->i_quant_offset + 0.5);
 422     }
 423
 424     qmin= av_clip(qmin, 1, FF_LAMBDA_MAX);
 425     qmax= av_clip(qmax, 1, FF_LAMBDA_MAX);
 426
 427     if(qmax<qmin) qmax= qmin;
 428
 429     *qmin_ret= qmin;
 430     *qmax_ret= qmax;
 431 }
 432
 433 static double modify_qscale(MpegEncContext *s, RateControlEntry *rce, double q, int frame_num){
 434     RateControlContext *rcc= &s->rc_context;
 435     int qmin, qmax;
 436     const int pict_type= rce->new_pict_type;
 437     const double buffer_size= s->avctx->rc_buffer_size;
 438     const double fps= 1/av_q2d(s->avctx->time_base);
 439     const double min_rate= s->avctx->rc_min_rate / fps;
 440     const double max_rate= s->avctx->rc_max_rate / fps;
 441
 442     get_qminmax(&qmin, &qmax, s, pict_type);
 443
 444     /* modulation */
 445     if(s->avctx->rc_qmod_freq && frame_num%s->avctx->rc_qmod_freq==0 && pict_type==AV_PICTURE_TYPE_P)
 446         q*= s->avctx->rc_qmod_amp;
 447
 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     av_dlog(s, "q:%f max:%f min:%f size:%f index:%f agr:%f\n",
 484             q, max_rate, min_rate, buffer_size, rcc->buffer_index,
 485             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 static void update_predictor(Predictor *p, double q, double var, double size)
 514 {
 515     double new_coeff= size*q / (var + 1);
 516     if(var<10) return;
 517
 518     p->count*= p->decay;
 519     p->coeff*= p->decay;
 520     p->count++;
 521     p->coeff+= new_coeff;
 522 }
 523
 524 static void adaptive_quantization(MpegEncContext *s, double q){
 525     int i;
 526     const float lumi_masking= s->avctx->lumi_masking / (128.0*128.0);
 527     const float dark_masking= s->avctx->dark_masking / (128.0*128.0);
 528     const float temp_cplx_masking= s->avctx->temporal_cplx_masking;
 529     const float spatial_cplx_masking = s->avctx->spatial_cplx_masking;
 530     const float p_masking = s->avctx->p_masking;
 531     const float border_masking = s->avctx->border_masking;
 532     float bits_sum= 0.0;
 533     float cplx_sum= 0.0;
 534     float *cplx_tab = s->cplx_tab;
 535     float *bits_tab = s->bits_tab;
 536     const int qmin= s->avctx->mb_lmin;
 537     const int qmax= s->avctx->mb_lmax;
 538     Picture * const pic= &s->current_picture;
 539     const int mb_width = s->mb_width;
 540     const int mb_height = s->mb_height;
 541
 542     for(i=0; i<s->mb_num; i++){
 543         const int mb_xy= s->mb_index2xy[i];
 544         float temp_cplx= sqrt(pic->mc_mb_var[mb_xy]); //FIXME merge in pow()
 545         float spat_cplx= sqrt(pic->mb_var[mb_xy]);
 546         const int lumi= pic->mb_mean[mb_xy];
 547         float bits, cplx, factor;
 548         int mb_x = mb_xy % s->mb_stride;
 549         int mb_y = mb_xy / s->mb_stride;
 550         int mb_distance;
 551         float mb_factor = 0.0;
 552         if(spat_cplx < 4) spat_cplx= 4; //FIXME finetune
 553         if(temp_cplx < 4) temp_cplx= 4; //FIXME finetune
 554
 555         if((s->mb_type[mb_xy]&CANDIDATE_MB_TYPE_INTRA)){//FIXME hq mode
 556             cplx= spat_cplx;
 557             factor= 1.0 + p_masking;
 558         }else{
 559             cplx= temp_cplx;
 560             factor= pow(temp_cplx, - temp_cplx_masking);
 561         }
 562         factor*=pow(spat_cplx, - spatial_cplx_masking);
 563
 564         if(lumi>127)
 565             factor*= (1.0 - (lumi-128)*(lumi-128)*lumi_masking);
 566         else
 567             factor*= (1.0 - (lumi-128)*(lumi-128)*dark_masking);
 568
 569         if(mb_x < mb_width/5){
 570             mb_distance = mb_width/5 - mb_x;
 571             mb_factor = (float)mb_distance / (float)(mb_width/5);
 572         }else if(mb_x > 4*mb_width/5){
 573             mb_distance = mb_x - 4*mb_width/5;
 574             mb_factor = (float)mb_distance / (float)(mb_width/5);
 575         }
 576         if(mb_y < mb_height/5){
 577             mb_distance = mb_height/5 - mb_y;
 578             mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
 579         }else if(mb_y > 4*mb_height/5){
 580             mb_distance = mb_y - 4*mb_height/5;
 581             mb_factor = FFMAX(mb_factor, (float)mb_distance / (float)(mb_height/5));
 582         }
 583
 584         factor*= 1.0 - border_masking*mb_factor;
 585
 586         if(factor<0.00001) factor= 0.00001;
 587
 588         bits= cplx*factor;
 589         cplx_sum+= cplx;
 590         bits_sum+= bits;
 591         cplx_tab[i]= cplx;
 592         bits_tab[i]= bits;
 593     }
 594
 595     /* handle qmin/qmax clipping */
 596     if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
 597         float factor= bits_sum/cplx_sum;
 598         for(i=0; i<s->mb_num; i++){
 599             float newq= q*cplx_tab[i]/bits_tab[i];
 600             newq*= factor;
 601
 602             if     (newq > qmax){
 603                 bits_sum -= bits_tab[i];
 604                 cplx_sum -= cplx_tab[i]*q/qmax;
 605             }
 606             else if(newq < qmin){
 607                 bits_sum -= bits_tab[i];
 608                 cplx_sum -= cplx_tab[i]*q/qmin;
 609             }
 610         }
 611         if(bits_sum < 0.001) bits_sum= 0.001;
 612         if(cplx_sum < 0.001) cplx_sum= 0.001;
 613     }
 614
 615     for(i=0; i<s->mb_num; i++){
 616         const int mb_xy= s->mb_index2xy[i];
 617         float newq= q*cplx_tab[i]/bits_tab[i];
 618         int intq;
 619
 620         if(s->flags&CODEC_FLAG_NORMALIZE_AQP){
 621             newq*= bits_sum/cplx_sum;
 622         }
 623
 624         intq= (int)(newq + 0.5);
 625
 626         if     (intq > qmax) intq= qmax;
 627         else if(intq < qmin) intq= qmin;
 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         /* update predictors */
 673     if(picture_number>2 && !dry_run){
 674         const int last_var= s->last_pict_type == AV_PICTURE_TYPE_I ? rcc->last_mb_var_sum : rcc->last_mc_mb_var_sum;
 675         update_predictor(&rcc->pred[s->last_pict_type], rcc->last_qscale, sqrt(last_var), s->frame_bits);
 676     }
 677
 678     if(s->flags&CODEC_FLAG_PASS2){
 679         assert(picture_number>=0);
 680         assert(picture_number<rcc->num_entries);
 681         rce= &rcc->entry[picture_number];
 682         wanted_bits= rce->expected_bits;
 683     }else{
 684         Picture *dts_pic;
 685         rce= &local_rce;
 686
 687         //FIXME add a dts field to AVFrame and ensure its set and use it here instead of reordering
 688         //but the reordering is simpler for now until h.264 b pyramid must be handeld
 689         if(s->pict_type == AV_PICTURE_TYPE_B || s->low_delay)
 690             dts_pic= s->current_picture_ptr;
 691         else
 692             dts_pic= s->last_picture_ptr;
 693
 694         if (!dts_pic || dts_pic->f.pts == AV_NOPTS_VALUE)
 695             wanted_bits= (uint64_t)(s->bit_rate*(double)picture_number/fps);
 696         else
 697             wanted_bits = (uint64_t)(s->bit_rate*(double)dts_pic->f.pts / 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 == AV_PICTURE_TYPE_I ? 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!=AV_PICTURE_TYPE_I)
 709             assert(pict_type == rce->new_pict_type);
 710
 711         q= rce->new_qscale / br_compensation;
 712         av_dlog(s, "%f %f %f last:%d var:%d type:%d//\n", q, rce->new_qscale,
 713                 br_compensation, s->frame_bits, var, pict_type);
 714     }else{
 715         rce->pict_type=
 716         rce->new_pict_type= pict_type;
 717         rce->mc_mb_var_sum= pic->mc_mb_var_sum;
 718         rce->mb_var_sum   = pic->   mb_var_sum;
 719         rce->qscale   = FF_QP2LAMBDA * 2;
 720         rce->f_code   = s->f_code;
 721         rce->b_code   = s->b_code;
 722         rce->misc_bits= 1;
 723
 724         bits= predict_size(&rcc->pred[pict_type], rce->qscale, sqrt(var));
 725         if(pict_type== AV_PICTURE_TYPE_I){
 726             rce->i_count   = s->mb_num;
 727             rce->i_tex_bits= bits;
 728             rce->p_tex_bits= 0;
 729             rce->mv_bits= 0;
 730         }else{
 731             rce->i_count   = 0; //FIXME we do know this approx
 732             rce->i_tex_bits= 0;
 733             rce->p_tex_bits= bits*0.9;
 734
 735             rce->mv_bits= bits*0.1;
 736         }
 737         rcc->i_cplx_sum [pict_type] += rce->i_tex_bits*rce->qscale;
 738         rcc->p_cplx_sum [pict_type] += rce->p_tex_bits*rce->qscale;
 739         rcc->mv_bits_sum[pict_type] += rce->mv_bits;
 740         rcc->frame_count[pict_type] ++;
 741
 742         bits= rce->i_tex_bits + rce->p_tex_bits;
 743         rate_factor= rcc->pass1_wanted_bits/rcc->pass1_rc_eq_output_sum * br_compensation;
 744
 745         q= get_qscale(s, rce, rate_factor, picture_number);
 746         if (q < 0)
 747             return -1;
 748
 749         assert(q>0.0);
 750         q= get_diff_limited_q(s, rce, q);
 751         assert(q>0.0);
 752
 753         if(pict_type==AV_PICTURE_TYPE_P || s->intra_only){ //FIXME type dependent blur like in 2-pass
 754             rcc->short_term_qsum*=a->qblur;
 755             rcc->short_term_qcount*=a->qblur;
 756
 757             rcc->short_term_qsum+= q;
 758             rcc->short_term_qcount++;
 759             q= short_term_q= rcc->short_term_qsum/rcc->short_term_qcount;
 760         }
 761         assert(q>0.0);
 762
 763         q= modify_qscale(s, rce, q, picture_number);
 764
 765         rcc->pass1_wanted_bits+= s->bit_rate/fps;
 766
 767         assert(q>0.0);
 768     }
 769
 770     if(s->avctx->debug&FF_DEBUG_RC){
 771         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",
 772         av_get_picture_type_char(pict_type), qmin, q, qmax, picture_number, (int)wanted_bits/1000, (int)s->total_bits/1000,
 773         br_compensation, short_term_q, s->frame_bits, pic->mb_var_sum, pic->mc_mb_var_sum, s->bit_rate/1000, (int)fps
 774         );
 775     }
 776
 777     if     (q<qmin) q=qmin;
 778     else if(q>qmax) q=qmax;
 779
 780     if(s->adaptive_quant)
 781         adaptive_quantization(s, q);
 782     else
 783         q= (int)(q + 0.5);
 784
 785     if(!dry_run){
 786         rcc->last_qscale= q;
 787         rcc->last_mc_mb_var_sum= pic->mc_mb_var_sum;
 788         rcc->last_mb_var_sum= pic->mb_var_sum;
 789     }
 790     return q;
 791 }
 792
 793 //----------------------------------------------
 794 // 2-Pass code
 795
 796 static int init_pass2(MpegEncContext *s)
 797 {
 798     RateControlContext *rcc= &s->rc_context;
 799     AVCodecContext *a= s->avctx;
 800     int i, toobig;
 801     double fps= 1/av_q2d(s->avctx->time_base);
 802     double complexity[5]={0,0,0,0,0};   // aproximate bits at quant=1
 803     uint64_t const_bits[5]={0,0,0,0,0}; // quantizer independent bits
 804     uint64_t all_const_bits;
 805     uint64_t all_available_bits= (uint64_t)(s->bit_rate*(double)rcc->num_entries/fps);
 806     double rate_factor=0;
 807     double step;
 808     //int last_i_frame=-10000000;
 809     const int filter_size= (int)(a->qblur*4) | 1;
 810     double expected_bits;
 811     double *qscale, *blurred_qscale, qscale_sum;
 812
 813     /* find complexity & const_bits & decide the pict_types */
 814     for(i=0; i<rcc->num_entries; i++){
 815         RateControlEntry *rce= &rcc->entry[i];
 816
 817         rce->new_pict_type= rce->pict_type;
 818         rcc->i_cplx_sum [rce->pict_type] += rce->i_tex_bits*rce->qscale;
 819         rcc->p_cplx_sum [rce->pict_type] += rce->p_tex_bits*rce->qscale;
 820         rcc->mv_bits_sum[rce->pict_type] += rce->mv_bits;
 821         rcc->frame_count[rce->pict_type] ++;
 822
 823         complexity[rce->new_pict_type]+= (rce->i_tex_bits+ rce->p_tex_bits)*(double)rce->qscale;
 824         const_bits[rce->new_pict_type]+= rce->mv_bits + rce->misc_bits;
 825     }
 826     all_const_bits= const_bits[AV_PICTURE_TYPE_I] + const_bits[AV_PICTURE_TYPE_P] + const_bits[AV_PICTURE_TYPE_B];
 827
 828     if(all_available_bits < all_const_bits){
 829         av_log(s->avctx, AV_LOG_ERROR, "requested bitrate is too low\n");
 830         return -1;
 831     }
 832
 833     qscale= av_malloc(sizeof(double)*rcc->num_entries);
 834     blurred_qscale= av_malloc(sizeof(double)*rcc->num_entries);
 835     toobig = 0;
 836
 837     for(step=256*256; step>0.0000001; step*=0.5){
 838         expected_bits=0;
 839         rate_factor+= step;
 840
 841         rcc->buffer_index= s->avctx->rc_buffer_size/2;
 842
 843         /* find qscale */
 844         for(i=0; i<rcc->num_entries; i++){
 845             RateControlEntry *rce= &rcc->entry[i];
 846             qscale[i]= get_qscale(s, &rcc->entry[i], rate_factor, i);
 847             rcc->last_qscale_for[rce->pict_type] = qscale[i];
 848         }
 849         assert(filter_size%2==1);
 850
 851         /* fixed I/B QP relative to P mode */
 852         for(i=rcc->num_entries-1; i>=0; i--){
 853             RateControlEntry *rce= &rcc->entry[i];
 854
 855             qscale[i]= get_diff_limited_q(s, rce, qscale[i]);
 856         }
 857
 858         /* smooth curve */
 859         for(i=0; i<rcc->num_entries; i++){
 860             RateControlEntry *rce= &rcc->entry[i];
 861             const int pict_type= rce->new_pict_type;
 862             int j;
 863             double q=0.0, sum=0.0;
 864
 865             for(j=0; j<filter_size; j++){
 866                 int index= i+j-filter_size/2;
 867                 double d= index-i;
 868                 double coeff= a->qblur==0 ? 1.0 : exp(-d*d/(a->qblur * a->qblur));
 869
 870                 if(index < 0 || index >= rcc->num_entries) continue;
 871                 if(pict_type != rcc->entry[index].new_pict_type) continue;
 872                 q+= qscale[index] * coeff;
 873                 sum+= coeff;
 874             }
 875             blurred_qscale[i]= q/sum;
 876         }
 877
 878         /* find expected bits */
 879         for(i=0; i<rcc->num_entries; i++){
 880             RateControlEntry *rce= &rcc->entry[i];
 881             double bits;
 882             rce->new_qscale= modify_qscale(s, rce, blurred_qscale[i], i);
 883             bits= qp2bits(rce, rce->new_qscale) + rce->mv_bits + rce->misc_bits;
 884             bits += 8*ff_vbv_update(s, bits);
 885
 886             rce->expected_bits= expected_bits;
 887             expected_bits += bits;
 888         }
 889
 890         av_dlog(s->avctx,
 891                 "expected_bits: %f all_available_bits: %d rate_factor: %f\n",
 892                 expected_bits, (int)all_available_bits, rate_factor);
 893         if(expected_bits > all_available_bits) {
 894             rate_factor-= step;
 895             ++toobig;
 896         }
 897     }
 898     av_free(qscale);
 899     av_free(blurred_qscale);
 900
 901     /* check bitrate calculations and print info */
 902     qscale_sum = 0.0;
 903     for(i=0; i<rcc->num_entries; i++){
 904         av_dlog(s, "[lavc rc] entry[%d].new_qscale = %.3f  qp = %.3f\n",
 905                 i,
 906                 rcc->entry[i].new_qscale,
 907                 rcc->entry[i].new_qscale / FF_QP2LAMBDA);
 908         qscale_sum += av_clip(rcc->entry[i].new_qscale / FF_QP2LAMBDA, s->avctx->qmin, s->avctx->qmax);
 909     }
 910     assert(toobig <= 40);
 911     av_log(s->avctx, AV_LOG_DEBUG,
 912         "[lavc rc] requested bitrate: %d bps  expected bitrate: %d bps\n",
 913         s->bit_rate,
 914         (int)(expected_bits / ((double)all_available_bits/s->bit_rate)));
 915     av_log(s->avctx, AV_LOG_DEBUG,
 916         "[lavc rc] estimated target average qp: %.3f\n",
 917         (float)qscale_sum / rcc->num_entries);
 918     if (toobig == 0) {
 919         av_log(s->avctx, AV_LOG_INFO,
 920             "[lavc rc] Using all of requested bitrate is not "
 921             "necessary for this video with these parameters.\n");
 922     } else if (toobig == 40) {
 923         av_log(s->avctx, AV_LOG_ERROR,
 924             "[lavc rc] Error: bitrate too low for this video "
 925             "with these parameters.\n");
 926         return -1;
 927     } else if (fabs(expected_bits/all_available_bits - 1.0) > 0.01) {
 928         av_log(s->avctx, AV_LOG_ERROR,
 929             "[lavc rc] Error: 2pass curve failed to converge\n");
 930         return -1;
 931     }
 932
 933     return 0;
 934 }