git.sesse.net Git - x264/blob - encoder/encoder.c

   1 /*****************************************************************************
   2  * x264: h264 encoder
   3  *****************************************************************************
   4  * Copyright (C) 2003 Laurent Aimar
   5  * $Id: encoder.c,v 1.1 2004/06/03 19:27:08 fenrir Exp $
   6  *
   7  * Authors: Laurent Aimar <fenrir@via.ecp.fr>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, write to the Free Software
  21  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
  22  *****************************************************************************/
  23
  24 #include <stdlib.h>
  25 #include <stdio.h>
  26 #include <string.h>
  27 #include <math.h>
  28
  29 #ifdef __WIN32__
  30 #include <windows.h>
  31 #define pthread_t               HANDLE
  32 #define pthread_create(t,u,f,d) *(t)=CreateThread(NULL,0,f,d,0,NULL)
  33 #define pthread_join(t,s)       { WaitForSingleObject(t,INFINITE); \
  34                                   CloseHandle(t); }
  35 #define HAVE_PTHREAD 1
  36
  37 #elif defined(SYS_BEOS)
  38 #include <kernel/OS.h>
  39 #define pthread_t               thread_id
  40 #define pthread_create(t,u,f,d) { *(t)=spawn_thread(f,"",10,d); \
  41                                   resume_thread(*(t)); }
  42 #define pthread_join(t,s)       wait_for_thread(t,(long*)s)
  43 #define HAVE_PTHREAD 1
  44
  45 #elif HAVE_PTHREAD
  46 #include <pthread.h>
  47 #endif
  48
  49 #include "common/common.h"
  50 #include "common/cpu.h"
  51
  52 #include "set.h"
  53 #include "analyse.h"
  54 #include "ratecontrol.h"
  55 #include "macroblock.h"
  56
  57 #if VISUALIZE
  58 #include "common/visualize.h"
  59 #endif
  60
  61 //#define DEBUG_MB_TYPE
  62 //#define DEBUG_DUMP_FRAME
  63 //#define DEBUG_BENCHMARK
  64
  65 #ifdef DEBUG_BENCHMARK
  66 static int64_t i_mtime_encode_frame = 0;
  67 static int64_t i_mtime_analyse = 0;
  68 static int64_t i_mtime_encode = 0;
  69 static int64_t i_mtime_write = 0;
  70 static int64_t i_mtime_filter = 0;
  71 #define TIMER_START( d ) \
  72     { \
  73         int64_t d##start = x264_mdate();
  74
  75 #define TIMER_STOP( d ) \
  76         d += x264_mdate() - d##start;\
  77     }
  78 #else
  79 #define TIMER_START( d )
  80 #define TIMER_STOP( d )
  81 #endif
  82
  83 #define NALU_OVERHEAD 5 // startcode + NAL type costs 5 bytes per frame
  84
  85 /****************************************************************************
  86  *
  87  ******************************* x264 libs **********************************
  88  *
  89  ****************************************************************************/
  90 static float x264_psnr( int64_t i_sqe, int64_t i_size )
  91 {
  92     double f_mse = (double)i_sqe / ((double)65025.0 * (double)i_size);
  93     if( f_mse <= 0.0000000001 ) /* Max 100dB */
  94         return 100;
  95
  96     return (float)(-10.0 * log( f_mse ) / log( 10.0 ));
  97 }
  98
  99 #ifdef DEBUG_DUMP_FRAME
 100 static void x264_frame_dump( x264_t *h, x264_frame_t *fr, char *name )
 101 {
 102     FILE * f = fopen( name, "a" );
 103     int i, y;
 104
 105     fseek( f, 0, SEEK_END );
 106
 107     for( i = 0; i < fr->i_plane; i++ )
 108     {
 109         for( y = 0; y < h->param.i_height / ( i == 0 ? 1 : 2 ); y++ )
 110         {
 111             fwrite( &fr->plane[i][y*fr->i_stride[i]], 1, h->param.i_width / ( i == 0 ? 1 : 2 ), f );
 112         }
 113     }
 114     fclose( f );
 115 }
 116 #endif
 117
 118
 119 /* Fill "default" values */
 120 static void x264_slice_header_init( x264_t *h, x264_slice_header_t *sh,
 121                                     x264_sps_t *sps, x264_pps_t *pps,
 122                                     int i_type, int i_idr_pic_id, int i_frame, int i_qp )
 123 {
 124     x264_param_t *param = &h->param;
 125     int i;
 126
 127     /* First we fill all field */
 128     sh->sps = sps;
 129     sh->pps = pps;
 130
 131     sh->i_type      = i_type;
 132     sh->i_first_mb  = 0;
 133     sh->i_last_mb   = h->sps->i_mb_width * h->sps->i_mb_height;
 134     sh->i_pps_id    = pps->i_id;
 135
 136     sh->i_frame_num = i_frame;
 137
 138     sh->b_field_pic = 0;    /* Not field support for now */
 139     sh->b_bottom_field = 1; /* not yet used */
 140
 141     sh->i_idr_pic_id = i_idr_pic_id;
 142
 143     /* poc stuff, fixed later */
 144     sh->i_poc_lsb = 0;
 145     sh->i_delta_poc_bottom = 0;
 146     sh->i_delta_poc[0] = 0;
 147     sh->i_delta_poc[1] = 0;
 148
 149     sh->i_redundant_pic_cnt = 0;
 150
 151     sh->b_direct_spatial_mv_pred = ( param->analyse.i_direct_mv_pred == X264_DIRECT_PRED_SPATIAL );
 152
 153     sh->b_num_ref_idx_override = 0;
 154     sh->i_num_ref_idx_l0_active = 1;
 155     sh->i_num_ref_idx_l1_active = 1;
 156
 157     sh->b_ref_pic_list_reordering_l0 = h->b_ref_reorder[0];
 158     sh->b_ref_pic_list_reordering_l1 = h->b_ref_reorder[1];
 159
 160     /* If the ref list isn't in the default order, construct reordering header */
 161     /* List1 reordering isn't needed yet */
 162     if( sh->b_ref_pic_list_reordering_l0 )
 163     {
 164         int pred_frame_num = i_frame;
 165         for( i = 0; i < h->i_ref0; i++ )
 166         {
 167             int diff = h->fref0[i]->i_frame_num - pred_frame_num;
 168             if( diff == 0 )
 169                 x264_log( h, X264_LOG_ERROR, "diff frame num == 0\n" );
 170             sh->ref_pic_list_order[0][i].idc = ( diff > 0 );
 171             sh->ref_pic_list_order[0][i].arg = abs( diff ) - 1;
 172             pred_frame_num = h->fref0[i]->i_frame_num;
 173         }
 174     }
 175
 176     sh->i_cabac_init_idc = param->i_cabac_init_idc;
 177
 178     sh->i_qp = i_qp;
 179     sh->i_qp_delta = i_qp - pps->i_pic_init_qp;
 180     sh->b_sp_for_swidth = 0;
 181     sh->i_qs_delta = 0;
 182
 183     /* If effective qp <= 15, deblocking would have no effect anyway */
 184     if( param->b_deblocking_filter
 185         && ( h->mb.b_variable_qp
 186         || 15 < i_qp + X264_MAX(param->i_deblocking_filter_alphac0, param->i_deblocking_filter_beta) ) )
 187     {
 188         sh->i_disable_deblocking_filter_idc = 0;
 189     }
 190     else
 191     {
 192         sh->i_disable_deblocking_filter_idc = 1;
 193     }
 194     sh->i_alpha_c0_offset = param->i_deblocking_filter_alphac0 << 1;
 195     sh->i_beta_offset = param->i_deblocking_filter_beta << 1;
 196 }
 197
 198 static void x264_slice_header_write( bs_t *s, x264_slice_header_t *sh, int i_nal_ref_idc )
 199 {
 200     int i;
 201
 202     bs_write_ue( s, sh->i_first_mb );
 203     bs_write_ue( s, sh->i_type + 5 );   /* same type things */
 204     bs_write_ue( s, sh->i_pps_id );
 205     bs_write( s, sh->sps->i_log2_max_frame_num, sh->i_frame_num );
 206
 207     if( sh->i_idr_pic_id >= 0 ) /* NAL IDR */
 208     {
 209         bs_write_ue( s, sh->i_idr_pic_id );
 210     }
 211
 212     if( sh->sps->i_poc_type == 0 )
 213     {
 214         bs_write( s, sh->sps->i_log2_max_poc_lsb, sh->i_poc_lsb );
 215         if( sh->pps->b_pic_order && !sh->b_field_pic )
 216         {
 217             bs_write_se( s, sh->i_delta_poc_bottom );
 218         }
 219     }
 220     else if( sh->sps->i_poc_type == 1 && !sh->sps->b_delta_pic_order_always_zero )
 221     {
 222         bs_write_se( s, sh->i_delta_poc[0] );
 223         if( sh->pps->b_pic_order && !sh->b_field_pic )
 224         {
 225             bs_write_se( s, sh->i_delta_poc[1] );
 226         }
 227     }
 228
 229     if( sh->pps->b_redundant_pic_cnt )
 230     {
 231         bs_write_ue( s, sh->i_redundant_pic_cnt );
 232     }
 233
 234     if( sh->i_type == SLICE_TYPE_B )
 235     {
 236         bs_write1( s, sh->b_direct_spatial_mv_pred );
 237     }
 238     if( sh->i_type == SLICE_TYPE_P || sh->i_type == SLICE_TYPE_SP || sh->i_type == SLICE_TYPE_B )
 239     {
 240         bs_write1( s, sh->b_num_ref_idx_override );
 241         if( sh->b_num_ref_idx_override )
 242         {
 243             bs_write_ue( s, sh->i_num_ref_idx_l0_active - 1 );
 244             if( sh->i_type == SLICE_TYPE_B )
 245             {
 246                 bs_write_ue( s, sh->i_num_ref_idx_l1_active - 1 );
 247             }
 248         }
 249     }
 250
 251     /* ref pic list reordering */
 252     if( sh->i_type != SLICE_TYPE_I )
 253     {
 254         bs_write1( s, sh->b_ref_pic_list_reordering_l0 );
 255         if( sh->b_ref_pic_list_reordering_l0 )
 256         {
 257             for( i = 0; i < sh->i_num_ref_idx_l0_active; i++ )
 258             {
 259                 bs_write_ue( s, sh->ref_pic_list_order[0][i].idc );
 260                 bs_write_ue( s, sh->ref_pic_list_order[0][i].arg );
 261
 262             }
 263             bs_write_ue( s, 3 );
 264         }
 265     }
 266     if( sh->i_type == SLICE_TYPE_B )
 267     {
 268         bs_write1( s, sh->b_ref_pic_list_reordering_l1 );
 269         if( sh->b_ref_pic_list_reordering_l1 )
 270         {
 271             for( i = 0; i < sh->i_num_ref_idx_l1_active; i++ )
 272             {
 273                 bs_write_ue( s, sh->ref_pic_list_order[1][i].idc );
 274                 bs_write_ue( s, sh->ref_pic_list_order[1][i].arg );
 275             }
 276             bs_write_ue( s, 3 );
 277         }
 278     }
 279
 280     if( ( sh->pps->b_weighted_pred && ( sh->i_type == SLICE_TYPE_P || sh->i_type == SLICE_TYPE_SP ) ) ||
 281         ( sh->pps->b_weighted_bipred == 1 && sh->i_type == SLICE_TYPE_B ) )
 282     {
 283         /* FIXME */
 284     }
 285
 286     if( i_nal_ref_idc != 0 )
 287     {
 288         if( sh->i_idr_pic_id >= 0 )
 289         {
 290             bs_write1( s, 0 );  /* no output of prior pics flag */
 291             bs_write1( s, 0 );  /* long term reference flag */
 292         }
 293         else
 294         {
 295             bs_write1( s, 0 );  /* adaptive_ref_pic_marking_mode_flag */
 296         }
 297     }
 298
 299     if( sh->pps->b_cabac && sh->i_type != SLICE_TYPE_I )
 300     {
 301         bs_write_ue( s, sh->i_cabac_init_idc );
 302     }
 303     bs_write_se( s, sh->i_qp_delta );      /* slice qp delta */
 304
 305     if( sh->pps->b_deblocking_filter_control )
 306     {
 307         bs_write_ue( s, sh->i_disable_deblocking_filter_idc );
 308         if( sh->i_disable_deblocking_filter_idc != 1 )
 309         {
 310             bs_write_se( s, sh->i_alpha_c0_offset >> 1 );
 311             bs_write_se( s, sh->i_beta_offset >> 1 );
 312         }
 313     }
 314 }
 315
 316 /****************************************************************************
 317  *
 318  ****************************************************************************
 319  ****************************** External API*********************************
 320  ****************************************************************************
 321  *
 322  ****************************************************************************/
 323
 324 static int x264_validate_parameters( x264_t *h )
 325 {
 326     if( h->param.i_width <= 0 || h->param.i_height <= 0 )
 327     {
 328         x264_log( h, X264_LOG_ERROR, "invalid width x height (%dx%d)\n",
 329                   h->param.i_width, h->param.i_height );
 330         return -1;
 331     }
 332
 333     if( h->param.i_width % 2 || h->param.i_height % 2 )
 334     {
 335         x264_log( h, X264_LOG_ERROR, "width or height not divisible by 2 (%dx%d)\n",
 336                   h->param.i_width, h->param.i_height );
 337         return -1;
 338     }
 339     if( h->param.i_csp != X264_CSP_I420 )
 340     {
 341         x264_log( h, X264_LOG_ERROR, "invalid CSP (only I420 supported)\n" );
 342         return -1;
 343     }
 344
 345     h->param.i_threads = x264_clip3( h->param.i_threads, 1, X264_SLICE_MAX );
 346     h->param.i_threads = X264_MIN( h->param.i_threads, (h->param.i_height + 15) / 16 );
 347 #if !(HAVE_PTHREAD)
 348     if( h->param.i_threads > 1 )
 349     {
 350         x264_log( h, X264_LOG_WARNING, "not compiled with pthread support!\n");
 351         x264_log( h, X264_LOG_WARNING, "multislicing anyway, but you won't see any speed gain.\n" );
 352     }
 353 #endif
 354
 355     if( h->param.rc.b_cbr )
 356         h->param.rc.i_rf_constant = 0;
 357     if( h->param.rc.i_rf_constant > 0 )
 358         h->param.rc.i_qp_constant = h->param.rc.i_rf_constant;
 359     h->param.rc.i_rf_constant = x264_clip3( h->param.rc.i_rf_constant, 0, 51 );
 360     h->param.rc.i_qp_constant = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
 361     if( !h->param.rc.b_cbr && h->param.rc.i_qp_constant == 0 )
 362     {
 363         h->mb.b_lossless = 1;
 364         h->param.analyse.b_transform_8x8 = 0;
 365         h->param.i_cqm_preset = X264_CQM_FLAT;
 366         h->param.psz_cqm_file = NULL;
 367         h->param.rc.f_ip_factor = 1;
 368         h->param.rc.f_pb_factor = 1;
 369         h->param.analyse.b_psnr = 0;
 370         h->param.analyse.i_chroma_qp_offset = 0;
 371         h->param.analyse.i_trellis = 0;
 372         h->param.analyse.b_fast_pskip = 0;
 373         h->param.analyse.i_noise_reduction = 0;
 374     }
 375
 376     if( ( h->param.i_width % 16 || h->param.i_height % 16 ) && !h->mb.b_lossless )
 377     {
 378         x264_log( h, X264_LOG_WARNING,
 379                   "width or height not divisible by 16 (%dx%d), compression will suffer.\n",
 380                   h->param.i_width, h->param.i_height );
 381     }
 382
 383     h->param.i_frame_reference = x264_clip3( h->param.i_frame_reference, 1, 16 );
 384     if( h->param.i_keyint_max <= 0 )
 385         h->param.i_keyint_max = 1;
 386     h->param.i_keyint_min = x264_clip3( h->param.i_keyint_min, 1, h->param.i_keyint_max/2+1 );
 387
 388     h->param.i_bframe = x264_clip3( h->param.i_bframe, 0, X264_BFRAME_MAX );
 389     h->param.i_bframe_bias = x264_clip3( h->param.i_bframe_bias, -90, 100 );
 390     h->param.b_bframe_pyramid = h->param.b_bframe_pyramid && h->param.i_bframe > 1;
 391     h->param.b_bframe_adaptive = h->param.b_bframe_adaptive && h->param.i_bframe > 0;
 392
 393     h->param.i_deblocking_filter_alphac0 = x264_clip3( h->param.i_deblocking_filter_alphac0, -6, 6 );
 394     h->param.i_deblocking_filter_beta    = x264_clip3( h->param.i_deblocking_filter_beta, -6, 6 );
 395
 396     h->param.i_cabac_init_idc = x264_clip3( h->param.i_cabac_init_idc, 0, 2 );
 397
 398     if( h->param.i_cqm_preset < X264_CQM_FLAT || h->param.i_cqm_preset > X264_CQM_CUSTOM )
 399         h->param.i_cqm_preset = X264_CQM_FLAT;
 400
 401     if( h->param.analyse.i_me_method < X264_ME_DIA ||
 402         h->param.analyse.i_me_method > X264_ME_ESA )
 403         h->param.analyse.i_me_method = X264_ME_HEX;
 404     if( h->param.analyse.i_me_range < 4 )
 405         h->param.analyse.i_me_range = 4;
 406     if( h->param.analyse.i_me_range > 16 && h->param.analyse.i_me_method <= X264_ME_HEX )
 407         h->param.analyse.i_me_range = 16;
 408     h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 1, 6 );
 409     h->param.analyse.b_bframe_rdo = h->param.analyse.b_bframe_rdo && h->param.analyse.i_subpel_refine >= 6;
 410     h->param.analyse.b_mixed_references = h->param.analyse.b_mixed_references && h->param.i_frame_reference > 1;
 411     h->param.analyse.inter &= X264_ANALYSE_PSUB16x16|X264_ANALYSE_PSUB8x8|X264_ANALYSE_BSUB16x16|
 412                               X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 413     h->param.analyse.intra &= X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 414     if( !(h->param.analyse.inter & X264_ANALYSE_PSUB16x16) )
 415         h->param.analyse.inter &= ~X264_ANALYSE_PSUB8x8;
 416     if( !h->param.analyse.b_transform_8x8 )
 417     {
 418         h->param.analyse.inter &= ~X264_ANALYSE_I8x8;
 419         h->param.analyse.intra &= ~X264_ANALYSE_I8x8;
 420     }
 421     h->param.analyse.i_chroma_qp_offset = x264_clip3(h->param.analyse.i_chroma_qp_offset, -12, 12);
 422     if( !h->param.b_cabac )
 423         h->param.analyse.i_trellis = 0;
 424     h->param.analyse.i_trellis = x264_clip3( h->param.analyse.i_trellis, 0, 2 );
 425     h->param.analyse.i_noise_reduction = x264_clip3( h->param.analyse.i_noise_reduction, 0, 1<<16 );
 426
 427     {
 428         const x264_level_t *l = x264_levels;
 429         while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
 430             l++;
 431         if( l->level_idc == 0 )
 432         {
 433             x264_log( h, X264_LOG_ERROR, "invalid level_idc: %d\n", h->param.i_level_idc );
 434             return -1;
 435         }
 436         if( h->param.analyse.i_mv_range <= 0 )
 437             h->param.analyse.i_mv_range = l->mv_range;
 438         else
 439             h->param.analyse.i_mv_range = x264_clip3(h->param.analyse.i_mv_range, 32, 2048);
 440     }
 441
 442     if( h->param.rc.f_qblur < 0 )
 443         h->param.rc.f_qblur = 0;
 444     if( h->param.rc.f_complexity_blur < 0 )
 445         h->param.rc.f_complexity_blur = 0;
 446
 447     /* ensure the booleans are 0 or 1 so they can be used in math */
 448 #define BOOLIFY(x) h->param.x = !!h->param.x
 449     BOOLIFY( b_cabac );
 450     BOOLIFY( b_deblocking_filter );
 451     BOOLIFY( analyse.b_transform_8x8 );
 452     BOOLIFY( analyse.b_weighted_bipred );
 453     BOOLIFY( analyse.b_bidir_me );
 454     BOOLIFY( analyse.b_chroma_me );
 455     BOOLIFY( analyse.b_fast_pskip );
 456     BOOLIFY( rc.b_cbr );
 457     BOOLIFY( rc.b_stat_write );
 458     BOOLIFY( rc.b_stat_read );
 459 #undef BOOLIFY
 460
 461     return 0;
 462 }
 463
 464 /****************************************************************************
 465  * x264_encoder_open:
 466  ****************************************************************************/
 467 x264_t *x264_encoder_open   ( x264_param_t *param )
 468 {
 469     x264_t *h = x264_malloc( sizeof( x264_t ) );
 470     int i;
 471
 472     memset( h, 0, sizeof( x264_t ) );
 473
 474     /* Create a copy of param */
 475     memcpy( &h->param, param, sizeof( x264_param_t ) );
 476
 477     if( x264_validate_parameters( h ) < 0 )
 478     {
 479         x264_free( h );
 480         return NULL;
 481     }
 482
 483     if( h->param.psz_cqm_file )
 484         if( x264_cqm_parse_file( h, h->param.psz_cqm_file ) < 0 )
 485         {
 486             x264_free( h );
 487             return NULL;
 488         }
 489
 490     if( h->param.rc.psz_stat_out )
 491         h->param.rc.psz_stat_out = strdup( h->param.rc.psz_stat_out );
 492     if( h->param.rc.psz_stat_in )
 493         h->param.rc.psz_stat_in = strdup( h->param.rc.psz_stat_in );
 494     if( h->param.rc.psz_rc_eq )
 495         h->param.rc.psz_rc_eq = strdup( h->param.rc.psz_rc_eq );
 496
 497     /* VUI */
 498     if( h->param.vui.i_sar_width > 0 && h->param.vui.i_sar_height > 0 )
 499     {
 500         int i_w = param->vui.i_sar_width;
 501         int i_h = param->vui.i_sar_height;
 502         int a = i_w, b = i_h;
 503
 504         while( b != 0 )
 505         {
 506             int t = a;
 507
 508             a = b;
 509             b = t % b;
 510         }
 511
 512         i_w /= a;
 513         i_h /= a;
 514         while( i_w > 65535 || i_h > 65535 )
 515         {
 516             i_w /= 2;
 517             i_h /= 2;
 518         }
 519
 520         h->param.vui.i_sar_width = 0;
 521         h->param.vui.i_sar_height = 0;
 522         if( i_w == 0 || i_h == 0 )
 523         {
 524             x264_log( h, X264_LOG_ERROR, "cannot create valid sample aspect ratio\n" );
 525         }
 526         else if( i_w == i_h )
 527         {
 528             x264_log( h, X264_LOG_INFO, "no need for a SAR\n" );
 529         }
 530         else
 531         {
 532             x264_log( h, X264_LOG_INFO, "using SAR=%d/%d\n", i_w, i_h );
 533             h->param.vui.i_sar_width = i_w;
 534             h->param.vui.i_sar_height = i_h;
 535         }
 536     }
 537
 538
 539     /* Init x264_t */
 540     h->out.i_nal = 0;
 541     h->out.i_bitstream = X264_MAX( 1000000, h->param.i_width * h->param.i_height * 1.7
 542         * ( h->param.rc.b_cbr ? pow( 0.5, h->param.rc.i_qp_min )
 543           : pow( 0.5, h->param.rc.i_qp_constant ) * X264_MAX( 1, h->param.rc.f_ip_factor )));
 544     h->out.p_bitstream = x264_malloc( h->out.i_bitstream );
 545
 546     h->i_frame = 0;
 547     h->i_frame_num = 0;
 548     h->i_idr_pic_id = 0;
 549
 550     h->sps = &h->sps_array[0];
 551     x264_sps_init( h->sps, 0, &h->param );
 552
 553     h->pps = &h->pps_array[0];
 554     x264_pps_init( h->pps, 0, &h->param, h->sps);
 555
 556     x264_validate_levels( h );
 557
 558     x264_cqm_init( h );
 559
 560     h->mb.i_mb_count = h->sps->i_mb_width * h->sps->i_mb_height;
 561
 562     /* Init frames. */
 563     h->frames.i_delay = h->param.i_bframe;
 564     h->frames.i_max_ref0 = h->param.i_frame_reference;
 565     h->frames.i_max_ref1 = h->sps->vui.i_num_reorder_frames;
 566     h->frames.i_max_dpb  = h->sps->vui.i_max_dec_frame_buffering + 1;
 567     h->frames.b_have_lowres = !h->param.rc.b_stat_read
 568         && ( h->param.rc.b_cbr || h->param.rc.i_rf_constant || h->param.b_bframe_adaptive );
 569
 570     for( i = 0; i < X264_BFRAME_MAX + 3; i++ )
 571     {
 572         h->frames.current[i] = NULL;
 573         h->frames.next[i]    = NULL;
 574         h->frames.unused[i]  = NULL;
 575     }
 576     for( i = 0; i < 1 + h->frames.i_delay; i++ )
 577     {
 578         h->frames.unused[i] =  x264_frame_new( h );
 579     }
 580     for( i = 0; i < h->frames.i_max_dpb; i++ )
 581     {
 582         h->frames.reference[i] = x264_frame_new( h );
 583     }
 584     h->frames.reference[h->frames.i_max_dpb] = NULL;
 585     h->frames.i_last_idr = - h->param.i_keyint_max;
 586     h->frames.i_input    = 0;
 587     h->frames.last_nonb  = NULL;
 588
 589     h->i_ref0 = 0;
 590     h->i_ref1 = 0;
 591
 592     h->fdec = h->frames.reference[0];
 593
 594     x264_macroblock_cache_init( h );
 595     x264_rdo_init( );
 596
 597     /* init CPU functions */
 598     x264_predict_16x16_init( h->param.cpu, h->predict_16x16 );
 599     x264_predict_8x8c_init( h->param.cpu, h->predict_8x8c );
 600     x264_predict_8x8_init( h->param.cpu, h->predict_8x8 );
 601     x264_predict_4x4_init( h->param.cpu, h->predict_4x4 );
 602
 603     x264_pixel_init( h->param.cpu, &h->pixf );
 604     x264_dct_init( h->param.cpu, &h->dctf );
 605     x264_mc_init( h->param.cpu, &h->mc );
 606     x264_csp_init( h->param.cpu, h->param.i_csp, &h->csp );
 607     x264_quant_init( h, h->param.cpu, &h->quantf );
 608     x264_deblock_init( h->param.cpu, &h->loopf );
 609
 610     memcpy( h->pixf.mbcmp,
 611             ( h->mb.b_lossless || h->param.analyse.i_subpel_refine <= 1 ) ? h->pixf.sad : h->pixf.satd,
 612             sizeof(h->pixf.mbcmp) );
 613
 614     /* rate control */
 615     if( x264_ratecontrol_new( h ) < 0 )
 616         return NULL;
 617
 618     x264_log( h, X264_LOG_INFO, "using cpu capabilities %s%s%s%s%s%s\n",
 619              param->cpu&X264_CPU_MMX ? "MMX " : "",
 620              param->cpu&X264_CPU_MMXEXT ? "MMXEXT " : "",
 621              param->cpu&X264_CPU_SSE ? "SSE " : "",
 622              param->cpu&X264_CPU_SSE2 ? "SSE2 " : "",
 623              param->cpu&X264_CPU_3DNOW ? "3DNow! " : "",
 624              param->cpu&X264_CPU_ALTIVEC ? "Altivec " : "" );
 625
 626     h->thread[0] = h;
 627     for( i = 1; i < param->i_threads; i++ )
 628         h->thread[i] = x264_malloc( sizeof(x264_t) );
 629
 630     return h;
 631 }
 632
 633 /****************************************************************************
 634  * x264_encoder_reconfig:
 635  ****************************************************************************/
 636 int x264_encoder_reconfig( x264_t *h, x264_param_t *param )
 637 {
 638     h->param.i_bframe_bias = param->i_bframe_bias;
 639     h->param.i_deblocking_filter_alphac0 = param->i_deblocking_filter_alphac0;
 640     h->param.i_deblocking_filter_beta    = param->i_deblocking_filter_beta;
 641     h->param.analyse.i_me_method = param->analyse.i_me_method;
 642     h->param.analyse.i_me_range = param->analyse.i_me_range;
 643     h->param.analyse.i_subpel_refine = param->analyse.i_subpel_refine;
 644     h->param.analyse.i_trellis = param->analyse.i_trellis;
 645     h->param.analyse.intra = param->analyse.intra;
 646     h->param.analyse.inter = param->analyse.inter;
 647     if( h->sps->b_direct8x8_inference && h->param.i_bframe
 648         && h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_TEMPORAL )
 649         h->param.analyse.inter &= ~X264_ANALYSE_PSUB8x8;
 650
 651     memcpy( h->pixf.mbcmp,
 652             ( h->mb.b_lossless || h->param.analyse.i_subpel_refine <= 1 ) ? h->pixf.sad : h->pixf.satd,
 653             sizeof(h->pixf.mbcmp) );
 654
 655     return x264_validate_parameters( h );
 656 }
 657
 658 /* internal usage */
 659 static void x264_nal_start( x264_t *h, int i_type, int i_ref_idc )
 660 {
 661     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 662
 663     nal->i_ref_idc = i_ref_idc;
 664     nal->i_type    = i_type;
 665
 666     bs_align_0( &h->out.bs );   /* not needed */
 667
 668     nal->i_payload= 0;
 669     nal->p_payload= &h->out.p_bitstream[bs_pos( &h->out.bs) / 8];
 670 }
 671 static void x264_nal_end( x264_t *h )
 672 {
 673     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 674
 675     bs_align_0( &h->out.bs );   /* not needed */
 676
 677     nal->i_payload = &h->out.p_bitstream[bs_pos( &h->out.bs)/8] - nal->p_payload;
 678
 679     h->out.i_nal++;
 680 }
 681
 682 /****************************************************************************
 683  * x264_encoder_headers:
 684  ****************************************************************************/
 685 int x264_encoder_headers( x264_t *h, x264_nal_t **pp_nal, int *pi_nal )
 686 {
 687     /* init bitstream context */
 688     h->out.i_nal = 0;
 689     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
 690
 691     /* Put SPS and PPS */
 692     if( h->i_frame == 0 )
 693     {
 694         /* identify ourself */
 695         x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
 696         x264_sei_version_write( h, &h->out.bs );
 697         x264_nal_end( h );
 698
 699         /* generate sequence parameters */
 700         x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
 701         x264_sps_write( &h->out.bs, h->sps );
 702         x264_nal_end( h );
 703
 704         /* generate picture parameters */
 705         x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
 706         x264_pps_write( &h->out.bs, h->pps );
 707         x264_nal_end( h );
 708     }
 709     /* now set output*/
 710     *pi_nal = h->out.i_nal;
 711     *pp_nal = &h->out.nal[0];
 712
 713     return 0;
 714 }
 715
 716
 717 static void x264_frame_put( x264_frame_t *list[X264_BFRAME_MAX], x264_frame_t *frame )
 718 {
 719     int i = 0;
 720     while( list[i] ) i++;
 721     list[i] = frame;
 722 }
 723
 724 static void x264_frame_push( x264_frame_t *list[X264_BFRAME_MAX], x264_frame_t *frame )
 725 {
 726     int i = 0;
 727     while( list[i] ) i++;
 728     while( i-- )
 729         list[i+1] = list[i];
 730     list[0] = frame;
 731 }
 732
 733 static x264_frame_t *x264_frame_get( x264_frame_t *list[X264_BFRAME_MAX+1] )
 734 {
 735     x264_frame_t *frame = list[0];
 736     int i;
 737     for( i = 0; list[i]; i++ )
 738         list[i] = list[i+1];
 739     return frame;
 740 }
 741
 742 static void x264_frame_sort( x264_frame_t *list[X264_BFRAME_MAX+1], int b_dts )
 743 {
 744     int i, b_ok;
 745     do {
 746         b_ok = 1;
 747         for( i = 0; list[i+1]; i++ )
 748         {
 749             int dtype = list[i]->i_type - list[i+1]->i_type;
 750             int dtime = list[i]->i_frame - list[i+1]->i_frame;
 751             int swap = b_dts ? dtype > 0 || ( dtype == 0 && dtime > 0 )
 752                              : dtime > 0;
 753             if( swap )
 754             {
 755                 XCHG( x264_frame_t*, list[i], list[i+1] );
 756                 b_ok = 0;
 757             }
 758         }
 759     } while( !b_ok );
 760 }
 761 #define x264_frame_sort_dts(list) x264_frame_sort(list, 1)
 762 #define x264_frame_sort_pts(list) x264_frame_sort(list, 0)
 763
 764 static inline void x264_reference_build_list( x264_t *h, int i_poc, int i_slice_type )
 765 {
 766     int i;
 767     int b_ok;
 768
 769     /* build ref list 0/1 */
 770     h->i_ref0 = 0;
 771     h->i_ref1 = 0;
 772     for( i = 1; i < h->frames.i_max_dpb; i++ )
 773     {
 774         if( h->frames.reference[i]->i_poc >= 0 )
 775         {
 776             if( h->frames.reference[i]->i_poc < i_poc )
 777             {
 778                 h->fref0[h->i_ref0++] = h->frames.reference[i];
 779             }
 780             else if( h->frames.reference[i]->i_poc > i_poc )
 781             {
 782                 h->fref1[h->i_ref1++] = h->frames.reference[i];
 783             }
 784         }
 785     }
 786
 787     /* Order ref0 from higher to lower poc */
 788     do
 789     {
 790         b_ok = 1;
 791         for( i = 0; i < h->i_ref0 - 1; i++ )
 792         {
 793             if( h->fref0[i]->i_poc < h->fref0[i+1]->i_poc )
 794             {
 795                 XCHG( x264_frame_t*, h->fref0[i], h->fref0[i+1] );
 796                 b_ok = 0;
 797                 break;
 798             }
 799         }
 800     } while( !b_ok );
 801     /* Order ref1 from lower to higher poc (bubble sort) for B-frame */
 802     do
 803     {
 804         b_ok = 1;
 805         for( i = 0; i < h->i_ref1 - 1; i++ )
 806         {
 807             if( h->fref1[i]->i_poc > h->fref1[i+1]->i_poc )
 808             {
 809                 XCHG( x264_frame_t*, h->fref1[i], h->fref1[i+1] );
 810                 b_ok = 0;
 811                 break;
 812             }
 813         }
 814     } while( !b_ok );
 815
 816     /* In the standard, a P-frame's ref list is sorted by frame_num.
 817      * We use POC, but check whether explicit reordering is needed */
 818     h->b_ref_reorder[0] =
 819     h->b_ref_reorder[1] = 0;
 820     if( i_slice_type == SLICE_TYPE_P )
 821     {
 822         for( i = 0; i < h->i_ref0 - 1; i++ )
 823             if( h->fref0[i]->i_frame_num < h->fref0[i+1]->i_frame_num )
 824             {
 825                 h->b_ref_reorder[0] = 1;
 826                 break;
 827             }
 828     }
 829
 830     h->i_ref1 = X264_MIN( h->i_ref1, h->frames.i_max_ref1 );
 831     h->i_ref0 = X264_MIN( h->i_ref0, h->frames.i_max_ref0 );
 832     h->i_ref0 = X264_MIN( h->i_ref0, 16 - h->i_ref1 );
 833 }
 834
 835 static inline void x264_reference_update( x264_t *h )
 836 {
 837     int i;
 838
 839     /* apply deblocking filter to the current decoded picture */
 840     if( !h->sh.i_disable_deblocking_filter_idc )
 841     {
 842         TIMER_START( i_mtime_filter );
 843         x264_frame_deblocking_filter( h, h->sh.i_type );
 844         TIMER_STOP( i_mtime_filter );
 845     }
 846     /* expand border */
 847     x264_frame_expand_border( h->fdec );
 848
 849     /* create filtered images */
 850     x264_frame_filter( h->param.cpu, h->fdec );
 851
 852     /* expand border of filtered images */
 853     x264_frame_expand_border_filtered( h->fdec );
 854
 855     /* move lowres copy of the image to the ref frame */
 856     for( i = 0; i < 4; i++)
 857         XCHG( uint8_t*, h->fdec->lowres[i], h->fenc->lowres[i] );
 858
 859     /* adaptive B decision needs a pointer, since it can't use the ref lists */
 860     if( h->sh.i_type != SLICE_TYPE_B )
 861         h->frames.last_nonb = h->fdec;
 862
 863     /* move frame in the buffer */
 864     h->fdec = h->frames.reference[h->frames.i_max_dpb-1];
 865     for( i = h->frames.i_max_dpb-1; i > 0; i-- )
 866     {
 867         h->frames.reference[i] = h->frames.reference[i-1];
 868     }
 869     h->frames.reference[0] = h->fdec;
 870 }
 871
 872 static inline void x264_reference_reset( x264_t *h )
 873 {
 874     int i;
 875
 876     /* reset ref pictures */
 877     for( i = 1; i < h->frames.i_max_dpb; i++ )
 878     {
 879         h->frames.reference[i]->i_poc = -1;
 880     }
 881     h->frames.reference[0]->i_poc = 0;
 882 }
 883
 884 static inline void x264_slice_init( x264_t *h, int i_nal_type, int i_slice_type, int i_global_qp )
 885 {
 886     /* ------------------------ Create slice header  ----------------------- */
 887     if( i_nal_type == NAL_SLICE_IDR )
 888     {
 889         x264_slice_header_init( h, &h->sh, h->sps, h->pps, i_slice_type, h->i_idr_pic_id, h->i_frame_num - 1, i_global_qp );
 890
 891         /* increment id */
 892         h->i_idr_pic_id = ( h->i_idr_pic_id + 1 ) % 65536;
 893     }
 894     else
 895     {
 896         x264_slice_header_init( h, &h->sh, h->sps, h->pps, i_slice_type, -1, h->i_frame_num - 1, i_global_qp );
 897
 898         /* always set the real higher num of ref frame used */
 899         h->sh.b_num_ref_idx_override = 1;
 900         h->sh.i_num_ref_idx_l0_active = h->i_ref0 <= 0 ? 1 : h->i_ref0;
 901         h->sh.i_num_ref_idx_l1_active = h->i_ref1 <= 0 ? 1 : h->i_ref1;
 902     }
 903
 904     h->fdec->i_frame_num = h->sh.i_frame_num;
 905
 906     if( h->sps->i_poc_type == 0 )
 907     {
 908         h->sh.i_poc_lsb = h->fdec->i_poc & ( (1 << h->sps->i_log2_max_poc_lsb) - 1 );
 909         h->sh.i_delta_poc_bottom = 0;   /* XXX won't work for field */
 910     }
 911     else if( h->sps->i_poc_type == 1 )
 912     {
 913         /* FIXME TODO FIXME */
 914     }
 915     else
 916     {
 917         /* Nothing to do ? */
 918     }
 919
 920     x264_macroblock_slice_init( h );
 921 }
 922
 923 static int x264_slice_write( x264_t *h )
 924 {
 925     int i_skip;
 926     int mb_xy;
 927     int i;
 928
 929     /* init stats */
 930     memset( &h->stat.frame, 0, sizeof(h->stat.frame) );
 931
 932     /* Slice */
 933     x264_nal_start( h, h->i_nal_type, h->i_nal_ref_idc );
 934
 935     /* Slice header */
 936     x264_slice_header_write( &h->out.bs, &h->sh, h->i_nal_ref_idc );
 937     if( h->param.b_cabac )
 938     {
 939         /* alignment needed */
 940         bs_align_1( &h->out.bs );
 941
 942         /* init cabac */
 943         x264_cabac_context_init( &h->cabac, h->sh.i_type, h->sh.i_qp, h->sh.i_cabac_init_idc );
 944         x264_cabac_encode_init ( &h->cabac, &h->out.bs );
 945     }
 946     h->mb.i_last_qp = h->sh.i_qp;
 947     h->mb.i_last_dqp = 0;
 948
 949     for( mb_xy = h->sh.i_first_mb, i_skip = 0; mb_xy < h->sh.i_last_mb; mb_xy++ )
 950     {
 951         const int i_mb_y = mb_xy / h->sps->i_mb_width;
 952         const int i_mb_x = mb_xy % h->sps->i_mb_width;
 953
 954         int mb_spos = bs_pos(&h->out.bs);
 955
 956         /* load cache */
 957         x264_macroblock_cache_load( h, i_mb_x, i_mb_y );
 958
 959         /* analyse parameters
 960          * Slice I: choose I_4x4 or I_16x16 mode
 961          * Slice P: choose between using P mode or intra (4x4 or 16x16)
 962          * */
 963         TIMER_START( i_mtime_analyse );
 964         x264_macroblock_analyse( h );
 965         TIMER_STOP( i_mtime_analyse );
 966
 967         /* encode this macrobock -> be carefull it can change the mb type to P_SKIP if needed */
 968         TIMER_START( i_mtime_encode );
 969         x264_macroblock_encode( h );
 970         TIMER_STOP( i_mtime_encode );
 971
 972         TIMER_START( i_mtime_write );
 973         if( h->param.b_cabac )
 974         {
 975             if( mb_xy > h->sh.i_first_mb )
 976                 x264_cabac_encode_terminal( &h->cabac, 0 );
 977
 978             if( IS_SKIP( h->mb.i_type ) )
 979                 x264_cabac_mb_skip( h, 1 );
 980             else
 981             {
 982                 if( h->sh.i_type != SLICE_TYPE_I )
 983                     x264_cabac_mb_skip( h, 0 );
 984                 x264_macroblock_write_cabac( h, &h->cabac );
 985             }
 986         }
 987         else
 988         {
 989             if( IS_SKIP( h->mb.i_type ) )
 990                 i_skip++;
 991             else
 992             {
 993                 if( h->sh.i_type != SLICE_TYPE_I )
 994                 {
 995                     bs_write_ue( &h->out.bs, i_skip );  /* skip run */
 996                     i_skip = 0;
 997                 }
 998                 x264_macroblock_write_cavlc( h, &h->out.bs );
 999             }
1000         }
1001         TIMER_STOP( i_mtime_write );
1002
1003 #if VISUALIZE
1004         if( h->param.b_visualize )
1005             x264_visualize_mb( h );
1006 #endif
1007
1008         /* save cache */
1009         x264_macroblock_cache_save( h );
1010
1011         /* accumulate mb stats */
1012         h->stat.frame.i_mb_count[h->mb.i_type]++;
1013         if( !IS_SKIP(h->mb.i_type) && !IS_INTRA(h->mb.i_type) && !IS_DIRECT(h->mb.i_type) )
1014         {
1015             if( h->mb.i_partition != D_8x8 )
1016                 h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_partition ] ] += 4;
1017             else
1018                 for( i = 0; i < 4; i++ )
1019                     h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_sub_partition[i] ] ] ++;
1020             if( h->param.i_frame_reference > 1 )
1021             {
1022                 for( i = 0; i < 4; i++ )
1023                 {
1024                     int i_ref = h->mb.cache.ref[0][ x264_scan8[4*i] ];
1025                     if( i_ref >= 0 )
1026                         h->stat.frame.i_mb_count_ref[i_ref] ++;
1027                 }
1028             }
1029         }
1030         if( h->mb.i_cbp_luma && !IS_INTRA(h->mb.i_type) )
1031         {
1032             h->stat.frame.i_mb_count_8x8dct[0] ++;
1033             h->stat.frame.i_mb_count_8x8dct[1] += h->mb.b_transform_8x8;
1034         }
1035
1036         if( h->mb.b_variable_qp )
1037             x264_ratecontrol_mb(h, bs_pos(&h->out.bs) - mb_spos);
1038     }
1039
1040     if( h->param.b_cabac )
1041     {
1042         /* end of slice */
1043         x264_cabac_encode_terminal( &h->cabac, 1 );
1044     }
1045     else if( i_skip > 0 )
1046     {
1047         bs_write_ue( &h->out.bs, i_skip );  /* last skip run */
1048     }
1049
1050     if( h->param.b_cabac )
1051     {
1052         x264_cabac_encode_flush( &h->cabac );
1053
1054     }
1055     else
1056     {
1057         /* rbsp_slice_trailing_bits */
1058         bs_rbsp_trailing( &h->out.bs );
1059     }
1060
1061     x264_nal_end( h );
1062
1063     /* Compute misc bits */
1064     h->stat.frame.i_misc_bits = bs_pos( &h->out.bs )
1065                               + NALU_OVERHEAD * 8
1066                               - h->stat.frame.i_itex_bits
1067                               - h->stat.frame.i_ptex_bits
1068                               - h->stat.frame.i_hdr_bits;
1069
1070     return 0;
1071 }
1072
1073 static inline int x264_slices_write( x264_t *h )
1074 {
1075     int i_frame_size;
1076
1077 #if VISUALIZE
1078     if( h->param.b_visualize )
1079         x264_visualize_init( h );
1080 #endif
1081
1082     if( h->param.i_threads == 1 )
1083     {
1084         x264_slice_write( h );
1085         i_frame_size = h->out.nal[h->out.i_nal-1].i_payload;
1086     }
1087     else
1088     {
1089         int i_nal = h->out.i_nal;
1090         int i_bs_size = h->out.i_bitstream / h->param.i_threads;
1091         int i;
1092         /* duplicate contexts */
1093         for( i = 0; i < h->param.i_threads; i++ )
1094         {
1095             x264_t *t = h->thread[i];
1096             if( i > 0 )
1097             {
1098                 memcpy( t, h, sizeof(x264_t) );
1099                 t->out.p_bitstream += i*i_bs_size;
1100                 bs_init( &t->out.bs, t->out.p_bitstream, i_bs_size );
1101             }
1102             t->sh.i_first_mb = (i    * h->sps->i_mb_height / h->param.i_threads) * h->sps->i_mb_width;
1103             t->sh.i_last_mb = ((i+1) * h->sps->i_mb_height / h->param.i_threads) * h->sps->i_mb_width;
1104             t->out.i_nal = i_nal + i;
1105         }
1106
1107         /* dispatch */
1108 #if HAVE_PTHREAD
1109         {
1110             pthread_t handles[X264_SLICE_MAX];
1111             void *status;
1112             for( i = 0; i < h->param.i_threads; i++ )
1113                 pthread_create( &handles[i], NULL, (void*)x264_slice_write, (void*)h->thread[i] );
1114             for( i = 0; i < h->param.i_threads; i++ )
1115                 pthread_join( handles[i], &status );
1116         }
1117 #else
1118         for( i = 0; i < h->param.i_threads; i++ )
1119             x264_slice_write( h->thread[i] );
1120 #endif
1121
1122         /* merge contexts */
1123         i_frame_size = h->out.nal[i_nal].i_payload;
1124         for( i = 1; i < h->param.i_threads; i++ )
1125         {
1126             int j;
1127             x264_t *t = h->thread[i];
1128             h->out.nal[i_nal+i] = t->out.nal[i_nal+i];
1129             i_frame_size += t->out.nal[i_nal+i].i_payload;
1130             // all entries in stat.frame are ints
1131             for( j = 0; j < sizeof(h->stat.frame) / sizeof(int); j++ )
1132                 ((int*)&h->stat.frame)[j] += ((int*)&t->stat.frame)[j];
1133         }
1134         h->out.i_nal = i_nal + h->param.i_threads;
1135     }
1136
1137 #if VISUALIZE
1138     if( h->param.b_visualize )
1139     {
1140         x264_visualize_show( h );
1141         x264_visualize_close( h );
1142     }
1143 #endif
1144
1145     return i_frame_size;
1146 }
1147
1148 /****************************************************************************
1149  * x264_encoder_encode:
1150  *  XXX: i_poc   : is the poc of the current given picture
1151  *       i_frame : is the number of the frame being coded
1152  *  ex:  type frame poc
1153  *       I      0   2*0
1154  *       P      1   2*3
1155  *       B      2   2*1
1156  *       B      3   2*2
1157  *       P      4   2*6
1158  *       B      5   2*4
1159  *       B      6   2*5
1160  ****************************************************************************/
1161 int     x264_encoder_encode( x264_t *h,
1162                              x264_nal_t **pp_nal, int *pi_nal,
1163                              x264_picture_t *pic_in,
1164                              x264_picture_t *pic_out )
1165 {
1166     x264_frame_t   *frame_psnr = h->fdec; /* just to keep the current decoded frame for psnr calculation */
1167     int     i_nal_type;
1168     int     i_nal_ref_idc;
1169     int     i_slice_type;
1170     int     i_frame_size;
1171
1172     int i;
1173
1174     int   i_global_qp;
1175
1176     char psz_message[80];
1177
1178     /* no data out */
1179     *pi_nal = 0;
1180     *pp_nal = NULL;
1181
1182
1183     /* ------------------- Setup new frame from picture -------------------- */
1184     TIMER_START( i_mtime_encode_frame );
1185     if( pic_in != NULL )
1186     {
1187         /* 1: Copy the picture to a frame and move it to a buffer */
1188         x264_frame_t *fenc = x264_frame_get( h->frames.unused );
1189
1190         x264_frame_copy_picture( h, fenc, pic_in );
1191
1192         if( h->param.i_width % 16 || h->param.i_height % 16 )
1193             x264_frame_expand_border_mod16( h, fenc );
1194
1195         fenc->i_frame = h->frames.i_input++;
1196
1197         x264_frame_put( h->frames.next, fenc );
1198
1199         if( h->frames.b_have_lowres )
1200             x264_frame_init_lowres( h->param.cpu, fenc );
1201
1202         if( h->frames.i_input <= h->frames.i_delay )
1203         {
1204             /* Nothing yet to encode */
1205             /* waiting for filling bframe buffer */
1206             pic_out->i_type = X264_TYPE_AUTO;
1207             return 0;
1208         }
1209     }
1210
1211     if( h->frames.current[0] == NULL )
1212     {
1213         int bframes = 0;
1214         /* 2: Select frame types */
1215         if( h->frames.next[0] == NULL )
1216             return 0;
1217
1218         x264_slicetype_decide( h );
1219
1220         /* 3: move some B-frames and 1 non-B to encode queue */
1221         while( IS_X264_TYPE_B( h->frames.next[bframes]->i_type ) )
1222             bframes++;
1223         x264_frame_put( h->frames.current, x264_frame_get( &h->frames.next[bframes] ) );
1224         /* FIXME: when max B-frames > 3, BREF may no longer be centered after GOP closing */
1225         if( h->param.b_bframe_pyramid && bframes > 1 )
1226         {
1227             x264_frame_t *mid = x264_frame_get( &h->frames.next[bframes/2] );
1228             mid->i_type = X264_TYPE_BREF;
1229             x264_frame_put( h->frames.current, mid );
1230             bframes--;
1231         }
1232         while( bframes-- )
1233             x264_frame_put( h->frames.current, x264_frame_get( h->frames.next ) );
1234     }
1235     TIMER_STOP( i_mtime_encode_frame );
1236
1237     /* ------------------- Get frame to be encoded ------------------------- */
1238     /* 4: get picture to encode */
1239     h->fenc = x264_frame_get( h->frames.current );
1240     if( h->fenc == NULL )
1241     {
1242         /* Nothing yet to encode (ex: waiting for I/P with B frames) */
1243         /* waiting for filling bframe buffer */
1244         pic_out->i_type = X264_TYPE_AUTO;
1245         return 0;
1246     }
1247
1248 do_encode:
1249
1250     if( h->fenc->i_type == X264_TYPE_IDR )
1251     {
1252         h->frames.i_last_idr = h->fenc->i_frame;
1253     }
1254
1255     /* ------------------- Setup frame context ----------------------------- */
1256     /* 5: Init data dependant of frame type */
1257     TIMER_START( i_mtime_encode_frame );
1258     if( h->fenc->i_type == X264_TYPE_IDR )
1259     {
1260         /* reset ref pictures */
1261         x264_reference_reset( h );
1262
1263         i_nal_type    = NAL_SLICE_IDR;
1264         i_nal_ref_idc = NAL_PRIORITY_HIGHEST;
1265         i_slice_type = SLICE_TYPE_I;
1266     }
1267     else if( h->fenc->i_type == X264_TYPE_I )
1268     {
1269         i_nal_type    = NAL_SLICE;
1270         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
1271         i_slice_type = SLICE_TYPE_I;
1272     }
1273     else if( h->fenc->i_type == X264_TYPE_P )
1274     {
1275         i_nal_type    = NAL_SLICE;
1276         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
1277         i_slice_type = SLICE_TYPE_P;
1278     }
1279     else if( h->fenc->i_type == X264_TYPE_BREF )
1280     {
1281         i_nal_type    = NAL_SLICE;
1282         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* maybe add MMCO to forget it? -> low */
1283         i_slice_type = SLICE_TYPE_B;
1284     }
1285     else    /* B frame */
1286     {
1287         i_nal_type    = NAL_SLICE;
1288         i_nal_ref_idc = NAL_PRIORITY_DISPOSABLE;
1289         i_slice_type = SLICE_TYPE_B;
1290     }
1291
1292     h->fdec->i_poc =
1293     h->fenc->i_poc = 2 * (h->fenc->i_frame - h->frames.i_last_idr);
1294     h->fdec->i_type = h->fenc->i_type;
1295     h->fdec->i_frame = h->fenc->i_frame;
1296     h->fenc->b_kept_as_ref =
1297     h->fdec->b_kept_as_ref = i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE;
1298
1299
1300
1301     /* ------------------- Init                ----------------------------- */
1302     /* build ref list 0/1 */
1303     x264_reference_build_list( h, h->fdec->i_poc, i_slice_type );
1304
1305     /* Init the rate control */
1306     x264_ratecontrol_start( h, i_slice_type, h->fenc->i_qpplus1 );
1307     i_global_qp = x264_ratecontrol_qp( h );
1308
1309     pic_out->i_qpplus1 =
1310     h->fdec->i_qpplus1 = i_global_qp + 1;
1311
1312     if( i_slice_type == SLICE_TYPE_B )
1313         x264_macroblock_bipred_init( h );
1314
1315     if( h->fenc->b_kept_as_ref )
1316         h->i_frame_num++;
1317
1318     /* ------------------------ Create slice header  ----------------------- */
1319     x264_slice_init( h, i_nal_type, i_slice_type, i_global_qp );
1320
1321     /* ---------------------- Write the bitstream -------------------------- */
1322     /* Init bitstream context */
1323     h->out.i_nal = 0;
1324     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
1325
1326     if(h->param.b_aud){
1327         int pic_type;
1328
1329         if(i_slice_type == SLICE_TYPE_I)
1330             pic_type = 0;
1331         else if(i_slice_type == SLICE_TYPE_P)
1332             pic_type = 1;
1333         else if(i_slice_type == SLICE_TYPE_B)
1334             pic_type = 2;
1335         else
1336             pic_type = 7;
1337
1338         x264_nal_start(h, NAL_AUD, NAL_PRIORITY_DISPOSABLE);
1339         bs_write(&h->out.bs, 3, pic_type);
1340         bs_rbsp_trailing(&h->out.bs);
1341         x264_nal_end(h);
1342     }
1343
1344     h->i_nal_type = i_nal_type;
1345     h->i_nal_ref_idc = i_nal_ref_idc;
1346
1347     /* Write SPS and PPS */
1348     if( i_nal_type == NAL_SLICE_IDR && h->param.b_repeat_headers )
1349     {
1350         if( h->fenc->i_frame == 0 )
1351         {
1352             /* identify ourself */
1353             x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
1354             x264_sei_version_write( h, &h->out.bs );
1355             x264_nal_end( h );
1356         }
1357
1358         /* generate sequence parameters */
1359         x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
1360         x264_sps_write( &h->out.bs, h->sps );
1361         x264_nal_end( h );
1362
1363         /* generate picture parameters */
1364         x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
1365         x264_pps_write( &h->out.bs, h->pps );
1366         x264_nal_end( h );
1367     }
1368
1369     /* Write frame */
1370     i_frame_size = x264_slices_write( h );
1371
1372     /* restore CPU state (before using float again) */
1373     x264_cpu_restore( h->param.cpu );
1374
1375     if( i_slice_type == SLICE_TYPE_P && !h->param.rc.b_stat_read
1376         && h->param.i_scenecut_threshold >= 0 )
1377     {
1378         const int *mbs = h->stat.frame.i_mb_count;
1379         int i_mb_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1380         int i_mb_p = mbs[P_L0] + mbs[P_8x8];
1381         int i_mb_s = mbs[P_SKIP];
1382         int i_mb   = h->sps->i_mb_width * h->sps->i_mb_height;
1383         int64_t i_inter_cost = h->stat.frame.i_inter_cost;
1384         int64_t i_intra_cost = h->stat.frame.i_intra_cost;
1385
1386         float f_bias;
1387         int i_gop_size = h->fenc->i_frame - h->frames.i_last_idr;
1388         float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
1389         /* magic numbers pulled out of thin air */
1390         float f_thresh_min = f_thresh_max * h->param.i_keyint_min
1391                              / ( h->param.i_keyint_max * 4 );
1392         if( h->param.i_keyint_min == h->param.i_keyint_max )
1393              f_thresh_min= f_thresh_max;
1394
1395         /* macroblock_analyse() doesn't further analyse skipped mbs,
1396          * so we have to guess their cost */
1397         if( i_mb_s < i_mb )
1398             i_intra_cost = i_intra_cost * i_mb / (i_mb - i_mb_s);
1399
1400         if( i_gop_size < h->param.i_keyint_min / 4 )
1401             f_bias = f_thresh_min / 4;
1402         else if( i_gop_size <= h->param.i_keyint_min )
1403             f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
1404         else
1405         {
1406             f_bias = f_thresh_min
1407                      + ( f_thresh_max - f_thresh_min )
1408                        * ( i_gop_size - h->param.i_keyint_min )
1409                        / ( h->param.i_keyint_max - h->param.i_keyint_min );
1410         }
1411         f_bias = X264_MIN( f_bias, 1.0 );
1412
1413         /* Bad P will be reencoded as I */
1414         if( i_mb_s < i_mb &&
1415             i_inter_cost >= (1.0 - f_bias) * i_intra_cost )
1416         {
1417             int b;
1418
1419             x264_log( h, X264_LOG_DEBUG, "scene cut at %d Icost:%.0f Pcost:%.0f ratio:%.3f bias=%.3f lastIDR:%d (I:%d P:%d S:%d)\n",
1420                       h->fenc->i_frame,
1421                       (double)i_intra_cost, (double)i_inter_cost,
1422                       (double)i_inter_cost / i_intra_cost,
1423                       f_bias, i_gop_size,
1424                       i_mb_i, i_mb_p, i_mb_s );
1425
1426             /* Restore frame num */
1427             h->i_frame_num--;
1428
1429             for( b = 0; h->frames.current[b] && IS_X264_TYPE_B( h->frames.current[b]->i_type ); b++ );
1430             if( b > 0 )
1431             {
1432                 /* If using B-frames, force GOP to be closed.
1433                  * Even if this frame is going to be I and not IDR, forcing a
1434                  * P-frame before the scenecut will probably help compression.
1435                  *
1436                  * We don't yet know exactly which frame is the scene cut, so
1437                  * we can't assign an I-frame. Instead, change the previous
1438                  * B-frame to P, and rearrange coding order. */
1439
1440                 if( h->param.b_bframe_adaptive || b > 1 )
1441                     h->fenc->i_type = X264_TYPE_AUTO;
1442                 x264_frame_sort_pts( h->frames.current );
1443                 x264_frame_push( h->frames.next, h->fenc );
1444                 h->fenc = h->frames.current[b-1];
1445                 h->frames.current[b-1] = NULL;
1446                 h->fenc->i_type = X264_TYPE_P;
1447                 x264_frame_sort_dts( h->frames.current );
1448             }
1449             /* Do IDR if needed */
1450             else if( i_gop_size >= h->param.i_keyint_min )
1451             {
1452                 x264_frame_t *tmp;
1453
1454                 /* Reset */
1455                 h->i_frame_num = 0;
1456
1457                 /* Reinit field of fenc */
1458                 h->fenc->i_type = X264_TYPE_IDR;
1459                 h->fenc->i_poc = 0;
1460
1461                 /* Put enqueued frames back in the pool */
1462                 while( (tmp = x264_frame_get( h->frames.current ) ) != NULL )
1463                     x264_frame_put( h->frames.next, tmp );
1464                 x264_frame_sort_pts( h->frames.next );
1465             }
1466             else
1467             {
1468                 h->fenc->i_type = X264_TYPE_I;
1469             }
1470             goto do_encode;
1471         }
1472     }
1473
1474     /* End bitstream, set output  */
1475     *pi_nal = h->out.i_nal;
1476     *pp_nal = h->out.nal;
1477
1478     /* Set output picture properties */
1479     if( i_slice_type == SLICE_TYPE_I )
1480         pic_out->i_type = i_nal_type == NAL_SLICE_IDR ? X264_TYPE_IDR : X264_TYPE_I;
1481     else if( i_slice_type == SLICE_TYPE_P )
1482         pic_out->i_type = X264_TYPE_P;
1483     else
1484         pic_out->i_type = X264_TYPE_B;
1485     pic_out->i_pts = h->fenc->i_pts;
1486
1487     pic_out->img.i_plane = h->fdec->i_plane;
1488     for(i = 0; i < 4; i++){
1489         pic_out->img.i_stride[i] = h->fdec->i_stride[i];
1490         pic_out->img.plane[i] = h->fdec->plane[i];
1491     }
1492
1493     /* ---------------------- Update encoder state ------------------------- */
1494
1495     /* handle references */
1496     if( i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE )
1497     {
1498         x264_reference_update( h );
1499     }
1500
1501     /* increase frame count */
1502     h->i_frame++;
1503
1504     /* restore CPU state (before using float again) */
1505     /* XXX: not needed? (done above) */
1506     x264_cpu_restore( h->param.cpu );
1507
1508     /* update rc */
1509     x264_ratecontrol_end( h, i_frame_size * 8 );
1510
1511     x264_frame_put( h->frames.unused, h->fenc );
1512
1513     x264_noise_reduction_update( h );
1514
1515     TIMER_STOP( i_mtime_encode_frame );
1516
1517     /* ---------------------- Compute/Print statistics --------------------- */
1518     /* Slice stat */
1519     h->stat.i_slice_count[i_slice_type]++;
1520     h->stat.i_slice_size[i_slice_type] += i_frame_size + NALU_OVERHEAD;
1521     h->stat.i_slice_qp[i_slice_type] += i_global_qp;
1522
1523     for( i = 0; i < 19; i++ )
1524         h->stat.i_mb_count[h->sh.i_type][i] += h->stat.frame.i_mb_count[i];
1525     for( i = 0; i < 2; i++ )
1526         h->stat.i_mb_count_8x8dct[i] += h->stat.frame.i_mb_count_8x8dct[i];
1527     if( h->sh.i_type != SLICE_TYPE_I )
1528     {
1529         for( i = 0; i < 7; i++ )
1530             h->stat.i_mb_count_size[h->sh.i_type][i] += h->stat.frame.i_mb_count_size[i];
1531         for( i = 0; i < 16; i++ )
1532             h->stat.i_mb_count_ref[h->sh.i_type][i] += h->stat.frame.i_mb_count_ref[i];
1533     }
1534
1535     if( h->param.analyse.b_psnr )
1536     {
1537         int64_t i_sqe_y, i_sqe_u, i_sqe_v;
1538
1539         /* PSNR */
1540         i_sqe_y = x264_pixel_ssd_wxh( &h->pixf, frame_psnr->plane[0], frame_psnr->i_stride[0], h->fenc->plane[0], h->fenc->i_stride[0], h->param.i_width, h->param.i_height );
1541         i_sqe_u = x264_pixel_ssd_wxh( &h->pixf, frame_psnr->plane[1], frame_psnr->i_stride[1], h->fenc->plane[1], h->fenc->i_stride[1], h->param.i_width/2, h->param.i_height/2);
1542         i_sqe_v = x264_pixel_ssd_wxh( &h->pixf, frame_psnr->plane[2], frame_psnr->i_stride[2], h->fenc->plane[2], h->fenc->i_stride[2], h->param.i_width/2, h->param.i_height/2);
1543         x264_cpu_restore( h->param.cpu );
1544
1545         h->stat.i_sqe_global[i_slice_type] += i_sqe_y + i_sqe_u + i_sqe_v;
1546         h->stat.f_psnr_average[i_slice_type] += x264_psnr( i_sqe_y + i_sqe_u + i_sqe_v, 3 * h->param.i_width * h->param.i_height / 2 );
1547         h->stat.f_psnr_mean_y[i_slice_type] += x264_psnr( i_sqe_y, h->param.i_width * h->param.i_height );
1548         h->stat.f_psnr_mean_u[i_slice_type] += x264_psnr( i_sqe_u, h->param.i_width * h->param.i_height / 4 );
1549         h->stat.f_psnr_mean_v[i_slice_type] += x264_psnr( i_sqe_v, h->param.i_width * h->param.i_height / 4 );
1550
1551         snprintf( psz_message, 80, " PSNR Y:%2.2f U:%2.2f V:%2.2f",
1552                   x264_psnr( i_sqe_y, h->param.i_width * h->param.i_height ),
1553                   x264_psnr( i_sqe_u, h->param.i_width * h->param.i_height / 4),
1554                   x264_psnr( i_sqe_v, h->param.i_width * h->param.i_height / 4) );
1555         psz_message[79] = '\0';
1556     }
1557     else
1558     {
1559         psz_message[0] = '\0';
1560     }
1561
1562     x264_log( h, X264_LOG_DEBUG,
1563                   "frame=%4d QP=%i NAL=%d Slice:%c Poc:%-3d I:%-4d P:%-4d SKIP:%-4d size=%d bytes%s\n",
1564               h->i_frame - 1,
1565               i_global_qp,
1566               i_nal_ref_idc,
1567               i_slice_type == SLICE_TYPE_I ? 'I' : (i_slice_type == SLICE_TYPE_P ? 'P' : 'B' ),
1568               frame_psnr->i_poc,
1569               h->stat.frame.i_mb_count_i,
1570               h->stat.frame.i_mb_count_p,
1571               h->stat.frame.i_mb_count_skip,
1572               i_frame_size,
1573               psz_message );
1574
1575
1576 #ifdef DEBUG_MB_TYPE
1577 {
1578     static const char mb_chars[] = { 'i', 'i', 'I', 'C', 'P', '8', 'S',
1579         'D', '<', 'X', 'B', 'X', '>', 'B', 'B', 'B', 'B', '8', 'S' };
1580     int mb_xy;
1581     for( mb_xy = 0; mb_xy < h->sps->i_mb_width * h->sps->i_mb_height; mb_xy++ )
1582     {
1583         if( h->mb.type[mb_xy] < 19 && h->mb.type[mb_xy] >= 0 )
1584             fprintf( stderr, "%c ", mb_chars[ h->mb.type[mb_xy] ] );
1585         else
1586             fprintf( stderr, "? " );
1587
1588         if( (mb_xy+1) % h->sps->i_mb_width == 0 )
1589             fprintf( stderr, "\n" );
1590     }
1591 }
1592 #endif
1593
1594 #ifdef DEBUG_DUMP_FRAME
1595     /* Dump reconstructed frame */
1596     x264_frame_dump( h, frame_psnr, "fdec.yuv" );
1597 #endif
1598     return 0;
1599 }
1600
1601 /****************************************************************************
1602  * x264_encoder_close:
1603  ****************************************************************************/
1604 void    x264_encoder_close  ( x264_t *h )
1605 {
1606 #ifdef DEBUG_BENCHMARK
1607     int64_t i_mtime_total = i_mtime_analyse + i_mtime_encode + i_mtime_write + i_mtime_filter + 1;
1608 #endif
1609     int64_t i_yuv_size = 3 * h->param.i_width * h->param.i_height / 2;
1610     int i;
1611
1612 #ifdef DEBUG_BENCHMARK
1613     x264_log( h, X264_LOG_INFO,
1614               "analyse=%d(%lldms) encode=%d(%lldms) write=%d(%lldms) filter=%d(%lldms)\n",
1615               (int)(100*i_mtime_analyse/i_mtime_total), i_mtime_analyse/1000,
1616               (int)(100*i_mtime_encode/i_mtime_total), i_mtime_encode/1000,
1617               (int)(100*i_mtime_write/i_mtime_total), i_mtime_write/1000,
1618               (int)(100*i_mtime_filter/i_mtime_total), i_mtime_filter/1000 );
1619 #endif
1620
1621     /* Slices used and PSNR */
1622     for( i=0; i<5; i++ )
1623     {
1624         static const int slice_order[] = { SLICE_TYPE_I, SLICE_TYPE_SI, SLICE_TYPE_P, SLICE_TYPE_SP, SLICE_TYPE_B };
1625         static const char *slice_name[] = { "P", "B", "I", "SP", "SI" };
1626         int i_slice = slice_order[i];
1627
1628         if( h->stat.i_slice_count[i_slice] > 0 )
1629         {
1630             const int i_count = h->stat.i_slice_count[i_slice];
1631             if( h->param.analyse.b_psnr )
1632             {
1633                 x264_log( h, X264_LOG_INFO,
1634                           "slice %s:%-5d Avg QP:%5.2f  size:%6.0f  PSNR Mean Y:%5.2f U:%5.2f V:%5.2f Avg:%5.2f Global:%5.2f\n",
1635                           slice_name[i_slice],
1636                           i_count,
1637                           (double)h->stat.i_slice_qp[i_slice] / i_count,
1638                           (double)h->stat.i_slice_size[i_slice] / i_count,
1639                           h->stat.f_psnr_mean_y[i_slice] / i_count, h->stat.f_psnr_mean_u[i_slice] / i_count, h->stat.f_psnr_mean_v[i_slice] / i_count,
1640                           h->stat.f_psnr_average[i_slice] / i_count,
1641                           x264_psnr( h->stat.i_sqe_global[i_slice], i_count * i_yuv_size ) );
1642             }
1643             else
1644             {
1645                 x264_log( h, X264_LOG_INFO,
1646                           "slice %s:%-5d Avg QP:%5.2f  size:%6.0f\n",
1647                           slice_name[i_slice],
1648                           i_count,
1649                           (double)h->stat.i_slice_qp[i_slice] / i_count,
1650                           (double)h->stat.i_slice_size[i_slice] / i_count );
1651             }
1652         }
1653     }
1654
1655     /* MB types used */
1656     if( h->stat.i_slice_count[SLICE_TYPE_I] > 0 )
1657     {
1658         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_I];
1659         const double i_count = h->stat.i_slice_count[SLICE_TYPE_I] * h->mb.i_mb_count / 100.0;
1660         x264_log( h, X264_LOG_INFO,
1661                   "mb I  I16..4: %4.1f%% %4.1f%% %4.1f%%\n",
1662                   i_mb_count[I_16x16]/ i_count,
1663                   i_mb_count[I_8x8]  / i_count,
1664                   i_mb_count[I_4x4]  / i_count );
1665     }
1666     if( h->stat.i_slice_count[SLICE_TYPE_P] > 0 )
1667     {
1668         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_P];
1669         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_P];
1670         const double i_count = h->stat.i_slice_count[SLICE_TYPE_P] * h->mb.i_mb_count / 100.0;
1671         x264_log( h, X264_LOG_INFO,
1672                   "mb P  I16..4: %4.1f%% %4.1f%% %4.1f%%  P16..4: %4.1f%% %4.1f%% %4.1f%% %4.1f%% %4.1f%%    skip:%4.1f%%\n",
1673                   i_mb_count[I_16x16]/ i_count,
1674                   i_mb_count[I_8x8]  / i_count,
1675                   i_mb_count[I_4x4]  / i_count,
1676                   i_mb_size[PIXEL_16x16] / (i_count*4),
1677                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
1678                   i_mb_size[PIXEL_8x8] / (i_count*4),
1679                   (i_mb_size[PIXEL_8x4] + i_mb_size[PIXEL_4x8]) / (i_count*4),
1680                   i_mb_size[PIXEL_4x4] / (i_count*4),
1681                   i_mb_count[P_SKIP] / i_count );
1682     }
1683     if( h->stat.i_slice_count[SLICE_TYPE_B] > 0 )
1684     {
1685         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_B];
1686         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_B];
1687         const double i_count = h->stat.i_slice_count[SLICE_TYPE_B] * h->mb.i_mb_count / 100.0;
1688         x264_log( h, X264_LOG_INFO,
1689                   "mb B  I16..4: %4.1f%% %4.1f%% %4.1f%%  B16..8: %4.1f%% %4.1f%% %4.1f%%  direct:%4.1f%%  skip:%4.1f%%\n",
1690                   i_mb_count[I_16x16]  / i_count,
1691                   i_mb_count[I_8x8]    / i_count,
1692                   i_mb_count[I_4x4]    / i_count,
1693                   i_mb_size[PIXEL_16x16] / (i_count*4),
1694                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
1695                   i_mb_size[PIXEL_8x8] / (i_count*4),
1696                   i_mb_count[B_DIRECT] / i_count,
1697                   i_mb_count[B_SKIP]   / i_count );
1698     }
1699
1700     x264_ratecontrol_summary( h );
1701
1702     if( h->stat.i_slice_count[SLICE_TYPE_I] + h->stat.i_slice_count[SLICE_TYPE_P] + h->stat.i_slice_count[SLICE_TYPE_B] > 0 )
1703     {
1704         const int i_count = h->stat.i_slice_count[SLICE_TYPE_I] +
1705                             h->stat.i_slice_count[SLICE_TYPE_P] +
1706                             h->stat.i_slice_count[SLICE_TYPE_B];
1707         float fps = (float) h->param.i_fps_num / h->param.i_fps_den;
1708 #define SUM3(p) (p[SLICE_TYPE_I] + p[SLICE_TYPE_P] + p[SLICE_TYPE_B])
1709 #define SUM3b(p,o) (p[SLICE_TYPE_I][o] + p[SLICE_TYPE_P][o] + p[SLICE_TYPE_B][o])
1710         float f_bitrate = fps * SUM3(h->stat.i_slice_size) / i_count / 125;
1711
1712         if( h->param.analyse.b_transform_8x8 )
1713         {
1714             int64_t i_i8x8 = SUM3b( h->stat.i_mb_count, I_8x8 );
1715             int64_t i_intra = i_i8x8 + SUM3b( h->stat.i_mb_count, I_4x4 )
1716                                      + SUM3b( h->stat.i_mb_count, I_16x16 );
1717             x264_log( h, X264_LOG_INFO, "8x8 transform  intra:%.1f%%  inter:%.1f%%\n",
1718                       100. * i_i8x8 / i_intra,
1719                       100. * h->stat.i_mb_count_8x8dct[1] / h->stat.i_mb_count_8x8dct[0] );
1720         }
1721
1722         if( h->param.i_frame_reference > 1 )
1723         {
1724             int i_slice;
1725             for( i_slice = 0; i_slice < 2; i_slice++ )
1726             {
1727                 char buf[200];
1728                 char *p = buf;
1729                 int64_t i_den = 0;
1730                 int i_max = 0;
1731                 for( i = 0; i < h->param.i_frame_reference; i++ )
1732                     if( h->stat.i_mb_count_ref[i_slice][i] )
1733                     {
1734                         i_den += h->stat.i_mb_count_ref[i_slice][i];
1735                         i_max = i;
1736                     }
1737                 if( i_max == 0 )
1738                     continue;
1739                 for( i = 0; i <= i_max; i++ )
1740                     p += sprintf( p, " %4.1f%%", 100. * h->stat.i_mb_count_ref[i_slice][i] / i_den );
1741                 x264_log( h, X264_LOG_INFO, "ref %c %s\n", i_slice==SLICE_TYPE_P ? 'P' : 'B', buf );
1742             }
1743         }
1744
1745         if( h->param.analyse.b_psnr )
1746             x264_log( h, X264_LOG_INFO,
1747                       "PSNR Mean Y:%6.3f U:%6.3f V:%6.3f Avg:%6.3f Global:%6.3f kb/s:%.2f\n",
1748                       SUM3( h->stat.f_psnr_mean_y ) / i_count,
1749                       SUM3( h->stat.f_psnr_mean_u ) / i_count,
1750                       SUM3( h->stat.f_psnr_mean_v ) / i_count,
1751                       SUM3( h->stat.f_psnr_average ) / i_count,
1752                       x264_psnr( SUM3( h->stat.i_sqe_global ), i_count * i_yuv_size ),
1753                       f_bitrate );
1754         else
1755             x264_log( h, X264_LOG_INFO, "kb/s:%.1f\n", f_bitrate );
1756     }
1757
1758     /* frames */
1759     for( i = 0; i < X264_BFRAME_MAX + 3; i++ )
1760     {
1761         if( h->frames.current[i] ) x264_frame_delete( h->frames.current[i] );
1762         if( h->frames.next[i] )    x264_frame_delete( h->frames.next[i] );
1763         if( h->frames.unused[i] )  x264_frame_delete( h->frames.unused[i] );
1764     }
1765     /* ref frames */
1766     for( i = 0; i < h->frames.i_max_dpb; i++ )
1767     {
1768         x264_frame_delete( h->frames.reference[i] );
1769     }
1770
1771     /* rc */
1772     x264_ratecontrol_delete( h );
1773
1774     /* param */
1775     if( h->param.rc.psz_stat_out )
1776         free( h->param.rc.psz_stat_out );
1777     if( h->param.rc.psz_stat_in )
1778         free( h->param.rc.psz_stat_in );
1779     if( h->param.rc.psz_rc_eq )
1780         free( h->param.rc.psz_rc_eq );
1781
1782     x264_macroblock_cache_end( h );
1783     x264_free( h->out.p_bitstream );
1784     for( i = 1; i < h->param.i_threads; i++ )
1785         x264_free( h->thread[i] );
1786     x264_free( h );
1787 }
1788