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 <stdio.h>
  25 #include <string.h>
  26 #include <math.h>
  27
  28 #include "common/common.h"
  29 #include "common/cpu.h"
  30
  31 #include "set.h"
  32 #include "analyse.h"
  33 #include "ratecontrol.h"
  34 #include "macroblock.h"
  35
  36 #if VISUALIZE
  37 #include "common/visualize.h"
  38 #endif
  39
  40 //#define DEBUG_MB_TYPE
  41 //#define DEBUG_DUMP_FRAME
  42 //#define DEBUG_BENCHMARK
  43
  44 #ifdef DEBUG_BENCHMARK
  45 static int64_t i_mtime_encode_frame = 0;
  46 static int64_t i_mtime_analyse = 0;
  47 static int64_t i_mtime_encode = 0;
  48 static int64_t i_mtime_write = 0;
  49 static int64_t i_mtime_filter = 0;
  50 #define TIMER_START( d ) \
  51     { \
  52         int64_t d##start = x264_mdate();
  53
  54 #define TIMER_STOP( d ) \
  55         d += x264_mdate() - d##start;\
  56     }
  57 #else
  58 #define TIMER_START( d )
  59 #define TIMER_STOP( d )
  60 #endif
  61
  62 #define NALU_OVERHEAD 5 // startcode + NAL type costs 5 bytes per frame
  63
  64 /****************************************************************************
  65  *
  66  ******************************* x264 libs **********************************
  67  *
  68  ****************************************************************************/
  69 static float x264_psnr( int64_t i_sqe, int64_t i_size )
  70 {
  71     double f_mse = (double)i_sqe / ((double)65025.0 * (double)i_size);
  72     if( f_mse <= 0.0000000001 ) /* Max 100dB */
  73         return 100;
  74
  75     return (float)(-10.0 * log( f_mse ) / log( 10.0 ));
  76 }
  77
  78 #ifdef DEBUG_DUMP_FRAME
  79 static void x264_frame_dump( x264_t *h, x264_frame_t *fr, char *name )
  80 {
  81     FILE *f = fopen( name, "r+b" );
  82     int i, y;
  83     if( !f )
  84         return;
  85
  86     /* Write the frame in display order */
  87     fseek( f, fr->i_frame * h->param.i_height * h->param.i_width * 3 / 2, SEEK_SET );
  88
  89     for( i = 0; i < fr->i_plane; i++ )
  90     {
  91         for( y = 0; y < h->param.i_height / ( i == 0 ? 1 : 2 ); y++ )
  92         {
  93             fwrite( &fr->plane[i][y*fr->i_stride[i]], 1, h->param.i_width / ( i == 0 ? 1 : 2 ), f );
  94         }
  95     }
  96     fclose( f );
  97 }
  98 #endif
  99
 100
 101 /* Fill "default" values */
 102 static void x264_slice_header_init( x264_t *h, x264_slice_header_t *sh,
 103                                     x264_sps_t *sps, x264_pps_t *pps,
 104                                     int i_type, int i_idr_pic_id, int i_frame, int i_qp )
 105 {
 106     x264_param_t *param = &h->param;
 107     int i;
 108
 109     /* First we fill all field */
 110     sh->sps = sps;
 111     sh->pps = pps;
 112
 113     sh->i_type      = i_type;
 114     sh->i_first_mb  = 0;
 115     sh->i_last_mb   = h->sps->i_mb_width * h->sps->i_mb_height;
 116     sh->i_pps_id    = pps->i_id;
 117
 118     sh->i_frame_num = i_frame;
 119
 120     sh->b_field_pic = 0;    /* Not field support for now */
 121     sh->b_bottom_field = 1; /* not yet used */
 122
 123     sh->i_idr_pic_id = i_idr_pic_id;
 124
 125     /* poc stuff, fixed later */
 126     sh->i_poc_lsb = 0;
 127     sh->i_delta_poc_bottom = 0;
 128     sh->i_delta_poc[0] = 0;
 129     sh->i_delta_poc[1] = 0;
 130
 131     sh->i_redundant_pic_cnt = 0;
 132
 133     if( !h->mb.b_direct_auto_read )
 134     {
 135         if( h->mb.b_direct_auto_write )
 136             sh->b_direct_spatial_mv_pred = ( h->stat.i_direct_score[1] > h->stat.i_direct_score[0] );
 137         else
 138             sh->b_direct_spatial_mv_pred = ( param->analyse.i_direct_mv_pred == X264_DIRECT_PRED_SPATIAL );
 139     }
 140     /* else b_direct_spatial_mv_pred was read from the 2pass statsfile */
 141
 142     sh->b_num_ref_idx_override = 0;
 143     sh->i_num_ref_idx_l0_active = 1;
 144     sh->i_num_ref_idx_l1_active = 1;
 145
 146     sh->b_ref_pic_list_reordering_l0 = h->b_ref_reorder[0];
 147     sh->b_ref_pic_list_reordering_l1 = h->b_ref_reorder[1];
 148
 149     /* If the ref list isn't in the default order, construct reordering header */
 150     /* List1 reordering isn't needed yet */
 151     if( sh->b_ref_pic_list_reordering_l0 )
 152     {
 153         int pred_frame_num = i_frame;
 154         for( i = 0; i < h->i_ref0; i++ )
 155         {
 156             int diff = h->fref0[i]->i_frame_num - pred_frame_num;
 157             if( diff == 0 )
 158                 x264_log( h, X264_LOG_ERROR, "diff frame num == 0\n" );
 159             sh->ref_pic_list_order[0][i].idc = ( diff > 0 );
 160             sh->ref_pic_list_order[0][i].arg = abs( diff ) - 1;
 161             pred_frame_num = h->fref0[i]->i_frame_num;
 162         }
 163     }
 164
 165     sh->i_cabac_init_idc = param->i_cabac_init_idc;
 166
 167     sh->i_qp = i_qp;
 168     sh->i_qp_delta = i_qp - pps->i_pic_init_qp;
 169     sh->b_sp_for_swidth = 0;
 170     sh->i_qs_delta = 0;
 171
 172     /* If effective qp <= 15, deblocking would have no effect anyway */
 173     if( param->b_deblocking_filter
 174         && ( h->mb.b_variable_qp
 175         || 15 < i_qp + 2 * X264_MAX(param->i_deblocking_filter_alphac0, param->i_deblocking_filter_beta) ) )
 176     {
 177         sh->i_disable_deblocking_filter_idc = 0;
 178     }
 179     else
 180     {
 181         sh->i_disable_deblocking_filter_idc = 1;
 182     }
 183     sh->i_alpha_c0_offset = param->i_deblocking_filter_alphac0 << 1;
 184     sh->i_beta_offset = param->i_deblocking_filter_beta << 1;
 185 }
 186
 187 static void x264_slice_header_write( bs_t *s, x264_slice_header_t *sh, int i_nal_ref_idc )
 188 {
 189     int i;
 190
 191     bs_write_ue( s, sh->i_first_mb );
 192     bs_write_ue( s, sh->i_type + 5 );   /* same type things */
 193     bs_write_ue( s, sh->i_pps_id );
 194     bs_write( s, sh->sps->i_log2_max_frame_num, sh->i_frame_num );
 195
 196     if( sh->i_idr_pic_id >= 0 ) /* NAL IDR */
 197     {
 198         bs_write_ue( s, sh->i_idr_pic_id );
 199     }
 200
 201     if( sh->sps->i_poc_type == 0 )
 202     {
 203         bs_write( s, sh->sps->i_log2_max_poc_lsb, sh->i_poc_lsb );
 204         if( sh->pps->b_pic_order && !sh->b_field_pic )
 205         {
 206             bs_write_se( s, sh->i_delta_poc_bottom );
 207         }
 208     }
 209     else if( sh->sps->i_poc_type == 1 && !sh->sps->b_delta_pic_order_always_zero )
 210     {
 211         bs_write_se( s, sh->i_delta_poc[0] );
 212         if( sh->pps->b_pic_order && !sh->b_field_pic )
 213         {
 214             bs_write_se( s, sh->i_delta_poc[1] );
 215         }
 216     }
 217
 218     if( sh->pps->b_redundant_pic_cnt )
 219     {
 220         bs_write_ue( s, sh->i_redundant_pic_cnt );
 221     }
 222
 223     if( sh->i_type == SLICE_TYPE_B )
 224     {
 225         bs_write1( s, sh->b_direct_spatial_mv_pred );
 226     }
 227     if( sh->i_type == SLICE_TYPE_P || sh->i_type == SLICE_TYPE_SP || sh->i_type == SLICE_TYPE_B )
 228     {
 229         bs_write1( s, sh->b_num_ref_idx_override );
 230         if( sh->b_num_ref_idx_override )
 231         {
 232             bs_write_ue( s, sh->i_num_ref_idx_l0_active - 1 );
 233             if( sh->i_type == SLICE_TYPE_B )
 234             {
 235                 bs_write_ue( s, sh->i_num_ref_idx_l1_active - 1 );
 236             }
 237         }
 238     }
 239
 240     /* ref pic list reordering */
 241     if( sh->i_type != SLICE_TYPE_I )
 242     {
 243         bs_write1( s, sh->b_ref_pic_list_reordering_l0 );
 244         if( sh->b_ref_pic_list_reordering_l0 )
 245         {
 246             for( i = 0; i < sh->i_num_ref_idx_l0_active; i++ )
 247             {
 248                 bs_write_ue( s, sh->ref_pic_list_order[0][i].idc );
 249                 bs_write_ue( s, sh->ref_pic_list_order[0][i].arg );
 250
 251             }
 252             bs_write_ue( s, 3 );
 253         }
 254     }
 255     if( sh->i_type == SLICE_TYPE_B )
 256     {
 257         bs_write1( s, sh->b_ref_pic_list_reordering_l1 );
 258         if( sh->b_ref_pic_list_reordering_l1 )
 259         {
 260             for( i = 0; i < sh->i_num_ref_idx_l1_active; i++ )
 261             {
 262                 bs_write_ue( s, sh->ref_pic_list_order[1][i].idc );
 263                 bs_write_ue( s, sh->ref_pic_list_order[1][i].arg );
 264             }
 265             bs_write_ue( s, 3 );
 266         }
 267     }
 268
 269     if( ( sh->pps->b_weighted_pred && ( sh->i_type == SLICE_TYPE_P || sh->i_type == SLICE_TYPE_SP ) ) ||
 270         ( sh->pps->b_weighted_bipred == 1 && sh->i_type == SLICE_TYPE_B ) )
 271     {
 272         /* FIXME */
 273     }
 274
 275     if( i_nal_ref_idc != 0 )
 276     {
 277         if( sh->i_idr_pic_id >= 0 )
 278         {
 279             bs_write1( s, 0 );  /* no output of prior pics flag */
 280             bs_write1( s, 0 );  /* long term reference flag */
 281         }
 282         else
 283         {
 284             bs_write1( s, 0 );  /* adaptive_ref_pic_marking_mode_flag */
 285         }
 286     }
 287
 288     if( sh->pps->b_cabac && sh->i_type != SLICE_TYPE_I )
 289     {
 290         bs_write_ue( s, sh->i_cabac_init_idc );
 291     }
 292     bs_write_se( s, sh->i_qp_delta );      /* slice qp delta */
 293
 294     if( sh->pps->b_deblocking_filter_control )
 295     {
 296         bs_write_ue( s, sh->i_disable_deblocking_filter_idc );
 297         if( sh->i_disable_deblocking_filter_idc != 1 )
 298         {
 299             bs_write_se( s, sh->i_alpha_c0_offset >> 1 );
 300             bs_write_se( s, sh->i_beta_offset >> 1 );
 301         }
 302     }
 303 }
 304
 305 /****************************************************************************
 306  *
 307  ****************************************************************************
 308  ****************************** External API*********************************
 309  ****************************************************************************
 310  *
 311  ****************************************************************************/
 312
 313 static int x264_validate_parameters( x264_t *h )
 314 {
 315     if( h->param.i_width <= 0 || h->param.i_height <= 0 )
 316     {
 317         x264_log( h, X264_LOG_ERROR, "invalid width x height (%dx%d)\n",
 318                   h->param.i_width, h->param.i_height );
 319         return -1;
 320     }
 321
 322     if( h->param.i_width % 2 || h->param.i_height % 2 )
 323     {
 324         x264_log( h, X264_LOG_ERROR, "width or height not divisible by 2 (%dx%d)\n",
 325                   h->param.i_width, h->param.i_height );
 326         return -1;
 327     }
 328     if( h->param.i_csp != X264_CSP_I420 )
 329     {
 330         x264_log( h, X264_LOG_ERROR, "invalid CSP (only I420 supported)\n" );
 331         return -1;
 332     }
 333
 334     h->param.i_threads = x264_clip3( h->param.i_threads, 1, X264_SLICE_MAX );
 335     h->param.i_threads = X264_MIN( h->param.i_threads, (h->param.i_height + 15) / 16 );
 336 #ifndef HAVE_PTHREAD
 337     if( h->param.i_threads > 1 )
 338     {
 339         x264_log( h, X264_LOG_WARNING, "not compiled with pthread support!\n");
 340         x264_log( h, X264_LOG_WARNING, "multislicing anyway, but you won't see any speed gain.\n" );
 341     }
 342 #endif
 343
 344     if( h->param.rc.b_cbr )
 345         h->param.rc.i_rf_constant = 0;
 346     if( h->param.rc.i_rf_constant > 0 )
 347         h->param.rc.i_qp_constant = h->param.rc.i_rf_constant;
 348     h->param.rc.i_rf_constant = x264_clip3( h->param.rc.i_rf_constant, 0, 51 );
 349     h->param.rc.i_qp_constant = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
 350     if( !h->param.rc.b_cbr && h->param.rc.i_qp_constant == 0 )
 351     {
 352         h->mb.b_lossless = 1;
 353         h->param.analyse.b_transform_8x8 = 0;
 354         h->param.i_cqm_preset = X264_CQM_FLAT;
 355         h->param.psz_cqm_file = NULL;
 356         h->param.rc.f_ip_factor = 1;
 357         h->param.rc.f_pb_factor = 1;
 358         h->param.analyse.b_psnr = 0;
 359         h->param.analyse.i_chroma_qp_offset = 0;
 360         h->param.analyse.i_trellis = 0;
 361         h->param.analyse.b_fast_pskip = 0;
 362         h->param.analyse.i_noise_reduction = 0;
 363         h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 1, 6 );
 364     }
 365
 366     if( ( h->param.i_width % 16 || h->param.i_height % 16 ) && !h->mb.b_lossless )
 367     {
 368         x264_log( h, X264_LOG_WARNING,
 369                   "width or height not divisible by 16 (%dx%d), compression will suffer.\n",
 370                   h->param.i_width, h->param.i_height );
 371     }
 372
 373     h->param.i_frame_reference = x264_clip3( h->param.i_frame_reference, 1, 16 );
 374     if( h->param.i_keyint_max <= 0 )
 375         h->param.i_keyint_max = 1;
 376     h->param.i_keyint_min = x264_clip3( h->param.i_keyint_min, 1, h->param.i_keyint_max/2+1 );
 377
 378     h->param.i_bframe = x264_clip3( h->param.i_bframe, 0, X264_BFRAME_MAX );
 379     h->param.i_bframe_bias = x264_clip3( h->param.i_bframe_bias, -90, 100 );
 380     h->param.b_bframe_pyramid = h->param.b_bframe_pyramid && h->param.i_bframe > 1;
 381     h->param.b_bframe_adaptive = h->param.b_bframe_adaptive && h->param.i_bframe > 0;
 382     h->param.analyse.b_weighted_bipred = h->param.analyse.b_weighted_bipred && h->param.i_bframe > 0;
 383     h->mb.b_direct_auto_write = h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
 384                                 && h->param.i_bframe
 385                                 && ( h->param.rc.b_stat_write || !h->param.rc.b_stat_read );
 386
 387     h->param.i_deblocking_filter_alphac0 = x264_clip3( h->param.i_deblocking_filter_alphac0, -6, 6 );
 388     h->param.i_deblocking_filter_beta    = x264_clip3( h->param.i_deblocking_filter_beta, -6, 6 );
 389
 390     h->param.i_cabac_init_idc = x264_clip3( h->param.i_cabac_init_idc, 0, 2 );
 391
 392     if( h->param.i_cqm_preset < X264_CQM_FLAT || h->param.i_cqm_preset > X264_CQM_CUSTOM )
 393         h->param.i_cqm_preset = X264_CQM_FLAT;
 394
 395     if( h->param.analyse.i_me_method < X264_ME_DIA ||
 396         h->param.analyse.i_me_method > X264_ME_ESA )
 397         h->param.analyse.i_me_method = X264_ME_HEX;
 398     if( h->param.analyse.i_me_range < 4 )
 399         h->param.analyse.i_me_range = 4;
 400     if( h->param.analyse.i_me_range > 16 && h->param.analyse.i_me_method <= X264_ME_HEX )
 401         h->param.analyse.i_me_range = 16;
 402     h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 1, 7 );
 403     h->param.analyse.b_bframe_rdo = h->param.analyse.b_bframe_rdo && h->param.analyse.i_subpel_refine >= 6;
 404     h->param.analyse.b_mixed_references = h->param.analyse.b_mixed_references && h->param.i_frame_reference > 1;
 405     h->param.analyse.inter &= X264_ANALYSE_PSUB16x16|X264_ANALYSE_PSUB8x8|X264_ANALYSE_BSUB16x16|
 406                               X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 407     h->param.analyse.intra &= X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 408     if( !(h->param.analyse.inter & X264_ANALYSE_PSUB16x16) )
 409         h->param.analyse.inter &= ~X264_ANALYSE_PSUB8x8;
 410     if( !h->param.analyse.b_transform_8x8 )
 411     {
 412         h->param.analyse.inter &= ~X264_ANALYSE_I8x8;
 413         h->param.analyse.intra &= ~X264_ANALYSE_I8x8;
 414     }
 415     h->param.analyse.i_chroma_qp_offset = x264_clip3(h->param.analyse.i_chroma_qp_offset, -12, 12);
 416     if( !h->param.b_cabac )
 417         h->param.analyse.i_trellis = 0;
 418     h->param.analyse.i_trellis = x264_clip3( h->param.analyse.i_trellis, 0, 2 );
 419     h->param.analyse.i_noise_reduction = x264_clip3( h->param.analyse.i_noise_reduction, 0, 1<<16 );
 420
 421     {
 422         const x264_level_t *l = x264_levels;
 423         while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
 424             l++;
 425         if( l->level_idc == 0 )
 426         {
 427             x264_log( h, X264_LOG_ERROR, "invalid level_idc: %d\n", h->param.i_level_idc );
 428             return -1;
 429         }
 430         if( h->param.analyse.i_mv_range <= 0 )
 431             h->param.analyse.i_mv_range = l->mv_range;
 432         else
 433             h->param.analyse.i_mv_range = x264_clip3(h->param.analyse.i_mv_range, 32, 2048);
 434     }
 435
 436     if( h->param.rc.f_qblur < 0 )
 437         h->param.rc.f_qblur = 0;
 438     if( h->param.rc.f_complexity_blur < 0 )
 439         h->param.rc.f_complexity_blur = 0;
 440
 441     h->param.i_sps_id &= 31;
 442
 443     /* ensure the booleans are 0 or 1 so they can be used in math */
 444 #define BOOLIFY(x) h->param.x = !!h->param.x
 445     BOOLIFY( b_cabac );
 446     BOOLIFY( b_deblocking_filter );
 447     BOOLIFY( analyse.b_transform_8x8 );
 448     BOOLIFY( analyse.b_bidir_me );
 449     BOOLIFY( analyse.b_chroma_me );
 450     BOOLIFY( analyse.b_fast_pskip );
 451     BOOLIFY( rc.b_cbr );
 452     BOOLIFY( rc.b_stat_write );
 453     BOOLIFY( rc.b_stat_read );
 454 #undef BOOLIFY
 455
 456     return 0;
 457 }
 458
 459 /****************************************************************************
 460  * x264_encoder_open:
 461  ****************************************************************************/
 462 x264_t *x264_encoder_open   ( x264_param_t *param )
 463 {
 464     x264_t *h = x264_malloc( sizeof( x264_t ) );
 465     int i;
 466
 467     memset( h, 0, sizeof( x264_t ) );
 468
 469     /* Create a copy of param */
 470     memcpy( &h->param, param, sizeof( x264_param_t ) );
 471
 472     if( x264_validate_parameters( h ) < 0 )
 473     {
 474         x264_free( h );
 475         return NULL;
 476     }
 477
 478     if( h->param.psz_cqm_file )
 479         if( x264_cqm_parse_file( h, h->param.psz_cqm_file ) < 0 )
 480         {
 481             x264_free( h );
 482             return NULL;
 483         }
 484
 485     if( h->param.rc.psz_stat_out )
 486         h->param.rc.psz_stat_out = strdup( h->param.rc.psz_stat_out );
 487     if( h->param.rc.psz_stat_in )
 488         h->param.rc.psz_stat_in = strdup( h->param.rc.psz_stat_in );
 489     if( h->param.rc.psz_rc_eq )
 490         h->param.rc.psz_rc_eq = strdup( h->param.rc.psz_rc_eq );
 491
 492     /* VUI */
 493     if( h->param.vui.i_sar_width > 0 && h->param.vui.i_sar_height > 0 )
 494     {
 495         int i_w = param->vui.i_sar_width;
 496         int i_h = param->vui.i_sar_height;
 497
 498         x264_reduce_fraction( &i_w, &i_h );
 499
 500         while( i_w > 65535 || i_h > 65535 )
 501         {
 502             i_w /= 2;
 503             i_h /= 2;
 504         }
 505
 506         h->param.vui.i_sar_width = 0;
 507         h->param.vui.i_sar_height = 0;
 508         if( i_w == 0 || i_h == 0 )
 509         {
 510             x264_log( h, X264_LOG_WARNING, "cannot create valid sample aspect ratio\n" );
 511         }
 512         else
 513         {
 514             x264_log( h, X264_LOG_INFO, "using SAR=%d/%d\n", i_w, i_h );
 515             h->param.vui.i_sar_width = i_w;
 516             h->param.vui.i_sar_height = i_h;
 517         }
 518     }
 519
 520     x264_reduce_fraction( &h->param.i_fps_num, &h->param.i_fps_den );
 521
 522     /* Init x264_t */
 523     h->out.i_nal = 0;
 524     h->out.i_bitstream = X264_MAX( 1000000, h->param.i_width * h->param.i_height * 1.7
 525         * ( h->param.rc.b_cbr ? pow( 0.5, h->param.rc.i_qp_min )
 526           : pow( 0.5, h->param.rc.i_qp_constant ) * X264_MAX( 1, h->param.rc.f_ip_factor )));
 527     h->out.p_bitstream = x264_malloc( h->out.i_bitstream );
 528
 529     h->i_frame = 0;
 530     h->i_frame_num = 0;
 531     h->i_idr_pic_id = 0;
 532
 533     h->sps = &h->sps_array[0];
 534     x264_sps_init( h->sps, h->param.i_sps_id, &h->param );
 535
 536     h->pps = &h->pps_array[0];
 537     x264_pps_init( h->pps, h->param.i_sps_id, &h->param, h->sps);
 538
 539     x264_validate_levels( h );
 540
 541     x264_cqm_init( h );
 542
 543     h->mb.i_mb_count = h->sps->i_mb_width * h->sps->i_mb_height;
 544
 545     /* Init frames. */
 546     h->frames.i_delay = h->param.i_bframe;
 547     h->frames.i_max_ref0 = h->param.i_frame_reference;
 548     h->frames.i_max_ref1 = h->sps->vui.i_num_reorder_frames;
 549     h->frames.i_max_dpb  = h->sps->vui.i_max_dec_frame_buffering + 1;
 550     h->frames.b_have_lowres = !h->param.rc.b_stat_read
 551         && ( h->param.rc.b_cbr || h->param.rc.i_rf_constant || h->param.b_bframe_adaptive );
 552
 553     for( i = 0; i < X264_BFRAME_MAX + 3; i++ )
 554     {
 555         h->frames.current[i] = NULL;
 556         h->frames.next[i]    = NULL;
 557         h->frames.unused[i]  = NULL;
 558     }
 559     for( i = 0; i < 1 + h->frames.i_delay; i++ )
 560     {
 561         h->frames.unused[i] =  x264_frame_new( h );
 562         if( !h->frames.unused[i] )
 563             return NULL;
 564     }
 565     for( i = 0; i < h->frames.i_max_dpb; i++ )
 566     {
 567         h->frames.reference[i] = x264_frame_new( h );
 568         if( !h->frames.reference[i] )
 569             return NULL;
 570     }
 571     h->frames.reference[h->frames.i_max_dpb] = NULL;
 572     h->frames.i_last_idr = - h->param.i_keyint_max;
 573     h->frames.i_input    = 0;
 574     h->frames.last_nonb  = NULL;
 575
 576     h->i_ref0 = 0;
 577     h->i_ref1 = 0;
 578
 579     h->fdec = h->frames.reference[0];
 580
 581     if( x264_macroblock_cache_init( h ) < 0 );
 582         return NULL;
 583     x264_rdo_init( );
 584
 585     /* init CPU functions */
 586     x264_predict_16x16_init( h->param.cpu, h->predict_16x16 );
 587     x264_predict_8x8c_init( h->param.cpu, h->predict_8x8c );
 588     x264_predict_8x8_init( h->param.cpu, h->predict_8x8 );
 589     x264_predict_4x4_init( h->param.cpu, h->predict_4x4 );
 590
 591     x264_pixel_init( h->param.cpu, &h->pixf );
 592     x264_dct_init( h->param.cpu, &h->dctf );
 593     x264_mc_init( h->param.cpu, &h->mc );
 594     x264_csp_init( h->param.cpu, h->param.i_csp, &h->csp );
 595     x264_quant_init( h, h->param.cpu, &h->quantf );
 596     x264_deblock_init( h->param.cpu, &h->loopf );
 597
 598     memcpy( h->pixf.mbcmp,
 599             ( h->mb.b_lossless || h->param.analyse.i_subpel_refine <= 1 ) ? h->pixf.sad : h->pixf.satd,
 600             sizeof(h->pixf.mbcmp) );
 601
 602     /* rate control */
 603     if( x264_ratecontrol_new( h ) < 0 )
 604         return NULL;
 605
 606     x264_log( h, X264_LOG_INFO, "using cpu capabilities %s%s%s%s%s%s\n",
 607              param->cpu&X264_CPU_MMX ? "MMX " : "",
 608              param->cpu&X264_CPU_MMXEXT ? "MMXEXT " : "",
 609              param->cpu&X264_CPU_SSE ? "SSE " : "",
 610              param->cpu&X264_CPU_SSE2 ? "SSE2 " : "",
 611              param->cpu&X264_CPU_3DNOW ? "3DNow! " : "",
 612              param->cpu&X264_CPU_ALTIVEC ? "Altivec " : "" );
 613
 614     h->thread[0] = h;
 615     h->i_thread_num = 0;
 616     for( i = 1; i < param->i_threads; i++ )
 617         h->thread[i] = x264_malloc( sizeof(x264_t) );
 618
 619 #ifdef DEBUG_DUMP_FRAME
 620     {
 621         /* create or truncate the reconstructed video file */
 622         FILE *f = fopen( "fdec.yuv", "w" );
 623         if( f )
 624             fclose( f );
 625         else
 626         {
 627             x264_log( h, X264_LOG_ERROR, "can't write to fdec.yuv\n" );
 628             x264_free( h );
 629             return NULL;
 630         }
 631     }
 632 #endif
 633
 634     return h;
 635 }
 636
 637 /****************************************************************************
 638  * x264_encoder_reconfig:
 639  ****************************************************************************/
 640 int x264_encoder_reconfig( x264_t *h, x264_param_t *param )
 641 {
 642     h->param.i_bframe_bias = param->i_bframe_bias;
 643     h->param.i_deblocking_filter_alphac0 = param->i_deblocking_filter_alphac0;
 644     h->param.i_deblocking_filter_beta    = param->i_deblocking_filter_beta;
 645     h->param.analyse.i_me_method = param->analyse.i_me_method;
 646     h->param.analyse.i_me_range = param->analyse.i_me_range;
 647     h->param.analyse.i_subpel_refine = param->analyse.i_subpel_refine;
 648     h->param.analyse.i_trellis = param->analyse.i_trellis;
 649     h->param.analyse.intra = param->analyse.intra;
 650     h->param.analyse.inter = param->analyse.inter;
 651
 652     memcpy( h->pixf.mbcmp,
 653             ( h->mb.b_lossless || h->param.analyse.i_subpel_refine <= 1 ) ? h->pixf.sad : h->pixf.satd,
 654             sizeof(h->pixf.mbcmp) );
 655
 656     return x264_validate_parameters( h );
 657 }
 658
 659 /* internal usage */
 660 static void x264_nal_start( x264_t *h, int i_type, int i_ref_idc )
 661 {
 662     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 663
 664     nal->i_ref_idc = i_ref_idc;
 665     nal->i_type    = i_type;
 666
 667     bs_align_0( &h->out.bs );   /* not needed */
 668
 669     nal->i_payload= 0;
 670     nal->p_payload= &h->out.p_bitstream[bs_pos( &h->out.bs) / 8];
 671 }
 672 static void x264_nal_end( x264_t *h )
 673 {
 674     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 675
 676     bs_align_0( &h->out.bs );   /* not needed */
 677
 678     nal->i_payload = &h->out.p_bitstream[bs_pos( &h->out.bs)/8] - nal->p_payload;
 679
 680     h->out.i_nal++;
 681 }
 682
 683 /****************************************************************************
 684  * x264_encoder_headers:
 685  ****************************************************************************/
 686 int x264_encoder_headers( x264_t *h, x264_nal_t **pp_nal, int *pi_nal )
 687 {
 688     /* init bitstream context */
 689     h->out.i_nal = 0;
 690     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
 691
 692     /* Put SPS and PPS */
 693     if( h->i_frame == 0 )
 694     {
 695         /* identify ourself */
 696         x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
 697         x264_sei_version_write( h, &h->out.bs );
 698         x264_nal_end( h );
 699
 700         /* generate sequence parameters */
 701         x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
 702         x264_sps_write( &h->out.bs, h->sps );
 703         x264_nal_end( h );
 704
 705         /* generate picture parameters */
 706         x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
 707         x264_pps_write( &h->out.bs, h->pps );
 708         x264_nal_end( h );
 709     }
 710     /* now set output*/
 711     *pi_nal = h->out.i_nal;
 712     *pp_nal = &h->out.nal[0];
 713
 714     return 0;
 715 }
 716
 717
 718 static void x264_frame_put( x264_frame_t *list[X264_BFRAME_MAX], x264_frame_t *frame )
 719 {
 720     int i = 0;
 721     while( list[i] ) i++;
 722     list[i] = frame;
 723 }
 724
 725 static void x264_frame_push( x264_frame_t *list[X264_BFRAME_MAX], x264_frame_t *frame )
 726 {
 727     int i = 0;
 728     while( list[i] ) i++;
 729     while( i-- )
 730         list[i+1] = list[i];
 731     list[0] = frame;
 732 }
 733
 734 static x264_frame_t *x264_frame_get( x264_frame_t *list[X264_BFRAME_MAX+1] )
 735 {
 736     x264_frame_t *frame = list[0];
 737     int i;
 738     for( i = 0; list[i]; i++ )
 739         list[i] = list[i+1];
 740     return frame;
 741 }
 742
 743 static void x264_frame_sort( x264_frame_t *list[X264_BFRAME_MAX+1], int b_dts )
 744 {
 745     int i, b_ok;
 746     do {
 747         b_ok = 1;
 748         for( i = 0; list[i+1]; i++ )
 749         {
 750             int dtype = list[i]->i_type - list[i+1]->i_type;
 751             int dtime = list[i]->i_frame - list[i+1]->i_frame;
 752             int swap = b_dts ? dtype > 0 || ( dtype == 0 && dtime > 0 )
 753                              : dtime > 0;
 754             if( swap )
 755             {
 756                 XCHG( x264_frame_t*, list[i], list[i+1] );
 757                 b_ok = 0;
 758             }
 759         }
 760     } while( !b_ok );
 761 }
 762 #define x264_frame_sort_dts(list) x264_frame_sort(list, 1)
 763 #define x264_frame_sort_pts(list) x264_frame_sort(list, 0)
 764
 765 static inline void x264_reference_build_list( x264_t *h, int i_poc, int i_slice_type )
 766 {
 767     int i;
 768     int b_ok;
 769
 770     /* build ref list 0/1 */
 771     h->i_ref0 = 0;
 772     h->i_ref1 = 0;
 773     for( i = 1; i < h->frames.i_max_dpb; i++ )
 774     {
 775         if( h->frames.reference[i]->i_poc >= 0 )
 776         {
 777             if( h->frames.reference[i]->i_poc < i_poc )
 778             {
 779                 h->fref0[h->i_ref0++] = h->frames.reference[i];
 780             }
 781             else if( h->frames.reference[i]->i_poc > i_poc )
 782             {
 783                 h->fref1[h->i_ref1++] = h->frames.reference[i];
 784             }
 785         }
 786     }
 787
 788     /* Order ref0 from higher to lower poc */
 789     do
 790     {
 791         b_ok = 1;
 792         for( i = 0; i < h->i_ref0 - 1; i++ )
 793         {
 794             if( h->fref0[i]->i_poc < h->fref0[i+1]->i_poc )
 795             {
 796                 XCHG( x264_frame_t*, h->fref0[i], h->fref0[i+1] );
 797                 b_ok = 0;
 798                 break;
 799             }
 800         }
 801     } while( !b_ok );
 802     /* Order ref1 from lower to higher poc (bubble sort) for B-frame */
 803     do
 804     {
 805         b_ok = 1;
 806         for( i = 0; i < h->i_ref1 - 1; i++ )
 807         {
 808             if( h->fref1[i]->i_poc > h->fref1[i+1]->i_poc )
 809             {
 810                 XCHG( x264_frame_t*, h->fref1[i], h->fref1[i+1] );
 811                 b_ok = 0;
 812                 break;
 813             }
 814         }
 815     } while( !b_ok );
 816
 817     /* In the standard, a P-frame's ref list is sorted by frame_num.
 818      * We use POC, but check whether explicit reordering is needed */
 819     h->b_ref_reorder[0] =
 820     h->b_ref_reorder[1] = 0;
 821     if( i_slice_type == SLICE_TYPE_P )
 822     {
 823         for( i = 0; i < h->i_ref0 - 1; i++ )
 824             if( h->fref0[i]->i_frame_num < h->fref0[i+1]->i_frame_num )
 825             {
 826                 h->b_ref_reorder[0] = 1;
 827                 break;
 828             }
 829     }
 830
 831     h->i_ref1 = X264_MIN( h->i_ref1, h->frames.i_max_ref1 );
 832     h->i_ref0 = X264_MIN( h->i_ref0, h->frames.i_max_ref0 );
 833     h->i_ref0 = X264_MIN( h->i_ref0, 16 - h->i_ref1 );
 834 }
 835
 836 static inline void x264_fdec_deblock( x264_t *h )
 837 {
 838     /* apply deblocking filter to the current decoded picture */
 839     if( !h->sh.i_disable_deblocking_filter_idc )
 840     {
 841         TIMER_START( i_mtime_filter );
 842         x264_frame_deblocking_filter( h, h->sh.i_type );
 843         TIMER_STOP( i_mtime_filter );
 844     }
 845 }
 846
 847 static inline void x264_reference_update( x264_t *h )
 848 {
 849     int i;
 850
 851     x264_fdec_deblock( h );
 852
 853     /* expand border */
 854     x264_frame_expand_border( h->fdec );
 855
 856     /* create filtered images */
 857     x264_frame_filter( h->param.cpu, h->fdec );
 858
 859     /* expand border of filtered images */
 860     x264_frame_expand_border_filtered( h->fdec );
 861
 862     /* move lowres copy of the image to the ref frame */
 863     for( i = 0; i < 4; i++)
 864         XCHG( uint8_t*, h->fdec->lowres[i], h->fenc->lowres[i] );
 865
 866     /* adaptive B decision needs a pointer, since it can't use the ref lists */
 867     if( h->sh.i_type != SLICE_TYPE_B )
 868         h->frames.last_nonb = h->fdec;
 869
 870     /* move frame in the buffer */
 871     h->fdec = h->frames.reference[h->frames.i_max_dpb-1];
 872     for( i = h->frames.i_max_dpb-1; i > 0; i-- )
 873     {
 874         h->frames.reference[i] = h->frames.reference[i-1];
 875     }
 876     h->frames.reference[0] = h->fdec;
 877 }
 878
 879 static inline void x264_reference_reset( x264_t *h )
 880 {
 881     int i;
 882
 883     /* reset ref pictures */
 884     for( i = 1; i < h->frames.i_max_dpb; i++ )
 885     {
 886         h->frames.reference[i]->i_poc = -1;
 887     }
 888     h->frames.reference[0]->i_poc = 0;
 889 }
 890
 891 static inline void x264_slice_init( x264_t *h, int i_nal_type, int i_slice_type, int i_global_qp )
 892 {
 893     /* ------------------------ Create slice header  ----------------------- */
 894     if( i_nal_type == NAL_SLICE_IDR )
 895     {
 896         x264_slice_header_init( h, &h->sh, h->sps, h->pps, i_slice_type, h->i_idr_pic_id, h->i_frame_num, i_global_qp );
 897
 898         /* increment id */
 899         h->i_idr_pic_id = ( h->i_idr_pic_id + 1 ) % 65536;
 900     }
 901     else
 902     {
 903         x264_slice_header_init( h, &h->sh, h->sps, h->pps, i_slice_type, -1, h->i_frame_num, i_global_qp );
 904
 905         /* always set the real higher num of ref frame used */
 906         h->sh.b_num_ref_idx_override = 1;
 907         h->sh.i_num_ref_idx_l0_active = h->i_ref0 <= 0 ? 1 : h->i_ref0;
 908         h->sh.i_num_ref_idx_l1_active = h->i_ref1 <= 0 ? 1 : h->i_ref1;
 909     }
 910
 911     h->fdec->i_frame_num = h->sh.i_frame_num;
 912
 913     if( h->sps->i_poc_type == 0 )
 914     {
 915         h->sh.i_poc_lsb = h->fdec->i_poc & ( (1 << h->sps->i_log2_max_poc_lsb) - 1 );
 916         h->sh.i_delta_poc_bottom = 0;   /* XXX won't work for field */
 917     }
 918     else if( h->sps->i_poc_type == 1 )
 919     {
 920         /* FIXME TODO FIXME */
 921     }
 922     else
 923     {
 924         /* Nothing to do ? */
 925     }
 926
 927     x264_macroblock_slice_init( h );
 928 }
 929
 930 static int x264_slice_write( x264_t *h )
 931 {
 932     int i_skip;
 933     int mb_xy;
 934     int i;
 935
 936     /* init stats */
 937     memset( &h->stat.frame, 0, sizeof(h->stat.frame) );
 938
 939     /* Slice */
 940     x264_nal_start( h, h->i_nal_type, h->i_nal_ref_idc );
 941
 942     /* Slice header */
 943     x264_slice_header_write( &h->out.bs, &h->sh, h->i_nal_ref_idc );
 944     if( h->param.b_cabac )
 945     {
 946         /* alignment needed */
 947         bs_align_1( &h->out.bs );
 948
 949         /* init cabac */
 950         x264_cabac_context_init( &h->cabac, h->sh.i_type, h->sh.i_qp, h->sh.i_cabac_init_idc );
 951         x264_cabac_encode_init ( &h->cabac, &h->out.bs );
 952     }
 953     h->mb.i_last_qp = h->sh.i_qp;
 954     h->mb.i_last_dqp = 0;
 955
 956     for( mb_xy = h->sh.i_first_mb, i_skip = 0; mb_xy < h->sh.i_last_mb; mb_xy++ )
 957     {
 958         const int i_mb_y = mb_xy / h->sps->i_mb_width;
 959         const int i_mb_x = mb_xy % h->sps->i_mb_width;
 960
 961         int mb_spos = bs_pos(&h->out.bs);
 962
 963         /* load cache */
 964         x264_macroblock_cache_load( h, i_mb_x, i_mb_y );
 965
 966         /* analyse parameters
 967          * Slice I: choose I_4x4 or I_16x16 mode
 968          * Slice P: choose between using P mode or intra (4x4 or 16x16)
 969          * */
 970         TIMER_START( i_mtime_analyse );
 971         x264_macroblock_analyse( h );
 972         TIMER_STOP( i_mtime_analyse );
 973
 974         /* encode this macrobock -> be carefull it can change the mb type to P_SKIP if needed */
 975         TIMER_START( i_mtime_encode );
 976         x264_macroblock_encode( h );
 977         TIMER_STOP( i_mtime_encode );
 978
 979         TIMER_START( i_mtime_write );
 980         if( h->param.b_cabac )
 981         {
 982             if( mb_xy > h->sh.i_first_mb )
 983                 x264_cabac_encode_terminal( &h->cabac, 0 );
 984
 985             if( IS_SKIP( h->mb.i_type ) )
 986                 x264_cabac_mb_skip( h, 1 );
 987             else
 988             {
 989                 if( h->sh.i_type != SLICE_TYPE_I )
 990                     x264_cabac_mb_skip( h, 0 );
 991                 x264_macroblock_write_cabac( h, &h->cabac );
 992             }
 993         }
 994         else
 995         {
 996             if( IS_SKIP( h->mb.i_type ) )
 997                 i_skip++;
 998             else
 999             {
1000                 if( h->sh.i_type != SLICE_TYPE_I )
1001                 {
1002                     bs_write_ue( &h->out.bs, i_skip );  /* skip run */
1003                     i_skip = 0;
1004                 }
1005                 x264_macroblock_write_cavlc( h, &h->out.bs );
1006             }
1007         }
1008         TIMER_STOP( i_mtime_write );
1009
1010 #if VISUALIZE
1011         if( h->param.b_visualize )
1012             x264_visualize_mb( h );
1013 #endif
1014
1015         /* save cache */
1016         x264_macroblock_cache_save( h );
1017
1018         /* accumulate mb stats */
1019         h->stat.frame.i_mb_count[h->mb.i_type]++;
1020         if( !IS_SKIP(h->mb.i_type) && !IS_INTRA(h->mb.i_type) && !IS_DIRECT(h->mb.i_type) )
1021         {
1022             if( h->mb.i_partition != D_8x8 )
1023                 h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_partition ] ] += 4;
1024             else
1025                 for( i = 0; i < 4; i++ )
1026                     h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_sub_partition[i] ] ] ++;
1027             if( h->param.i_frame_reference > 1 )
1028             {
1029                 for( i = 0; i < 4; i++ )
1030                 {
1031                     int i_ref = h->mb.cache.ref[0][ x264_scan8[4*i] ];
1032                     if( i_ref >= 0 )
1033                         h->stat.frame.i_mb_count_ref[i_ref] ++;
1034                 }
1035             }
1036         }
1037         if( h->mb.i_cbp_luma && !IS_INTRA(h->mb.i_type) )
1038         {
1039             h->stat.frame.i_mb_count_8x8dct[0] ++;
1040             h->stat.frame.i_mb_count_8x8dct[1] += h->mb.b_transform_8x8;
1041         }
1042
1043         if( h->mb.b_variable_qp )
1044             x264_ratecontrol_mb(h, bs_pos(&h->out.bs) - mb_spos);
1045     }
1046
1047     if( h->param.b_cabac )
1048     {
1049         /* end of slice */
1050         x264_cabac_encode_terminal( &h->cabac, 1 );
1051     }
1052     else if( i_skip > 0 )
1053     {
1054         bs_write_ue( &h->out.bs, i_skip );  /* last skip run */
1055     }
1056
1057     if( h->param.b_cabac )
1058     {
1059         x264_cabac_encode_flush( &h->cabac );
1060
1061     }
1062     else
1063     {
1064         /* rbsp_slice_trailing_bits */
1065         bs_rbsp_trailing( &h->out.bs );
1066     }
1067
1068     x264_nal_end( h );
1069
1070     /* Compute misc bits */
1071     h->stat.frame.i_misc_bits = bs_pos( &h->out.bs )
1072                               + NALU_OVERHEAD * 8
1073                               - h->stat.frame.i_itex_bits
1074                               - h->stat.frame.i_ptex_bits
1075                               - h->stat.frame.i_hdr_bits;
1076
1077     return 0;
1078 }
1079
1080 static inline int x264_slices_write( x264_t *h )
1081 {
1082     int i_frame_size;
1083
1084 #if VISUALIZE
1085     if( h->param.b_visualize )
1086         x264_visualize_init( h );
1087 #endif
1088
1089     if( h->param.i_threads == 1 )
1090     {
1091         x264_ratecontrol_threads_start( h );
1092         x264_slice_write( h );
1093         i_frame_size = h->out.nal[h->out.i_nal-1].i_payload;
1094     }
1095     else
1096     {
1097         int i_nal = h->out.i_nal;
1098         int i_bs_size = h->out.i_bitstream / h->param.i_threads;
1099         int i;
1100         /* duplicate contexts */
1101         for( i = 0; i < h->param.i_threads; i++ )
1102         {
1103             x264_t *t = h->thread[i];
1104             if( i > 0 )
1105             {
1106                 memcpy( t, h, sizeof(x264_t) );
1107                 t->out.p_bitstream += i*i_bs_size;
1108                 bs_init( &t->out.bs, t->out.p_bitstream, i_bs_size );
1109                 t->i_thread_num = i;
1110             }
1111             t->sh.i_first_mb = (i    * h->sps->i_mb_height / h->param.i_threads) * h->sps->i_mb_width;
1112             t->sh.i_last_mb = ((i+1) * h->sps->i_mb_height / h->param.i_threads) * h->sps->i_mb_width;
1113             t->out.i_nal = i_nal + i;
1114         }
1115         x264_ratecontrol_threads_start( h );
1116
1117         /* dispatch */
1118 #ifdef HAVE_PTHREAD
1119         {
1120             pthread_t handles[X264_SLICE_MAX];
1121             for( i = 0; i < h->param.i_threads; i++ )
1122                 pthread_create( &handles[i], NULL, (void*)x264_slice_write, (void*)h->thread[i] );
1123             for( i = 0; i < h->param.i_threads; i++ )
1124                 pthread_join( handles[i], NULL );
1125         }
1126 #else
1127         for( i = 0; i < h->param.i_threads; i++ )
1128             x264_slice_write( h->thread[i] );
1129 #endif
1130
1131         /* merge contexts */
1132         i_frame_size = h->out.nal[i_nal].i_payload;
1133         for( i = 1; i < h->param.i_threads; i++ )
1134         {
1135             int j;
1136             x264_t *t = h->thread[i];
1137             h->out.nal[i_nal+i] = t->out.nal[i_nal+i];
1138             i_frame_size += t->out.nal[i_nal+i].i_payload;
1139             // all entries in stat.frame are ints
1140             for( j = 0; j < sizeof(h->stat.frame) / sizeof(int); j++ )
1141                 ((int*)&h->stat.frame)[j] += ((int*)&t->stat.frame)[j];
1142         }
1143         h->out.i_nal = i_nal + h->param.i_threads;
1144     }
1145
1146 #if VISUALIZE
1147     if( h->param.b_visualize )
1148     {
1149         x264_visualize_show( h );
1150         x264_visualize_close( h );
1151     }
1152 #endif
1153
1154     return i_frame_size;
1155 }
1156
1157 /****************************************************************************
1158  * x264_encoder_encode:
1159  *  XXX: i_poc   : is the poc of the current given picture
1160  *       i_frame : is the number of the frame being coded
1161  *  ex:  type frame poc
1162  *       I      0   2*0
1163  *       P      1   2*3
1164  *       B      2   2*1
1165  *       B      3   2*2
1166  *       P      4   2*6
1167  *       B      5   2*4
1168  *       B      6   2*5
1169  ****************************************************************************/
1170 int     x264_encoder_encode( x264_t *h,
1171                              x264_nal_t **pp_nal, int *pi_nal,
1172                              x264_picture_t *pic_in,
1173                              x264_picture_t *pic_out )
1174 {
1175     x264_frame_t   *frame_psnr = h->fdec; /* just to keep the current decoded frame for psnr calculation */
1176     int     i_nal_type;
1177     int     i_nal_ref_idc;
1178     int     i_slice_type;
1179     int     i_frame_size;
1180
1181     int i;
1182
1183     int   i_global_qp;
1184
1185     char psz_message[80];
1186
1187     /* no data out */
1188     *pi_nal = 0;
1189     *pp_nal = NULL;
1190
1191
1192     /* ------------------- Setup new frame from picture -------------------- */
1193     TIMER_START( i_mtime_encode_frame );
1194     if( pic_in != NULL )
1195     {
1196         /* 1: Copy the picture to a frame and move it to a buffer */
1197         x264_frame_t *fenc = x264_frame_get( h->frames.unused );
1198
1199         x264_frame_copy_picture( h, fenc, pic_in );
1200
1201         if( h->param.i_width % 16 || h->param.i_height % 16 )
1202             x264_frame_expand_border_mod16( h, fenc );
1203
1204         fenc->i_frame = h->frames.i_input++;
1205
1206         x264_frame_put( h->frames.next, fenc );
1207
1208         if( h->frames.b_have_lowres )
1209             x264_frame_init_lowres( h->param.cpu, fenc );
1210
1211         if( h->frames.i_input <= h->frames.i_delay )
1212         {
1213             /* Nothing yet to encode */
1214             /* waiting for filling bframe buffer */
1215             pic_out->i_type = X264_TYPE_AUTO;
1216             return 0;
1217         }
1218     }
1219
1220     if( h->frames.current[0] == NULL )
1221     {
1222         int bframes = 0;
1223         /* 2: Select frame types */
1224         if( h->frames.next[0] == NULL )
1225             return 0;
1226
1227         x264_slicetype_decide( h );
1228
1229         /* 3: move some B-frames and 1 non-B to encode queue */
1230         while( IS_X264_TYPE_B( h->frames.next[bframes]->i_type ) )
1231             bframes++;
1232         x264_frame_put( h->frames.current, x264_frame_get( &h->frames.next[bframes] ) );
1233         /* FIXME: when max B-frames > 3, BREF may no longer be centered after GOP closing */
1234         if( h->param.b_bframe_pyramid && bframes > 1 )
1235         {
1236             x264_frame_t *mid = x264_frame_get( &h->frames.next[bframes/2] );
1237             mid->i_type = X264_TYPE_BREF;
1238             x264_frame_put( h->frames.current, mid );
1239             bframes--;
1240         }
1241         while( bframes-- )
1242             x264_frame_put( h->frames.current, x264_frame_get( h->frames.next ) );
1243     }
1244     TIMER_STOP( i_mtime_encode_frame );
1245
1246     /* ------------------- Get frame to be encoded ------------------------- */
1247     /* 4: get picture to encode */
1248     h->fenc = x264_frame_get( h->frames.current );
1249     if( h->fenc == NULL )
1250     {
1251         /* Nothing yet to encode (ex: waiting for I/P with B frames) */
1252         /* waiting for filling bframe buffer */
1253         pic_out->i_type = X264_TYPE_AUTO;
1254         return 0;
1255     }
1256
1257 do_encode:
1258
1259     if( h->fenc->i_type == X264_TYPE_IDR )
1260     {
1261         h->frames.i_last_idr = h->fenc->i_frame;
1262     }
1263
1264     /* ------------------- Setup frame context ----------------------------- */
1265     /* 5: Init data dependant of frame type */
1266     TIMER_START( i_mtime_encode_frame );
1267     if( h->fenc->i_type == X264_TYPE_IDR )
1268     {
1269         /* reset ref pictures */
1270         x264_reference_reset( h );
1271
1272         i_nal_type    = NAL_SLICE_IDR;
1273         i_nal_ref_idc = NAL_PRIORITY_HIGHEST;
1274         i_slice_type = SLICE_TYPE_I;
1275     }
1276     else if( h->fenc->i_type == X264_TYPE_I )
1277     {
1278         i_nal_type    = NAL_SLICE;
1279         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
1280         i_slice_type = SLICE_TYPE_I;
1281     }
1282     else if( h->fenc->i_type == X264_TYPE_P )
1283     {
1284         i_nal_type    = NAL_SLICE;
1285         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
1286         i_slice_type = SLICE_TYPE_P;
1287     }
1288     else if( h->fenc->i_type == X264_TYPE_BREF )
1289     {
1290         i_nal_type    = NAL_SLICE;
1291         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* maybe add MMCO to forget it? -> low */
1292         i_slice_type = SLICE_TYPE_B;
1293     }
1294     else    /* B frame */
1295     {
1296         i_nal_type    = NAL_SLICE;
1297         i_nal_ref_idc = NAL_PRIORITY_DISPOSABLE;
1298         i_slice_type = SLICE_TYPE_B;
1299     }
1300
1301     h->fdec->i_poc =
1302     h->fenc->i_poc = 2 * (h->fenc->i_frame - h->frames.i_last_idr);
1303     h->fdec->i_type = h->fenc->i_type;
1304     h->fdec->i_frame = h->fenc->i_frame;
1305     h->fenc->b_kept_as_ref =
1306     h->fdec->b_kept_as_ref = i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE;
1307
1308
1309
1310     /* ------------------- Init                ----------------------------- */
1311     /* build ref list 0/1 */
1312     x264_reference_build_list( h, h->fdec->i_poc, i_slice_type );
1313
1314     /* Init the rate control */
1315     x264_ratecontrol_start( h, i_slice_type, h->fenc->i_qpplus1 );
1316     i_global_qp = x264_ratecontrol_qp( h );
1317
1318     pic_out->i_qpplus1 =
1319     h->fdec->i_qpplus1 = i_global_qp + 1;
1320
1321     if( i_slice_type == SLICE_TYPE_B )
1322         x264_macroblock_bipred_init( h );
1323
1324     /* ------------------------ Create slice header  ----------------------- */
1325     x264_slice_init( h, i_nal_type, i_slice_type, i_global_qp );
1326
1327     if( h->fenc->b_kept_as_ref )
1328         h->i_frame_num++;
1329
1330     /* ---------------------- Write the bitstream -------------------------- */
1331     /* Init bitstream context */
1332     h->out.i_nal = 0;
1333     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
1334
1335     if(h->param.b_aud){
1336         int pic_type;
1337
1338         if(i_slice_type == SLICE_TYPE_I)
1339             pic_type = 0;
1340         else if(i_slice_type == SLICE_TYPE_P)
1341             pic_type = 1;
1342         else if(i_slice_type == SLICE_TYPE_B)
1343             pic_type = 2;
1344         else
1345             pic_type = 7;
1346
1347         x264_nal_start(h, NAL_AUD, NAL_PRIORITY_DISPOSABLE);
1348         bs_write(&h->out.bs, 3, pic_type);
1349         bs_rbsp_trailing(&h->out.bs);
1350         x264_nal_end(h);
1351     }
1352
1353     h->i_nal_type = i_nal_type;
1354     h->i_nal_ref_idc = i_nal_ref_idc;
1355
1356     /* Write SPS and PPS */
1357     if( i_nal_type == NAL_SLICE_IDR && h->param.b_repeat_headers )
1358     {
1359         if( h->fenc->i_frame == 0 )
1360         {
1361             /* identify ourself */
1362             x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
1363             x264_sei_version_write( h, &h->out.bs );
1364             x264_nal_end( h );
1365         }
1366
1367         /* generate sequence parameters */
1368         x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
1369         x264_sps_write( &h->out.bs, h->sps );
1370         x264_nal_end( h );
1371
1372         /* generate picture parameters */
1373         x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
1374         x264_pps_write( &h->out.bs, h->pps );
1375         x264_nal_end( h );
1376     }
1377
1378     /* Write frame */
1379     i_frame_size = x264_slices_write( h );
1380
1381     /* restore CPU state (before using float again) */
1382     x264_cpu_restore( h->param.cpu );
1383
1384     if( i_slice_type == SLICE_TYPE_P && !h->param.rc.b_stat_read
1385         && h->param.i_scenecut_threshold >= 0 )
1386     {
1387         const int *mbs = h->stat.frame.i_mb_count;
1388         int i_mb_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1389         int i_mb_p = mbs[P_L0] + mbs[P_8x8];
1390         int i_mb_s = mbs[P_SKIP];
1391         int i_mb   = h->sps->i_mb_width * h->sps->i_mb_height;
1392         int64_t i_inter_cost = h->stat.frame.i_inter_cost;
1393         int64_t i_intra_cost = h->stat.frame.i_intra_cost;
1394
1395         float f_bias;
1396         int i_gop_size = h->fenc->i_frame - h->frames.i_last_idr;
1397         float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
1398         /* magic numbers pulled out of thin air */
1399         float f_thresh_min = f_thresh_max * h->param.i_keyint_min
1400                              / ( h->param.i_keyint_max * 4 );
1401         if( h->param.i_keyint_min == h->param.i_keyint_max )
1402              f_thresh_min= f_thresh_max;
1403
1404         /* macroblock_analyse() doesn't further analyse skipped mbs,
1405          * so we have to guess their cost */
1406         if( i_mb_s < i_mb )
1407             i_intra_cost = i_intra_cost * i_mb / (i_mb - i_mb_s);
1408
1409         if( i_gop_size < h->param.i_keyint_min / 4 )
1410             f_bias = f_thresh_min / 4;
1411         else if( i_gop_size <= h->param.i_keyint_min )
1412             f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
1413         else
1414         {
1415             f_bias = f_thresh_min
1416                      + ( f_thresh_max - f_thresh_min )
1417                        * ( i_gop_size - h->param.i_keyint_min )
1418                        / ( h->param.i_keyint_max - h->param.i_keyint_min );
1419         }
1420         f_bias = X264_MIN( f_bias, 1.0 );
1421
1422         /* Bad P will be reencoded as I */
1423         if( i_mb_s < i_mb &&
1424             i_inter_cost >= (1.0 - f_bias) * i_intra_cost )
1425         {
1426             int b;
1427
1428             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",
1429                       h->fenc->i_frame,
1430                       (double)i_intra_cost, (double)i_inter_cost,
1431                       (double)i_inter_cost / i_intra_cost,
1432                       f_bias, i_gop_size,
1433                       i_mb_i, i_mb_p, i_mb_s );
1434
1435             /* Restore frame num */
1436             h->i_frame_num--;
1437
1438             for( b = 0; h->frames.current[b] && IS_X264_TYPE_B( h->frames.current[b]->i_type ); b++ );
1439             if( b > 0 )
1440             {
1441                 /* If using B-frames, force GOP to be closed.
1442                  * Even if this frame is going to be I and not IDR, forcing a
1443                  * P-frame before the scenecut will probably help compression.
1444                  *
1445                  * We don't yet know exactly which frame is the scene cut, so
1446                  * we can't assign an I-frame. Instead, change the previous
1447                  * B-frame to P, and rearrange coding order. */
1448
1449                 if( h->param.b_bframe_adaptive || b > 1 )
1450                     h->fenc->i_type = X264_TYPE_AUTO;
1451                 x264_frame_sort_pts( h->frames.current );
1452                 x264_frame_push( h->frames.next, h->fenc );
1453                 h->fenc = h->frames.current[b-1];
1454                 h->frames.current[b-1] = NULL;
1455                 h->fenc->i_type = X264_TYPE_P;
1456                 x264_frame_sort_dts( h->frames.current );
1457             }
1458             /* Do IDR if needed */
1459             else if( i_gop_size >= h->param.i_keyint_min )
1460             {
1461                 x264_frame_t *tmp;
1462
1463                 /* Reset */
1464                 h->i_frame_num = 0;
1465
1466                 /* Reinit field of fenc */
1467                 h->fenc->i_type = X264_TYPE_IDR;
1468                 h->fenc->i_poc = 0;
1469
1470                 /* Put enqueued frames back in the pool */
1471                 while( (tmp = x264_frame_get( h->frames.current ) ) != NULL )
1472                     x264_frame_put( h->frames.next, tmp );
1473                 x264_frame_sort_pts( h->frames.next );
1474             }
1475             else
1476             {
1477                 h->fenc->i_type = X264_TYPE_I;
1478             }
1479             goto do_encode;
1480         }
1481     }
1482
1483     /* End bitstream, set output  */
1484     *pi_nal = h->out.i_nal;
1485     *pp_nal = h->out.nal;
1486
1487     /* Set output picture properties */
1488     if( i_slice_type == SLICE_TYPE_I )
1489         pic_out->i_type = i_nal_type == NAL_SLICE_IDR ? X264_TYPE_IDR : X264_TYPE_I;
1490     else if( i_slice_type == SLICE_TYPE_P )
1491         pic_out->i_type = X264_TYPE_P;
1492     else
1493         pic_out->i_type = X264_TYPE_B;
1494     pic_out->i_pts = h->fenc->i_pts;
1495
1496     pic_out->img.i_plane = h->fdec->i_plane;
1497     for(i = 0; i < 4; i++){
1498         pic_out->img.i_stride[i] = h->fdec->i_stride[i];
1499         pic_out->img.plane[i] = h->fdec->plane[i];
1500     }
1501
1502     /* ---------------------- Update encoder state ------------------------- */
1503
1504     /* update rc */
1505     x264_cpu_restore( h->param.cpu );
1506     x264_ratecontrol_end( h, i_frame_size * 8 );
1507
1508     /* handle references */
1509     if( i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE )
1510         x264_reference_update( h );
1511 #ifdef DEBUG_DUMP_FRAME
1512     else
1513         x264_fdec_deblock( h );
1514 #endif
1515     x264_frame_put( h->frames.unused, h->fenc );
1516
1517     /* increase frame count */
1518     h->i_frame++;
1519
1520     /* restore CPU state (before using float again) */
1521     x264_cpu_restore( h->param.cpu );
1522
1523     x264_noise_reduction_update( h );
1524
1525     TIMER_STOP( i_mtime_encode_frame );
1526
1527     /* ---------------------- Compute/Print statistics --------------------- */
1528     /* Slice stat */
1529     h->stat.i_slice_count[i_slice_type]++;
1530     h->stat.i_slice_size[i_slice_type] += i_frame_size + NALU_OVERHEAD;
1531     h->stat.i_slice_qp[i_slice_type] += i_global_qp;
1532
1533     for( i = 0; i < 19; i++ )
1534         h->stat.i_mb_count[h->sh.i_type][i] += h->stat.frame.i_mb_count[i];
1535     for( i = 0; i < 2; i++ )
1536         h->stat.i_mb_count_8x8dct[i] += h->stat.frame.i_mb_count_8x8dct[i];
1537     if( h->sh.i_type != SLICE_TYPE_I )
1538     {
1539         for( i = 0; i < 7; i++ )
1540             h->stat.i_mb_count_size[h->sh.i_type][i] += h->stat.frame.i_mb_count_size[i];
1541         for( i = 0; i < 16; i++ )
1542             h->stat.i_mb_count_ref[h->sh.i_type][i] += h->stat.frame.i_mb_count_ref[i];
1543     }
1544     if( i_slice_type == SLICE_TYPE_B )
1545     {
1546         h->stat.i_direct_frames[ h->sh.b_direct_spatial_mv_pred ] ++;
1547         if( h->mb.b_direct_auto_write )
1548         {
1549             //FIXME somewhat arbitrary time constants
1550             if( h->stat.i_direct_score[0] + h->stat.i_direct_score[1] > h->mb.i_mb_count )
1551             {
1552                 for( i = 0; i < 2; i++ )
1553                     h->stat.i_direct_score[i] = h->stat.i_direct_score[i] * 9/10;
1554             }
1555             for( i = 0; i < 2; i++ )
1556                 h->stat.i_direct_score[i] += h->stat.frame.i_direct_score[i];
1557         }
1558     }
1559
1560     if( h->param.analyse.b_psnr )
1561     {
1562         int64_t i_sqe_y, i_sqe_u, i_sqe_v;
1563
1564         /* PSNR */
1565         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 );
1566         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);
1567         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);
1568         x264_cpu_restore( h->param.cpu );
1569
1570         h->stat.i_sqe_global[i_slice_type] += i_sqe_y + i_sqe_u + i_sqe_v;
1571         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 );
1572         h->stat.f_psnr_mean_y[i_slice_type] += x264_psnr( i_sqe_y, h->param.i_width * h->param.i_height );
1573         h->stat.f_psnr_mean_u[i_slice_type] += x264_psnr( i_sqe_u, h->param.i_width * h->param.i_height / 4 );
1574         h->stat.f_psnr_mean_v[i_slice_type] += x264_psnr( i_sqe_v, h->param.i_width * h->param.i_height / 4 );
1575
1576         snprintf( psz_message, 80, " PSNR Y:%2.2f U:%2.2f V:%2.2f",
1577                   x264_psnr( i_sqe_y, h->param.i_width * h->param.i_height ),
1578                   x264_psnr( i_sqe_u, h->param.i_width * h->param.i_height / 4),
1579                   x264_psnr( i_sqe_v, h->param.i_width * h->param.i_height / 4) );
1580         psz_message[79] = '\0';
1581     }
1582     else
1583     {
1584         psz_message[0] = '\0';
1585     }
1586
1587     x264_log( h, X264_LOG_DEBUG,
1588                   "frame=%4d QP=%i NAL=%d Slice:%c Poc:%-3d I:%-4d P:%-4d SKIP:%-4d size=%d bytes%s\n",
1589               h->i_frame - 1,
1590               i_global_qp,
1591               i_nal_ref_idc,
1592               i_slice_type == SLICE_TYPE_I ? 'I' : (i_slice_type == SLICE_TYPE_P ? 'P' : 'B' ),
1593               frame_psnr->i_poc,
1594               h->stat.frame.i_mb_count_i,
1595               h->stat.frame.i_mb_count_p,
1596               h->stat.frame.i_mb_count_skip,
1597               i_frame_size,
1598               psz_message );
1599
1600
1601 #ifdef DEBUG_MB_TYPE
1602 {
1603     static const char mb_chars[] = { 'i', 'i', 'I', 'C', 'P', '8', 'S',
1604         'D', '<', 'X', 'B', 'X', '>', 'B', 'B', 'B', 'B', '8', 'S' };
1605     int mb_xy;
1606     for( mb_xy = 0; mb_xy < h->sps->i_mb_width * h->sps->i_mb_height; mb_xy++ )
1607     {
1608         if( h->mb.type[mb_xy] < 19 && h->mb.type[mb_xy] >= 0 )
1609             fprintf( stderr, "%c ", mb_chars[ h->mb.type[mb_xy] ] );
1610         else
1611             fprintf( stderr, "? " );
1612
1613         if( (mb_xy+1) % h->sps->i_mb_width == 0 )
1614             fprintf( stderr, "\n" );
1615     }
1616 }
1617 #endif
1618
1619 #ifdef DEBUG_DUMP_FRAME
1620     /* Dump reconstructed frame */
1621     x264_frame_dump( h, frame_psnr, "fdec.yuv" );
1622 #endif
1623     return 0;
1624 }
1625
1626 /****************************************************************************
1627  * x264_encoder_close:
1628  ****************************************************************************/
1629 void    x264_encoder_close  ( x264_t *h )
1630 {
1631 #ifdef DEBUG_BENCHMARK
1632     int64_t i_mtime_total = i_mtime_analyse + i_mtime_encode + i_mtime_write + i_mtime_filter + 1;
1633 #endif
1634     int64_t i_yuv_size = 3 * h->param.i_width * h->param.i_height / 2;
1635     int i;
1636
1637 #ifdef DEBUG_BENCHMARK
1638     x264_log( h, X264_LOG_INFO,
1639               "analyse=%d(%lldms) encode=%d(%lldms) write=%d(%lldms) filter=%d(%lldms)\n",
1640               (int)(100*i_mtime_analyse/i_mtime_total), i_mtime_analyse/1000,
1641               (int)(100*i_mtime_encode/i_mtime_total), i_mtime_encode/1000,
1642               (int)(100*i_mtime_write/i_mtime_total), i_mtime_write/1000,
1643               (int)(100*i_mtime_filter/i_mtime_total), i_mtime_filter/1000 );
1644 #endif
1645
1646     /* Slices used and PSNR */
1647     for( i=0; i<5; i++ )
1648     {
1649         static const int slice_order[] = { SLICE_TYPE_I, SLICE_TYPE_SI, SLICE_TYPE_P, SLICE_TYPE_SP, SLICE_TYPE_B };
1650         static const char *slice_name[] = { "P", "B", "I", "SP", "SI" };
1651         int i_slice = slice_order[i];
1652
1653         if( h->stat.i_slice_count[i_slice] > 0 )
1654         {
1655             const int i_count = h->stat.i_slice_count[i_slice];
1656             if( h->param.analyse.b_psnr )
1657             {
1658                 x264_log( h, X264_LOG_INFO,
1659                           "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",
1660                           slice_name[i_slice],
1661                           i_count,
1662                           (double)h->stat.i_slice_qp[i_slice] / i_count,
1663                           (double)h->stat.i_slice_size[i_slice] / i_count,
1664                           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,
1665                           h->stat.f_psnr_average[i_slice] / i_count,
1666                           x264_psnr( h->stat.i_sqe_global[i_slice], i_count * i_yuv_size ) );
1667             }
1668             else
1669             {
1670                 x264_log( h, X264_LOG_INFO,
1671                           "slice %s:%-5d Avg QP:%5.2f  size:%6.0f\n",
1672                           slice_name[i_slice],
1673                           i_count,
1674                           (double)h->stat.i_slice_qp[i_slice] / i_count,
1675                           (double)h->stat.i_slice_size[i_slice] / i_count );
1676             }
1677         }
1678     }
1679
1680     /* MB types used */
1681     if( h->stat.i_slice_count[SLICE_TYPE_I] > 0 )
1682     {
1683         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_I];
1684         const double i_count = h->stat.i_slice_count[SLICE_TYPE_I] * h->mb.i_mb_count / 100.0;
1685         x264_log( h, X264_LOG_INFO,
1686                   "mb I  I16..4: %4.1f%% %4.1f%% %4.1f%%\n",
1687                   i_mb_count[I_16x16]/ i_count,
1688                   i_mb_count[I_8x8]  / i_count,
1689                   i_mb_count[I_4x4]  / i_count );
1690     }
1691     if( h->stat.i_slice_count[SLICE_TYPE_P] > 0 )
1692     {
1693         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_P];
1694         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_P];
1695         const double i_count = h->stat.i_slice_count[SLICE_TYPE_P] * h->mb.i_mb_count / 100.0;
1696         x264_log( h, X264_LOG_INFO,
1697                   "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",
1698                   i_mb_count[I_16x16]/ i_count,
1699                   i_mb_count[I_8x8]  / i_count,
1700                   i_mb_count[I_4x4]  / i_count,
1701                   i_mb_size[PIXEL_16x16] / (i_count*4),
1702                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
1703                   i_mb_size[PIXEL_8x8] / (i_count*4),
1704                   (i_mb_size[PIXEL_8x4] + i_mb_size[PIXEL_4x8]) / (i_count*4),
1705                   i_mb_size[PIXEL_4x4] / (i_count*4),
1706                   i_mb_count[P_SKIP] / i_count );
1707     }
1708     if( h->stat.i_slice_count[SLICE_TYPE_B] > 0 )
1709     {
1710         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_B];
1711         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_B];
1712         const double i_count = h->stat.i_slice_count[SLICE_TYPE_B] * h->mb.i_mb_count / 100.0;
1713         x264_log( h, X264_LOG_INFO,
1714                   "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",
1715                   i_mb_count[I_16x16]  / i_count,
1716                   i_mb_count[I_8x8]    / i_count,
1717                   i_mb_count[I_4x4]    / i_count,
1718                   i_mb_size[PIXEL_16x16] / (i_count*4),
1719                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
1720                   i_mb_size[PIXEL_8x8] / (i_count*4),
1721                   i_mb_count[B_DIRECT] / i_count,
1722                   i_mb_count[B_SKIP]   / i_count );
1723     }
1724
1725     x264_ratecontrol_summary( h );
1726
1727     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 )
1728     {
1729         const int i_count = h->stat.i_slice_count[SLICE_TYPE_I] +
1730                             h->stat.i_slice_count[SLICE_TYPE_P] +
1731                             h->stat.i_slice_count[SLICE_TYPE_B];
1732         float fps = (float) h->param.i_fps_num / h->param.i_fps_den;
1733 #define SUM3(p) (p[SLICE_TYPE_I] + p[SLICE_TYPE_P] + p[SLICE_TYPE_B])
1734 #define SUM3b(p,o) (p[SLICE_TYPE_I][o] + p[SLICE_TYPE_P][o] + p[SLICE_TYPE_B][o])
1735         float f_bitrate = fps * SUM3(h->stat.i_slice_size) / i_count / 125;
1736
1737         if( h->param.analyse.b_transform_8x8 )
1738         {
1739             int64_t i_i8x8 = SUM3b( h->stat.i_mb_count, I_8x8 );
1740             int64_t i_intra = i_i8x8 + SUM3b( h->stat.i_mb_count, I_4x4 )
1741                                      + SUM3b( h->stat.i_mb_count, I_16x16 );
1742             x264_log( h, X264_LOG_INFO, "8x8 transform  intra:%.1f%%  inter:%.1f%%\n",
1743                       100. * i_i8x8 / i_intra,
1744                       100. * h->stat.i_mb_count_8x8dct[1] / h->stat.i_mb_count_8x8dct[0] );
1745         }
1746
1747         if( h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
1748             && h->stat.i_slice_count[SLICE_TYPE_B] )
1749         {
1750             x264_log( h, X264_LOG_INFO, "direct mvs  spatial:%.1f%%  temporal:%.1f%%\n",
1751                       h->stat.i_direct_frames[1] * 100. / h->stat.i_slice_count[SLICE_TYPE_B],
1752                       h->stat.i_direct_frames[0] * 100. / h->stat.i_slice_count[SLICE_TYPE_B] );
1753         }
1754
1755         if( h->param.i_frame_reference > 1 )
1756         {
1757             int i_slice;
1758             for( i_slice = 0; i_slice < 2; i_slice++ )
1759             {
1760                 char buf[200];
1761                 char *p = buf;
1762                 int64_t i_den = 0;
1763                 int i_max = 0;
1764                 for( i = 0; i < h->param.i_frame_reference; i++ )
1765                     if( h->stat.i_mb_count_ref[i_slice][i] )
1766                     {
1767                         i_den += h->stat.i_mb_count_ref[i_slice][i];
1768                         i_max = i;
1769                     }
1770                 if( i_max == 0 )
1771                     continue;
1772                 for( i = 0; i <= i_max; i++ )
1773                     p += sprintf( p, " %4.1f%%", 100. * h->stat.i_mb_count_ref[i_slice][i] / i_den );
1774                 x264_log( h, X264_LOG_INFO, "ref %c %s\n", i_slice==SLICE_TYPE_P ? 'P' : 'B', buf );
1775             }
1776         }
1777
1778         if( h->param.analyse.b_psnr )
1779             x264_log( h, X264_LOG_INFO,
1780                       "PSNR Mean Y:%6.3f U:%6.3f V:%6.3f Avg:%6.3f Global:%6.3f kb/s:%.2f\n",
1781                       SUM3( h->stat.f_psnr_mean_y ) / i_count,
1782                       SUM3( h->stat.f_psnr_mean_u ) / i_count,
1783                       SUM3( h->stat.f_psnr_mean_v ) / i_count,
1784                       SUM3( h->stat.f_psnr_average ) / i_count,
1785                       x264_psnr( SUM3( h->stat.i_sqe_global ), i_count * i_yuv_size ),
1786                       f_bitrate );
1787         else
1788             x264_log( h, X264_LOG_INFO, "kb/s:%.1f\n", f_bitrate );
1789     }
1790
1791     /* frames */
1792     for( i = 0; i < X264_BFRAME_MAX + 3; i++ )
1793     {
1794         if( h->frames.current[i] ) x264_frame_delete( h->frames.current[i] );
1795         if( h->frames.next[i] )    x264_frame_delete( h->frames.next[i] );
1796         if( h->frames.unused[i] )  x264_frame_delete( h->frames.unused[i] );
1797     }
1798     /* ref frames */
1799     for( i = 0; i < h->frames.i_max_dpb; i++ )
1800     {
1801         x264_frame_delete( h->frames.reference[i] );
1802     }
1803
1804     /* rc */
1805     x264_ratecontrol_delete( h );
1806
1807     /* param */
1808     if( h->param.rc.psz_stat_out )
1809         free( h->param.rc.psz_stat_out );
1810     if( h->param.rc.psz_stat_in )
1811         free( h->param.rc.psz_stat_in );
1812     if( h->param.rc.psz_rc_eq )
1813         free( h->param.rc.psz_rc_eq );
1814
1815     x264_macroblock_cache_end( h );
1816     x264_free( h->out.p_bitstream );
1817     for( i = 1; i < h->param.i_threads; i++ )
1818         x264_free( h->thread[i] );
1819     x264_free( h );
1820 }
1821