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.i_rc_method < 0 || h->param.rc.i_rc_method > 2 )
 345     {
 346         x264_log( h, X264_LOG_ERROR, "invalid RC method\n" );
 347         return -1;
 348     }
 349     h->param.rc.i_rf_constant = x264_clip3( h->param.rc.i_rf_constant, 0, 51 );
 350     h->param.rc.i_qp_constant = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
 351     if( h->param.rc.i_rc_method == X264_RC_CRF )
 352         h->param.rc.i_qp_constant = h->param.rc.i_rf_constant;
 353     if( (h->param.rc.i_rc_method == X264_RC_CQP || h->param.rc.i_rc_method == X264_RC_CRF)
 354         && h->param.rc.i_qp_constant == 0 )
 355     {
 356         h->mb.b_lossless = 1;
 357         h->param.i_cqm_preset = X264_CQM_FLAT;
 358         h->param.psz_cqm_file = NULL;
 359         h->param.rc.i_rc_method = X264_RC_CQP;
 360         h->param.rc.f_ip_factor = 1;
 361         h->param.rc.f_pb_factor = 1;
 362         h->param.analyse.b_transform_8x8 = 0;
 363         h->param.analyse.b_psnr = 0;
 364         h->param.analyse.i_chroma_qp_offset = 0;
 365         h->param.analyse.i_trellis = 0;
 366         h->param.analyse.b_fast_pskip = 0;
 367         h->param.analyse.i_noise_reduction = 0;
 368         h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 1, 6 );
 369     }
 370
 371     if( ( h->param.i_width % 16 || h->param.i_height % 16 ) && !h->mb.b_lossless )
 372     {
 373         x264_log( h, X264_LOG_WARNING,
 374                   "width or height not divisible by 16 (%dx%d), compression will suffer.\n",
 375                   h->param.i_width, h->param.i_height );
 376     }
 377
 378     h->param.i_frame_reference = x264_clip3( h->param.i_frame_reference, 1, 16 );
 379     if( h->param.i_keyint_max <= 0 )
 380         h->param.i_keyint_max = 1;
 381     h->param.i_keyint_min = x264_clip3( h->param.i_keyint_min, 1, h->param.i_keyint_max/2+1 );
 382
 383     h->param.i_bframe = x264_clip3( h->param.i_bframe, 0, X264_BFRAME_MAX );
 384     h->param.i_bframe_bias = x264_clip3( h->param.i_bframe_bias, -90, 100 );
 385     h->param.b_bframe_pyramid = h->param.b_bframe_pyramid && h->param.i_bframe > 1;
 386     h->param.b_bframe_adaptive = h->param.b_bframe_adaptive && h->param.i_bframe > 0;
 387     h->param.analyse.b_weighted_bipred = h->param.analyse.b_weighted_bipred && h->param.i_bframe > 0;
 388     h->mb.b_direct_auto_write = h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
 389                                 && h->param.i_bframe
 390                                 && ( h->param.rc.b_stat_write || !h->param.rc.b_stat_read );
 391
 392     h->param.i_deblocking_filter_alphac0 = x264_clip3( h->param.i_deblocking_filter_alphac0, -6, 6 );
 393     h->param.i_deblocking_filter_beta    = x264_clip3( h->param.i_deblocking_filter_beta, -6, 6 );
 394
 395     h->param.i_cabac_init_idc = x264_clip3( h->param.i_cabac_init_idc, 0, 2 );
 396
 397     if( h->param.i_cqm_preset < X264_CQM_FLAT || h->param.i_cqm_preset > X264_CQM_CUSTOM )
 398         h->param.i_cqm_preset = X264_CQM_FLAT;
 399
 400     if( h->param.analyse.i_me_method < X264_ME_DIA ||
 401         h->param.analyse.i_me_method > X264_ME_ESA )
 402         h->param.analyse.i_me_method = X264_ME_HEX;
 403     if( h->param.analyse.i_me_range < 4 )
 404         h->param.analyse.i_me_range = 4;
 405     if( h->param.analyse.i_me_range > 16 && h->param.analyse.i_me_method <= X264_ME_HEX )
 406         h->param.analyse.i_me_range = 16;
 407     h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 1, 7 );
 408     h->param.analyse.b_bframe_rdo = h->param.analyse.b_bframe_rdo && h->param.analyse.i_subpel_refine >= 6;
 409     h->param.analyse.b_mixed_references = h->param.analyse.b_mixed_references && h->param.i_frame_reference > 1;
 410     h->param.analyse.inter &= X264_ANALYSE_PSUB16x16|X264_ANALYSE_PSUB8x8|X264_ANALYSE_BSUB16x16|
 411                               X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 412     h->param.analyse.intra &= X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 413     if( !(h->param.analyse.inter & X264_ANALYSE_PSUB16x16) )
 414         h->param.analyse.inter &= ~X264_ANALYSE_PSUB8x8;
 415     if( !h->param.analyse.b_transform_8x8 )
 416     {
 417         h->param.analyse.inter &= ~X264_ANALYSE_I8x8;
 418         h->param.analyse.intra &= ~X264_ANALYSE_I8x8;
 419     }
 420     h->param.analyse.i_chroma_qp_offset = x264_clip3(h->param.analyse.i_chroma_qp_offset, -12, 12);
 421     if( !h->param.b_cabac )
 422         h->param.analyse.i_trellis = 0;
 423     h->param.analyse.i_trellis = x264_clip3( h->param.analyse.i_trellis, 0, 2 );
 424     h->param.analyse.i_noise_reduction = x264_clip3( h->param.analyse.i_noise_reduction, 0, 1<<16 );
 425
 426     {
 427         const x264_level_t *l = x264_levels;
 428         while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
 429             l++;
 430         if( l->level_idc == 0 )
 431         {
 432             x264_log( h, X264_LOG_ERROR, "invalid level_idc: %d\n", h->param.i_level_idc );
 433             return -1;
 434         }
 435         if( h->param.analyse.i_mv_range <= 0 )
 436             h->param.analyse.i_mv_range = l->mv_range;
 437         else
 438             h->param.analyse.i_mv_range = x264_clip3(h->param.analyse.i_mv_range, 32, 2048);
 439     }
 440
 441     if( h->param.rc.f_qblur < 0 )
 442         h->param.rc.f_qblur = 0;
 443     if( h->param.rc.f_complexity_blur < 0 )
 444         h->param.rc.f_complexity_blur = 0;
 445
 446     h->param.i_sps_id &= 31;
 447
 448     /* ensure the booleans are 0 or 1 so they can be used in math */
 449 #define BOOLIFY(x) h->param.x = !!h->param.x
 450     BOOLIFY( b_cabac );
 451     BOOLIFY( b_deblocking_filter );
 452     BOOLIFY( analyse.b_transform_8x8 );
 453     BOOLIFY( analyse.b_bidir_me );
 454     BOOLIFY( analyse.b_chroma_me );
 455     BOOLIFY( analyse.b_fast_pskip );
 456     BOOLIFY( rc.b_stat_write );
 457     BOOLIFY( rc.b_stat_read );
 458 #undef BOOLIFY
 459
 460     return 0;
 461 }
 462
 463 /****************************************************************************
 464  * x264_encoder_open:
 465  ****************************************************************************/
 466 x264_t *x264_encoder_open   ( x264_param_t *param )
 467 {
 468     x264_t *h = x264_malloc( sizeof( x264_t ) );
 469     int i;
 470
 471     memset( h, 0, sizeof( x264_t ) );
 472
 473     /* Create a copy of param */
 474     memcpy( &h->param, param, sizeof( x264_param_t ) );
 475
 476     if( x264_validate_parameters( h ) < 0 )
 477     {
 478         x264_free( h );
 479         return NULL;
 480     }
 481
 482     if( h->param.psz_cqm_file )
 483         if( x264_cqm_parse_file( h, h->param.psz_cqm_file ) < 0 )
 484         {
 485             x264_free( h );
 486             return NULL;
 487         }
 488
 489     if( h->param.rc.psz_stat_out )
 490         h->param.rc.psz_stat_out = strdup( h->param.rc.psz_stat_out );
 491     if( h->param.rc.psz_stat_in )
 492         h->param.rc.psz_stat_in = strdup( h->param.rc.psz_stat_in );
 493     if( h->param.rc.psz_rc_eq )
 494         h->param.rc.psz_rc_eq = strdup( h->param.rc.psz_rc_eq );
 495
 496     /* VUI */
 497     if( h->param.vui.i_sar_width > 0 && h->param.vui.i_sar_height > 0 )
 498     {
 499         int i_w = param->vui.i_sar_width;
 500         int i_h = param->vui.i_sar_height;
 501
 502         x264_reduce_fraction( &i_w, &i_h );
 503
 504         while( i_w > 65535 || i_h > 65535 )
 505         {
 506             i_w /= 2;
 507             i_h /= 2;
 508         }
 509
 510         h->param.vui.i_sar_width = 0;
 511         h->param.vui.i_sar_height = 0;
 512         if( i_w == 0 || i_h == 0 )
 513         {
 514             x264_log( h, X264_LOG_WARNING, "cannot create valid sample aspect ratio\n" );
 515         }
 516         else
 517         {
 518             x264_log( h, X264_LOG_INFO, "using SAR=%d/%d\n", i_w, i_h );
 519             h->param.vui.i_sar_width = i_w;
 520             h->param.vui.i_sar_height = i_h;
 521         }
 522     }
 523
 524     x264_reduce_fraction( &h->param.i_fps_num, &h->param.i_fps_den );
 525
 526     /* Init x264_t */
 527     h->out.i_nal = 0;
 528     h->out.i_bitstream = X264_MAX( 1000000, h->param.i_width * h->param.i_height * 1.7
 529         * ( h->param.rc.i_rc_method == X264_RC_ABR ? pow( 0.5, h->param.rc.i_qp_min )
 530           : pow( 0.5, h->param.rc.i_qp_constant ) * X264_MAX( 1, h->param.rc.f_ip_factor )));
 531     h->out.p_bitstream = x264_malloc( h->out.i_bitstream );
 532
 533     h->i_frame = 0;
 534     h->i_frame_num = 0;
 535     h->i_idr_pic_id = 0;
 536
 537     h->sps = &h->sps_array[0];
 538     x264_sps_init( h->sps, h->param.i_sps_id, &h->param );
 539
 540     h->pps = &h->pps_array[0];
 541     x264_pps_init( h->pps, h->param.i_sps_id, &h->param, h->sps);
 542
 543     x264_validate_levels( h );
 544
 545     x264_cqm_init( h );
 546
 547     h->mb.i_mb_count = h->sps->i_mb_width * h->sps->i_mb_height;
 548
 549     /* Init frames. */
 550     h->frames.i_delay = h->param.i_bframe;
 551     h->frames.i_max_ref0 = h->param.i_frame_reference;
 552     h->frames.i_max_ref1 = h->sps->vui.i_num_reorder_frames;
 553     h->frames.i_max_dpb  = h->sps->vui.i_max_dec_frame_buffering + 1;
 554     h->frames.b_have_lowres = !h->param.rc.b_stat_read
 555         && ( h->param.rc.i_rc_method == X264_RC_ABR
 556           || h->param.rc.i_rc_method == X264_RC_CRF
 557           || h->param.b_bframe_adaptive );
 558
 559     for( i = 0; i < X264_BFRAME_MAX + 3; i++ )
 560     {
 561         h->frames.current[i] = NULL;
 562         h->frames.next[i]    = NULL;
 563         h->frames.unused[i]  = NULL;
 564     }
 565     for( i = 0; i < 1 + h->frames.i_delay; i++ )
 566     {
 567         h->frames.unused[i] =  x264_frame_new( h );
 568         if( !h->frames.unused[i] )
 569             return NULL;
 570     }
 571     for( i = 0; i < h->frames.i_max_dpb; i++ )
 572     {
 573         h->frames.reference[i] = x264_frame_new( h );
 574         if( !h->frames.reference[i] )
 575             return NULL;
 576     }
 577     h->frames.reference[h->frames.i_max_dpb] = NULL;
 578     h->frames.i_last_idr = - h->param.i_keyint_max;
 579     h->frames.i_input    = 0;
 580     h->frames.last_nonb  = NULL;
 581
 582     h->i_ref0 = 0;
 583     h->i_ref1 = 0;
 584
 585     h->fdec = h->frames.reference[0];
 586
 587     if( x264_macroblock_cache_init( h ) < 0 )
 588         return NULL;
 589     x264_rdo_init( );
 590
 591     /* init CPU functions */
 592     x264_predict_16x16_init( h->param.cpu, h->predict_16x16 );
 593     x264_predict_8x8c_init( h->param.cpu, h->predict_8x8c );
 594     x264_predict_8x8_init( h->param.cpu, h->predict_8x8 );
 595     x264_predict_4x4_init( h->param.cpu, h->predict_4x4 );
 596
 597     x264_pixel_init( h->param.cpu, &h->pixf );
 598     x264_dct_init( h->param.cpu, &h->dctf );
 599     x264_mc_init( h->param.cpu, &h->mc );
 600     x264_csp_init( h->param.cpu, h->param.i_csp, &h->csp );
 601     x264_quant_init( h, h->param.cpu, &h->quantf );
 602     x264_deblock_init( h->param.cpu, &h->loopf );
 603
 604     memcpy( h->pixf.mbcmp,
 605             ( h->mb.b_lossless || h->param.analyse.i_subpel_refine <= 1 ) ? h->pixf.sad : h->pixf.satd,
 606             sizeof(h->pixf.mbcmp) );
 607
 608     /* rate control */
 609     if( x264_ratecontrol_new( h ) < 0 )
 610         return NULL;
 611
 612     x264_log( h, X264_LOG_INFO, "using cpu capabilities %s%s%s%s%s%s\n",
 613              param->cpu&X264_CPU_MMX ? "MMX " : "",
 614              param->cpu&X264_CPU_MMXEXT ? "MMXEXT " : "",
 615              param->cpu&X264_CPU_SSE ? "SSE " : "",
 616              param->cpu&X264_CPU_SSE2 ? "SSE2 " : "",
 617              param->cpu&X264_CPU_3DNOW ? "3DNow! " : "",
 618              param->cpu&X264_CPU_ALTIVEC ? "Altivec " : "" );
 619
 620     h->thread[0] = h;
 621     h->i_thread_num = 0;
 622     for( i = 1; i < param->i_threads; i++ )
 623         h->thread[i] = x264_malloc( sizeof(x264_t) );
 624
 625 #ifdef DEBUG_DUMP_FRAME
 626     {
 627         /* create or truncate the reconstructed video file */
 628         FILE *f = fopen( "fdec.yuv", "w" );
 629         if( f )
 630             fclose( f );
 631         else
 632         {
 633             x264_log( h, X264_LOG_ERROR, "can't write to fdec.yuv\n" );
 634             x264_free( h );
 635             return NULL;
 636         }
 637     }
 638 #endif
 639
 640     return h;
 641 }
 642
 643 /****************************************************************************
 644  * x264_encoder_reconfig:
 645  ****************************************************************************/
 646 int x264_encoder_reconfig( x264_t *h, x264_param_t *param )
 647 {
 648     h->param.i_bframe_bias = param->i_bframe_bias;
 649     h->param.i_deblocking_filter_alphac0 = param->i_deblocking_filter_alphac0;
 650     h->param.i_deblocking_filter_beta    = param->i_deblocking_filter_beta;
 651     h->param.analyse.i_me_method = param->analyse.i_me_method;
 652     h->param.analyse.i_me_range = param->analyse.i_me_range;
 653     h->param.analyse.i_subpel_refine = param->analyse.i_subpel_refine;
 654     h->param.analyse.i_trellis = param->analyse.i_trellis;
 655     h->param.analyse.intra = param->analyse.intra;
 656     h->param.analyse.inter = param->analyse.inter;
 657
 658     memcpy( h->pixf.mbcmp,
 659             ( h->mb.b_lossless || h->param.analyse.i_subpel_refine <= 1 ) ? h->pixf.sad : h->pixf.satd,
 660             sizeof(h->pixf.mbcmp) );
 661
 662     return x264_validate_parameters( h );
 663 }
 664
 665 /* internal usage */
 666 static void x264_nal_start( x264_t *h, int i_type, int i_ref_idc )
 667 {
 668     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 669
 670     nal->i_ref_idc = i_ref_idc;
 671     nal->i_type    = i_type;
 672
 673     bs_align_0( &h->out.bs );   /* not needed */
 674
 675     nal->i_payload= 0;
 676     nal->p_payload= &h->out.p_bitstream[bs_pos( &h->out.bs) / 8];
 677 }
 678 static void x264_nal_end( x264_t *h )
 679 {
 680     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 681
 682     bs_align_0( &h->out.bs );   /* not needed */
 683
 684     nal->i_payload = &h->out.p_bitstream[bs_pos( &h->out.bs)/8] - nal->p_payload;
 685
 686     h->out.i_nal++;
 687 }
 688
 689 /****************************************************************************
 690  * x264_encoder_headers:
 691  ****************************************************************************/
 692 int x264_encoder_headers( x264_t *h, x264_nal_t **pp_nal, int *pi_nal )
 693 {
 694     /* init bitstream context */
 695     h->out.i_nal = 0;
 696     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
 697
 698     /* Put SPS and PPS */
 699     if( h->i_frame == 0 )
 700     {
 701         /* identify ourself */
 702         x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
 703         x264_sei_version_write( h, &h->out.bs );
 704         x264_nal_end( h );
 705
 706         /* generate sequence parameters */
 707         x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
 708         x264_sps_write( &h->out.bs, h->sps );
 709         x264_nal_end( h );
 710
 711         /* generate picture parameters */
 712         x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
 713         x264_pps_write( &h->out.bs, h->pps );
 714         x264_nal_end( h );
 715     }
 716     /* now set output*/
 717     *pi_nal = h->out.i_nal;
 718     *pp_nal = &h->out.nal[0];
 719
 720     return 0;
 721 }
 722
 723
 724 static void x264_frame_put( x264_frame_t *list[X264_BFRAME_MAX], x264_frame_t *frame )
 725 {
 726     int i = 0;
 727     while( list[i] ) i++;
 728     list[i] = frame;
 729 }
 730
 731 static void x264_frame_push( x264_frame_t *list[X264_BFRAME_MAX], x264_frame_t *frame )
 732 {
 733     int i = 0;
 734     while( list[i] ) i++;
 735     while( i-- )
 736         list[i+1] = list[i];
 737     list[0] = frame;
 738 }
 739
 740 static x264_frame_t *x264_frame_get( x264_frame_t *list[X264_BFRAME_MAX+1] )
 741 {
 742     x264_frame_t *frame = list[0];
 743     int i;
 744     for( i = 0; list[i]; i++ )
 745         list[i] = list[i+1];
 746     return frame;
 747 }
 748
 749 static void x264_frame_sort( x264_frame_t *list[X264_BFRAME_MAX+1], int b_dts )
 750 {
 751     int i, b_ok;
 752     do {
 753         b_ok = 1;
 754         for( i = 0; list[i+1]; i++ )
 755         {
 756             int dtype = list[i]->i_type - list[i+1]->i_type;
 757             int dtime = list[i]->i_frame - list[i+1]->i_frame;
 758             int swap = b_dts ? dtype > 0 || ( dtype == 0 && dtime > 0 )
 759                              : dtime > 0;
 760             if( swap )
 761             {
 762                 XCHG( x264_frame_t*, list[i], list[i+1] );
 763                 b_ok = 0;
 764             }
 765         }
 766     } while( !b_ok );
 767 }
 768 #define x264_frame_sort_dts(list) x264_frame_sort(list, 1)
 769 #define x264_frame_sort_pts(list) x264_frame_sort(list, 0)
 770
 771 static inline void x264_reference_build_list( x264_t *h, int i_poc, int i_slice_type )
 772 {
 773     int i;
 774     int b_ok;
 775
 776     /* build ref list 0/1 */
 777     h->i_ref0 = 0;
 778     h->i_ref1 = 0;
 779     for( i = 1; i < h->frames.i_max_dpb; i++ )
 780     {
 781         if( h->frames.reference[i]->i_poc >= 0 )
 782         {
 783             if( h->frames.reference[i]->i_poc < i_poc )
 784             {
 785                 h->fref0[h->i_ref0++] = h->frames.reference[i];
 786             }
 787             else if( h->frames.reference[i]->i_poc > i_poc )
 788             {
 789                 h->fref1[h->i_ref1++] = h->frames.reference[i];
 790             }
 791         }
 792     }
 793
 794     /* Order ref0 from higher to lower poc */
 795     do
 796     {
 797         b_ok = 1;
 798         for( i = 0; i < h->i_ref0 - 1; i++ )
 799         {
 800             if( h->fref0[i]->i_poc < h->fref0[i+1]->i_poc )
 801             {
 802                 XCHG( x264_frame_t*, h->fref0[i], h->fref0[i+1] );
 803                 b_ok = 0;
 804                 break;
 805             }
 806         }
 807     } while( !b_ok );
 808     /* Order ref1 from lower to higher poc (bubble sort) for B-frame */
 809     do
 810     {
 811         b_ok = 1;
 812         for( i = 0; i < h->i_ref1 - 1; i++ )
 813         {
 814             if( h->fref1[i]->i_poc > h->fref1[i+1]->i_poc )
 815             {
 816                 XCHG( x264_frame_t*, h->fref1[i], h->fref1[i+1] );
 817                 b_ok = 0;
 818                 break;
 819             }
 820         }
 821     } while( !b_ok );
 822
 823     /* In the standard, a P-frame's ref list is sorted by frame_num.
 824      * We use POC, but check whether explicit reordering is needed */
 825     h->b_ref_reorder[0] =
 826     h->b_ref_reorder[1] = 0;
 827     if( i_slice_type == SLICE_TYPE_P )
 828     {
 829         for( i = 0; i < h->i_ref0 - 1; i++ )
 830             if( h->fref0[i]->i_frame_num < h->fref0[i+1]->i_frame_num )
 831             {
 832                 h->b_ref_reorder[0] = 1;
 833                 break;
 834             }
 835     }
 836
 837     h->i_ref1 = X264_MIN( h->i_ref1, h->frames.i_max_ref1 );
 838     h->i_ref0 = X264_MIN( h->i_ref0, h->frames.i_max_ref0 );
 839     h->i_ref0 = X264_MIN( h->i_ref0, 16 - h->i_ref1 );
 840 }
 841
 842 static inline void x264_fdec_deblock( x264_t *h )
 843 {
 844     /* apply deblocking filter to the current decoded picture */
 845     if( !h->sh.i_disable_deblocking_filter_idc )
 846     {
 847         TIMER_START( i_mtime_filter );
 848         x264_frame_deblocking_filter( h, h->sh.i_type );
 849         TIMER_STOP( i_mtime_filter );
 850     }
 851 }
 852
 853 static inline void x264_reference_update( x264_t *h )
 854 {
 855     int i;
 856
 857     x264_fdec_deblock( h );
 858
 859     /* expand border */
 860     x264_frame_expand_border( h->fdec );
 861
 862     /* create filtered images */
 863     x264_frame_filter( h->param.cpu, h->fdec );
 864
 865     /* expand border of filtered images */
 866     x264_frame_expand_border_filtered( h->fdec );
 867
 868     /* move lowres copy of the image to the ref frame */
 869     for( i = 0; i < 4; i++)
 870         XCHG( uint8_t*, h->fdec->lowres[i], h->fenc->lowres[i] );
 871
 872     /* adaptive B decision needs a pointer, since it can't use the ref lists */
 873     if( h->sh.i_type != SLICE_TYPE_B )
 874         h->frames.last_nonb = h->fdec;
 875
 876     /* move frame in the buffer */
 877     h->fdec = h->frames.reference[h->frames.i_max_dpb-1];
 878     for( i = h->frames.i_max_dpb-1; i > 0; i-- )
 879     {
 880         h->frames.reference[i] = h->frames.reference[i-1];
 881     }
 882     h->frames.reference[0] = h->fdec;
 883 }
 884
 885 static inline void x264_reference_reset( x264_t *h )
 886 {
 887     int i;
 888
 889     /* reset ref pictures */
 890     for( i = 1; i < h->frames.i_max_dpb; i++ )
 891     {
 892         h->frames.reference[i]->i_poc = -1;
 893     }
 894     h->frames.reference[0]->i_poc = 0;
 895 }
 896
 897 static inline void x264_slice_init( x264_t *h, int i_nal_type, int i_slice_type, int i_global_qp )
 898 {
 899     /* ------------------------ Create slice header  ----------------------- */
 900     if( i_nal_type == NAL_SLICE_IDR )
 901     {
 902         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 );
 903
 904         /* increment id */
 905         h->i_idr_pic_id = ( h->i_idr_pic_id + 1 ) % 65536;
 906     }
 907     else
 908     {
 909         x264_slice_header_init( h, &h->sh, h->sps, h->pps, i_slice_type, -1, h->i_frame_num, i_global_qp );
 910
 911         /* always set the real higher num of ref frame used */
 912         h->sh.b_num_ref_idx_override = 1;
 913         h->sh.i_num_ref_idx_l0_active = h->i_ref0 <= 0 ? 1 : h->i_ref0;
 914         h->sh.i_num_ref_idx_l1_active = h->i_ref1 <= 0 ? 1 : h->i_ref1;
 915     }
 916
 917     h->fdec->i_frame_num = h->sh.i_frame_num;
 918
 919     if( h->sps->i_poc_type == 0 )
 920     {
 921         h->sh.i_poc_lsb = h->fdec->i_poc & ( (1 << h->sps->i_log2_max_poc_lsb) - 1 );
 922         h->sh.i_delta_poc_bottom = 0;   /* XXX won't work for field */
 923     }
 924     else if( h->sps->i_poc_type == 1 )
 925     {
 926         /* FIXME TODO FIXME */
 927     }
 928     else
 929     {
 930         /* Nothing to do ? */
 931     }
 932
 933     x264_macroblock_slice_init( h );
 934 }
 935
 936 static int x264_slice_write( x264_t *h )
 937 {
 938     int i_skip;
 939     int mb_xy;
 940     int i;
 941
 942     /* init stats */
 943     memset( &h->stat.frame, 0, sizeof(h->stat.frame) );
 944
 945     /* Slice */
 946     x264_nal_start( h, h->i_nal_type, h->i_nal_ref_idc );
 947
 948     /* Slice header */
 949     x264_slice_header_write( &h->out.bs, &h->sh, h->i_nal_ref_idc );
 950     if( h->param.b_cabac )
 951     {
 952         /* alignment needed */
 953         bs_align_1( &h->out.bs );
 954
 955         /* init cabac */
 956         x264_cabac_context_init( &h->cabac, h->sh.i_type, h->sh.i_qp, h->sh.i_cabac_init_idc );
 957         x264_cabac_encode_init ( &h->cabac, &h->out.bs );
 958     }
 959     h->mb.i_last_qp = h->sh.i_qp;
 960     h->mb.i_last_dqp = 0;
 961
 962     for( mb_xy = h->sh.i_first_mb, i_skip = 0; mb_xy < h->sh.i_last_mb; mb_xy++ )
 963     {
 964         const int i_mb_y = mb_xy / h->sps->i_mb_width;
 965         const int i_mb_x = mb_xy % h->sps->i_mb_width;
 966
 967         int mb_spos = bs_pos(&h->out.bs);
 968
 969         /* load cache */
 970         x264_macroblock_cache_load( h, i_mb_x, i_mb_y );
 971
 972         /* analyse parameters
 973          * Slice I: choose I_4x4 or I_16x16 mode
 974          * Slice P: choose between using P mode or intra (4x4 or 16x16)
 975          * */
 976         TIMER_START( i_mtime_analyse );
 977         x264_macroblock_analyse( h );
 978         TIMER_STOP( i_mtime_analyse );
 979
 980         /* encode this macrobock -> be carefull it can change the mb type to P_SKIP if needed */
 981         TIMER_START( i_mtime_encode );
 982         x264_macroblock_encode( h );
 983         TIMER_STOP( i_mtime_encode );
 984
 985         TIMER_START( i_mtime_write );
 986         if( h->param.b_cabac )
 987         {
 988             if( mb_xy > h->sh.i_first_mb )
 989                 x264_cabac_encode_terminal( &h->cabac, 0 );
 990
 991             if( IS_SKIP( h->mb.i_type ) )
 992                 x264_cabac_mb_skip( h, 1 );
 993             else
 994             {
 995                 if( h->sh.i_type != SLICE_TYPE_I )
 996                     x264_cabac_mb_skip( h, 0 );
 997                 x264_macroblock_write_cabac( h, &h->cabac );
 998             }
 999         }
1000         else
1001         {
1002             if( IS_SKIP( h->mb.i_type ) )
1003                 i_skip++;
1004             else
1005             {
1006                 if( h->sh.i_type != SLICE_TYPE_I )
1007                 {
1008                     bs_write_ue( &h->out.bs, i_skip );  /* skip run */
1009                     i_skip = 0;
1010                 }
1011                 x264_macroblock_write_cavlc( h, &h->out.bs );
1012             }
1013         }
1014         TIMER_STOP( i_mtime_write );
1015
1016 #if VISUALIZE
1017         if( h->param.b_visualize )
1018             x264_visualize_mb( h );
1019 #endif
1020
1021         /* save cache */
1022         x264_macroblock_cache_save( h );
1023
1024         /* accumulate mb stats */
1025         h->stat.frame.i_mb_count[h->mb.i_type]++;
1026         if( !IS_SKIP(h->mb.i_type) && !IS_INTRA(h->mb.i_type) && !IS_DIRECT(h->mb.i_type) )
1027         {
1028             if( h->mb.i_partition != D_8x8 )
1029                 h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_partition ] ] += 4;
1030             else
1031                 for( i = 0; i < 4; i++ )
1032                     h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_sub_partition[i] ] ] ++;
1033             if( h->param.i_frame_reference > 1 )
1034             {
1035                 for( i = 0; i < 4; i++ )
1036                 {
1037                     int i_ref = h->mb.cache.ref[0][ x264_scan8[4*i] ];
1038                     if( i_ref >= 0 )
1039                         h->stat.frame.i_mb_count_ref[i_ref] ++;
1040                 }
1041             }
1042         }
1043         if( h->mb.i_cbp_luma && !IS_INTRA(h->mb.i_type) )
1044         {
1045             h->stat.frame.i_mb_count_8x8dct[0] ++;
1046             h->stat.frame.i_mb_count_8x8dct[1] += h->mb.b_transform_8x8;
1047         }
1048
1049         if( h->mb.b_variable_qp )
1050             x264_ratecontrol_mb(h, bs_pos(&h->out.bs) - mb_spos);
1051     }
1052
1053     if( h->param.b_cabac )
1054     {
1055         /* end of slice */
1056         x264_cabac_encode_terminal( &h->cabac, 1 );
1057     }
1058     else if( i_skip > 0 )
1059     {
1060         bs_write_ue( &h->out.bs, i_skip );  /* last skip run */
1061     }
1062
1063     if( h->param.b_cabac )
1064     {
1065         x264_cabac_encode_flush( &h->cabac );
1066
1067     }
1068     else
1069     {
1070         /* rbsp_slice_trailing_bits */
1071         bs_rbsp_trailing( &h->out.bs );
1072     }
1073
1074     x264_nal_end( h );
1075
1076     /* Compute misc bits */
1077     h->stat.frame.i_misc_bits = bs_pos( &h->out.bs )
1078                               + NALU_OVERHEAD * 8
1079                               - h->stat.frame.i_itex_bits
1080                               - h->stat.frame.i_ptex_bits
1081                               - h->stat.frame.i_hdr_bits;
1082
1083     return 0;
1084 }
1085
1086 static inline int x264_slices_write( x264_t *h )
1087 {
1088     int i_frame_size;
1089
1090 #if VISUALIZE
1091     if( h->param.b_visualize )
1092         x264_visualize_init( h );
1093 #endif
1094
1095     if( h->param.i_threads == 1 )
1096     {
1097         x264_ratecontrol_threads_start( h );
1098         x264_slice_write( h );
1099         i_frame_size = h->out.nal[h->out.i_nal-1].i_payload;
1100     }
1101     else
1102     {
1103         int i_nal = h->out.i_nal;
1104         int i_bs_size = h->out.i_bitstream / h->param.i_threads;
1105         int i;
1106         /* duplicate contexts */
1107         for( i = 0; i < h->param.i_threads; i++ )
1108         {
1109             x264_t *t = h->thread[i];
1110             if( i > 0 )
1111             {
1112                 memcpy( t, h, sizeof(x264_t) );
1113                 t->out.p_bitstream += i*i_bs_size;
1114                 bs_init( &t->out.bs, t->out.p_bitstream, i_bs_size );
1115                 t->i_thread_num = i;
1116             }
1117             t->sh.i_first_mb = (i    * h->sps->i_mb_height / h->param.i_threads) * h->sps->i_mb_width;
1118             t->sh.i_last_mb = ((i+1) * h->sps->i_mb_height / h->param.i_threads) * h->sps->i_mb_width;
1119             t->out.i_nal = i_nal + i;
1120         }
1121         x264_ratecontrol_threads_start( h );
1122
1123         /* dispatch */
1124 #ifdef HAVE_PTHREAD
1125         {
1126             pthread_t handles[X264_SLICE_MAX];
1127             for( i = 0; i < h->param.i_threads; i++ )
1128                 pthread_create( &handles[i], NULL, (void*)x264_slice_write, (void*)h->thread[i] );
1129             for( i = 0; i < h->param.i_threads; i++ )
1130                 pthread_join( handles[i], NULL );
1131         }
1132 #else
1133         for( i = 0; i < h->param.i_threads; i++ )
1134             x264_slice_write( h->thread[i] );
1135 #endif
1136
1137         /* merge contexts */
1138         i_frame_size = h->out.nal[i_nal].i_payload;
1139         for( i = 1; i < h->param.i_threads; i++ )
1140         {
1141             int j;
1142             x264_t *t = h->thread[i];
1143             h->out.nal[i_nal+i] = t->out.nal[i_nal+i];
1144             i_frame_size += t->out.nal[i_nal+i].i_payload;
1145             // all entries in stat.frame are ints
1146             for( j = 0; j < sizeof(h->stat.frame) / sizeof(int); j++ )
1147                 ((int*)&h->stat.frame)[j] += ((int*)&t->stat.frame)[j];
1148         }
1149         h->out.i_nal = i_nal + h->param.i_threads;
1150     }
1151
1152 #if VISUALIZE
1153     if( h->param.b_visualize )
1154     {
1155         x264_visualize_show( h );
1156         x264_visualize_close( h );
1157     }
1158 #endif
1159
1160     return i_frame_size;
1161 }
1162
1163 /****************************************************************************
1164  * x264_encoder_encode:
1165  *  XXX: i_poc   : is the poc of the current given picture
1166  *       i_frame : is the number of the frame being coded
1167  *  ex:  type frame poc
1168  *       I      0   2*0
1169  *       P      1   2*3
1170  *       B      2   2*1
1171  *       B      3   2*2
1172  *       P      4   2*6
1173  *       B      5   2*4
1174  *       B      6   2*5
1175  ****************************************************************************/
1176 int     x264_encoder_encode( x264_t *h,
1177                              x264_nal_t **pp_nal, int *pi_nal,
1178                              x264_picture_t *pic_in,
1179                              x264_picture_t *pic_out )
1180 {
1181     x264_frame_t   *frame_psnr = h->fdec; /* just to keep the current decoded frame for psnr calculation */
1182     int     i_nal_type;
1183     int     i_nal_ref_idc;
1184     int     i_slice_type;
1185     int     i_frame_size;
1186
1187     int i;
1188
1189     int   i_global_qp;
1190
1191     char psz_message[80];
1192
1193     /* no data out */
1194     *pi_nal = 0;
1195     *pp_nal = NULL;
1196
1197
1198     /* ------------------- Setup new frame from picture -------------------- */
1199     TIMER_START( i_mtime_encode_frame );
1200     if( pic_in != NULL )
1201     {
1202         /* 1: Copy the picture to a frame and move it to a buffer */
1203         x264_frame_t *fenc = x264_frame_get( h->frames.unused );
1204
1205         x264_frame_copy_picture( h, fenc, pic_in );
1206
1207         if( h->param.i_width % 16 || h->param.i_height % 16 )
1208             x264_frame_expand_border_mod16( h, fenc );
1209
1210         fenc->i_frame = h->frames.i_input++;
1211
1212         x264_frame_put( h->frames.next, fenc );
1213
1214         if( h->frames.b_have_lowres )
1215             x264_frame_init_lowres( h->param.cpu, fenc );
1216
1217         if( h->frames.i_input <= h->frames.i_delay )
1218         {
1219             /* Nothing yet to encode */
1220             /* waiting for filling bframe buffer */
1221             pic_out->i_type = X264_TYPE_AUTO;
1222             return 0;
1223         }
1224     }
1225
1226     if( h->frames.current[0] == NULL )
1227     {
1228         int bframes = 0;
1229         /* 2: Select frame types */
1230         if( h->frames.next[0] == NULL )
1231             return 0;
1232
1233         x264_slicetype_decide( h );
1234
1235         /* 3: move some B-frames and 1 non-B to encode queue */
1236         while( IS_X264_TYPE_B( h->frames.next[bframes]->i_type ) )
1237             bframes++;
1238         x264_frame_put( h->frames.current, x264_frame_get( &h->frames.next[bframes] ) );
1239         /* FIXME: when max B-frames > 3, BREF may no longer be centered after GOP closing */
1240         if( h->param.b_bframe_pyramid && bframes > 1 )
1241         {
1242             x264_frame_t *mid = x264_frame_get( &h->frames.next[bframes/2] );
1243             mid->i_type = X264_TYPE_BREF;
1244             x264_frame_put( h->frames.current, mid );
1245             bframes--;
1246         }
1247         while( bframes-- )
1248             x264_frame_put( h->frames.current, x264_frame_get( h->frames.next ) );
1249     }
1250     TIMER_STOP( i_mtime_encode_frame );
1251
1252     /* ------------------- Get frame to be encoded ------------------------- */
1253     /* 4: get picture to encode */
1254     h->fenc = x264_frame_get( h->frames.current );
1255     if( h->fenc == NULL )
1256     {
1257         /* Nothing yet to encode (ex: waiting for I/P with B frames) */
1258         /* waiting for filling bframe buffer */
1259         pic_out->i_type = X264_TYPE_AUTO;
1260         return 0;
1261     }
1262
1263 do_encode:
1264
1265     if( h->fenc->i_type == X264_TYPE_IDR )
1266     {
1267         h->frames.i_last_idr = h->fenc->i_frame;
1268     }
1269
1270     /* ------------------- Setup frame context ----------------------------- */
1271     /* 5: Init data dependant of frame type */
1272     TIMER_START( i_mtime_encode_frame );
1273     if( h->fenc->i_type == X264_TYPE_IDR )
1274     {
1275         /* reset ref pictures */
1276         x264_reference_reset( h );
1277
1278         i_nal_type    = NAL_SLICE_IDR;
1279         i_nal_ref_idc = NAL_PRIORITY_HIGHEST;
1280         i_slice_type = SLICE_TYPE_I;
1281     }
1282     else if( h->fenc->i_type == X264_TYPE_I )
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_I;
1287     }
1288     else if( h->fenc->i_type == X264_TYPE_P )
1289     {
1290         i_nal_type    = NAL_SLICE;
1291         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
1292         i_slice_type = SLICE_TYPE_P;
1293     }
1294     else if( h->fenc->i_type == X264_TYPE_BREF )
1295     {
1296         i_nal_type    = NAL_SLICE;
1297         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* maybe add MMCO to forget it? -> low */
1298         i_slice_type = SLICE_TYPE_B;
1299     }
1300     else    /* B frame */
1301     {
1302         i_nal_type    = NAL_SLICE;
1303         i_nal_ref_idc = NAL_PRIORITY_DISPOSABLE;
1304         i_slice_type = SLICE_TYPE_B;
1305     }
1306
1307     h->fdec->i_poc =
1308     h->fenc->i_poc = 2 * (h->fenc->i_frame - h->frames.i_last_idr);
1309     h->fdec->i_type = h->fenc->i_type;
1310     h->fdec->i_frame = h->fenc->i_frame;
1311     h->fenc->b_kept_as_ref =
1312     h->fdec->b_kept_as_ref = i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE;
1313
1314
1315
1316     /* ------------------- Init                ----------------------------- */
1317     /* build ref list 0/1 */
1318     x264_reference_build_list( h, h->fdec->i_poc, i_slice_type );
1319
1320     /* Init the rate control */
1321     x264_ratecontrol_start( h, i_slice_type, h->fenc->i_qpplus1 );
1322     i_global_qp = x264_ratecontrol_qp( h );
1323
1324     pic_out->i_qpplus1 =
1325     h->fdec->i_qpplus1 = i_global_qp + 1;
1326
1327     if( i_slice_type == SLICE_TYPE_B )
1328         x264_macroblock_bipred_init( h );
1329
1330     /* ------------------------ Create slice header  ----------------------- */
1331     x264_slice_init( h, i_nal_type, i_slice_type, i_global_qp );
1332
1333     if( h->fenc->b_kept_as_ref )
1334         h->i_frame_num++;
1335
1336     /* ---------------------- Write the bitstream -------------------------- */
1337     /* Init bitstream context */
1338     h->out.i_nal = 0;
1339     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
1340
1341     if(h->param.b_aud){
1342         int pic_type;
1343
1344         if(i_slice_type == SLICE_TYPE_I)
1345             pic_type = 0;
1346         else if(i_slice_type == SLICE_TYPE_P)
1347             pic_type = 1;
1348         else if(i_slice_type == SLICE_TYPE_B)
1349             pic_type = 2;
1350         else
1351             pic_type = 7;
1352
1353         x264_nal_start(h, NAL_AUD, NAL_PRIORITY_DISPOSABLE);
1354         bs_write(&h->out.bs, 3, pic_type);
1355         bs_rbsp_trailing(&h->out.bs);
1356         x264_nal_end(h);
1357     }
1358
1359     h->i_nal_type = i_nal_type;
1360     h->i_nal_ref_idc = i_nal_ref_idc;
1361
1362     /* Write SPS and PPS */
1363     if( i_nal_type == NAL_SLICE_IDR && h->param.b_repeat_headers )
1364     {
1365         if( h->fenc->i_frame == 0 )
1366         {
1367             /* identify ourself */
1368             x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
1369             x264_sei_version_write( h, &h->out.bs );
1370             x264_nal_end( h );
1371         }
1372
1373         /* generate sequence parameters */
1374         x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
1375         x264_sps_write( &h->out.bs, h->sps );
1376         x264_nal_end( h );
1377
1378         /* generate picture parameters */
1379         x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
1380         x264_pps_write( &h->out.bs, h->pps );
1381         x264_nal_end( h );
1382     }
1383
1384     /* Write frame */
1385     i_frame_size = x264_slices_write( h );
1386
1387     /* restore CPU state (before using float again) */
1388     x264_cpu_restore( h->param.cpu );
1389
1390     if( i_slice_type == SLICE_TYPE_P && !h->param.rc.b_stat_read
1391         && h->param.i_scenecut_threshold >= 0 )
1392     {
1393         const int *mbs = h->stat.frame.i_mb_count;
1394         int i_mb_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1395         int i_mb_p = mbs[P_L0] + mbs[P_8x8];
1396         int i_mb_s = mbs[P_SKIP];
1397         int i_mb   = h->sps->i_mb_width * h->sps->i_mb_height;
1398         int64_t i_inter_cost = h->stat.frame.i_inter_cost;
1399         int64_t i_intra_cost = h->stat.frame.i_intra_cost;
1400
1401         float f_bias;
1402         int i_gop_size = h->fenc->i_frame - h->frames.i_last_idr;
1403         float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
1404         /* magic numbers pulled out of thin air */
1405         float f_thresh_min = f_thresh_max * h->param.i_keyint_min
1406                              / ( h->param.i_keyint_max * 4 );
1407         if( h->param.i_keyint_min == h->param.i_keyint_max )
1408              f_thresh_min= f_thresh_max;
1409
1410         /* macroblock_analyse() doesn't further analyse skipped mbs,
1411          * so we have to guess their cost */
1412         if( i_mb_s < i_mb )
1413             i_intra_cost = i_intra_cost * i_mb / (i_mb - i_mb_s);
1414
1415         if( i_gop_size < h->param.i_keyint_min / 4 )
1416             f_bias = f_thresh_min / 4;
1417         else if( i_gop_size <= h->param.i_keyint_min )
1418             f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
1419         else
1420         {
1421             f_bias = f_thresh_min
1422                      + ( f_thresh_max - f_thresh_min )
1423                        * ( i_gop_size - h->param.i_keyint_min )
1424                        / ( h->param.i_keyint_max - h->param.i_keyint_min );
1425         }
1426         f_bias = X264_MIN( f_bias, 1.0 );
1427
1428         /* Bad P will be reencoded as I */
1429         if( i_mb_s < i_mb &&
1430             i_inter_cost >= (1.0 - f_bias) * i_intra_cost )
1431         {
1432             int b;
1433
1434             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",
1435                       h->fenc->i_frame,
1436                       (double)i_intra_cost, (double)i_inter_cost,
1437                       (double)i_inter_cost / i_intra_cost,
1438                       f_bias, i_gop_size,
1439                       i_mb_i, i_mb_p, i_mb_s );
1440
1441             /* Restore frame num */
1442             h->i_frame_num--;
1443
1444             for( b = 0; h->frames.current[b] && IS_X264_TYPE_B( h->frames.current[b]->i_type ); b++ );
1445             if( b > 0 )
1446             {
1447                 /* If using B-frames, force GOP to be closed.
1448                  * Even if this frame is going to be I and not IDR, forcing a
1449                  * P-frame before the scenecut will probably help compression.
1450                  *
1451                  * We don't yet know exactly which frame is the scene cut, so
1452                  * we can't assign an I-frame. Instead, change the previous
1453                  * B-frame to P, and rearrange coding order. */
1454
1455                 if( h->param.b_bframe_adaptive || b > 1 )
1456                     h->fenc->i_type = X264_TYPE_AUTO;
1457                 x264_frame_sort_pts( h->frames.current );
1458                 x264_frame_push( h->frames.next, h->fenc );
1459                 h->fenc = h->frames.current[b-1];
1460                 h->frames.current[b-1] = NULL;
1461                 h->fenc->i_type = X264_TYPE_P;
1462                 x264_frame_sort_dts( h->frames.current );
1463             }
1464             /* Do IDR if needed */
1465             else if( i_gop_size >= h->param.i_keyint_min )
1466             {
1467                 x264_frame_t *tmp;
1468
1469                 /* Reset */
1470                 h->i_frame_num = 0;
1471
1472                 /* Reinit field of fenc */
1473                 h->fenc->i_type = X264_TYPE_IDR;
1474                 h->fenc->i_poc = 0;
1475
1476                 /* Put enqueued frames back in the pool */
1477                 while( (tmp = x264_frame_get( h->frames.current ) ) != NULL )
1478                     x264_frame_put( h->frames.next, tmp );
1479                 x264_frame_sort_pts( h->frames.next );
1480             }
1481             else
1482             {
1483                 h->fenc->i_type = X264_TYPE_I;
1484             }
1485             goto do_encode;
1486         }
1487     }
1488
1489     /* End bitstream, set output  */
1490     *pi_nal = h->out.i_nal;
1491     *pp_nal = h->out.nal;
1492
1493     /* Set output picture properties */
1494     if( i_slice_type == SLICE_TYPE_I )
1495         pic_out->i_type = i_nal_type == NAL_SLICE_IDR ? X264_TYPE_IDR : X264_TYPE_I;
1496     else if( i_slice_type == SLICE_TYPE_P )
1497         pic_out->i_type = X264_TYPE_P;
1498     else
1499         pic_out->i_type = X264_TYPE_B;
1500     pic_out->i_pts = h->fenc->i_pts;
1501
1502     pic_out->img.i_plane = h->fdec->i_plane;
1503     for(i = 0; i < 4; i++){
1504         pic_out->img.i_stride[i] = h->fdec->i_stride[i];
1505         pic_out->img.plane[i] = h->fdec->plane[i];
1506     }
1507
1508     /* ---------------------- Update encoder state ------------------------- */
1509
1510     /* update rc */
1511     x264_cpu_restore( h->param.cpu );
1512     x264_ratecontrol_end( h, i_frame_size * 8 );
1513
1514     /* handle references */
1515     if( i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE )
1516         x264_reference_update( h );
1517 #ifdef DEBUG_DUMP_FRAME
1518     else
1519         x264_fdec_deblock( h );
1520 #endif
1521     x264_frame_put( h->frames.unused, h->fenc );
1522
1523     /* increase frame count */
1524     h->i_frame++;
1525
1526     /* restore CPU state (before using float again) */
1527     x264_cpu_restore( h->param.cpu );
1528
1529     x264_noise_reduction_update( h );
1530
1531     TIMER_STOP( i_mtime_encode_frame );
1532
1533     /* ---------------------- Compute/Print statistics --------------------- */
1534     /* Slice stat */
1535     h->stat.i_slice_count[i_slice_type]++;
1536     h->stat.i_slice_size[i_slice_type] += i_frame_size + NALU_OVERHEAD;
1537     h->stat.i_slice_qp[i_slice_type] += i_global_qp;
1538
1539     for( i = 0; i < 19; i++ )
1540         h->stat.i_mb_count[h->sh.i_type][i] += h->stat.frame.i_mb_count[i];
1541     for( i = 0; i < 2; i++ )
1542         h->stat.i_mb_count_8x8dct[i] += h->stat.frame.i_mb_count_8x8dct[i];
1543     if( h->sh.i_type != SLICE_TYPE_I )
1544     {
1545         for( i = 0; i < 7; i++ )
1546             h->stat.i_mb_count_size[h->sh.i_type][i] += h->stat.frame.i_mb_count_size[i];
1547         for( i = 0; i < 16; i++ )
1548             h->stat.i_mb_count_ref[h->sh.i_type][i] += h->stat.frame.i_mb_count_ref[i];
1549     }
1550     if( i_slice_type == SLICE_TYPE_B )
1551     {
1552         h->stat.i_direct_frames[ h->sh.b_direct_spatial_mv_pred ] ++;
1553         if( h->mb.b_direct_auto_write )
1554         {
1555             //FIXME somewhat arbitrary time constants
1556             if( h->stat.i_direct_score[0] + h->stat.i_direct_score[1] > h->mb.i_mb_count )
1557             {
1558                 for( i = 0; i < 2; i++ )
1559                     h->stat.i_direct_score[i] = h->stat.i_direct_score[i] * 9/10;
1560             }
1561             for( i = 0; i < 2; i++ )
1562                 h->stat.i_direct_score[i] += h->stat.frame.i_direct_score[i];
1563         }
1564     }
1565
1566     if( h->param.analyse.b_psnr )
1567     {
1568         int64_t i_sqe_y, i_sqe_u, i_sqe_v;
1569
1570         /* PSNR */
1571         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 );
1572         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);
1573         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);
1574         x264_cpu_restore( h->param.cpu );
1575
1576         h->stat.i_sqe_global[i_slice_type] += i_sqe_y + i_sqe_u + i_sqe_v;
1577         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 );
1578         h->stat.f_psnr_mean_y[i_slice_type] += x264_psnr( i_sqe_y, h->param.i_width * h->param.i_height );
1579         h->stat.f_psnr_mean_u[i_slice_type] += x264_psnr( i_sqe_u, h->param.i_width * h->param.i_height / 4 );
1580         h->stat.f_psnr_mean_v[i_slice_type] += x264_psnr( i_sqe_v, h->param.i_width * h->param.i_height / 4 );
1581
1582         snprintf( psz_message, 80, " PSNR Y:%2.2f U:%2.2f V:%2.2f",
1583                   x264_psnr( i_sqe_y, h->param.i_width * h->param.i_height ),
1584                   x264_psnr( i_sqe_u, h->param.i_width * h->param.i_height / 4),
1585                   x264_psnr( i_sqe_v, h->param.i_width * h->param.i_height / 4) );
1586         psz_message[79] = '\0';
1587     }
1588     else
1589     {
1590         psz_message[0] = '\0';
1591     }
1592
1593     x264_log( h, X264_LOG_DEBUG,
1594                   "frame=%4d QP=%i NAL=%d Slice:%c Poc:%-3d I:%-4d P:%-4d SKIP:%-4d size=%d bytes%s\n",
1595               h->i_frame - 1,
1596               i_global_qp,
1597               i_nal_ref_idc,
1598               i_slice_type == SLICE_TYPE_I ? 'I' : (i_slice_type == SLICE_TYPE_P ? 'P' : 'B' ),
1599               frame_psnr->i_poc,
1600               h->stat.frame.i_mb_count_i,
1601               h->stat.frame.i_mb_count_p,
1602               h->stat.frame.i_mb_count_skip,
1603               i_frame_size,
1604               psz_message );
1605
1606
1607 #ifdef DEBUG_MB_TYPE
1608 {
1609     static const char mb_chars[] = { 'i', 'i', 'I', 'C', 'P', '8', 'S',
1610         'D', '<', 'X', 'B', 'X', '>', 'B', 'B', 'B', 'B', '8', 'S' };
1611     int mb_xy;
1612     for( mb_xy = 0; mb_xy < h->sps->i_mb_width * h->sps->i_mb_height; mb_xy++ )
1613     {
1614         if( h->mb.type[mb_xy] < 19 && h->mb.type[mb_xy] >= 0 )
1615             fprintf( stderr, "%c ", mb_chars[ h->mb.type[mb_xy] ] );
1616         else
1617             fprintf( stderr, "? " );
1618
1619         if( (mb_xy+1) % h->sps->i_mb_width == 0 )
1620             fprintf( stderr, "\n" );
1621     }
1622 }
1623 #endif
1624
1625 #ifdef DEBUG_DUMP_FRAME
1626     /* Dump reconstructed frame */
1627     x264_frame_dump( h, frame_psnr, "fdec.yuv" );
1628 #endif
1629     return 0;
1630 }
1631
1632 /****************************************************************************
1633  * x264_encoder_close:
1634  ****************************************************************************/
1635 void    x264_encoder_close  ( x264_t *h )
1636 {
1637 #ifdef DEBUG_BENCHMARK
1638     int64_t i_mtime_total = i_mtime_analyse + i_mtime_encode + i_mtime_write + i_mtime_filter + 1;
1639 #endif
1640     int64_t i_yuv_size = 3 * h->param.i_width * h->param.i_height / 2;
1641     int i;
1642
1643 #ifdef DEBUG_BENCHMARK
1644     x264_log( h, X264_LOG_INFO,
1645               "analyse=%d(%lldms) encode=%d(%lldms) write=%d(%lldms) filter=%d(%lldms)\n",
1646               (int)(100*i_mtime_analyse/i_mtime_total), i_mtime_analyse/1000,
1647               (int)(100*i_mtime_encode/i_mtime_total), i_mtime_encode/1000,
1648               (int)(100*i_mtime_write/i_mtime_total), i_mtime_write/1000,
1649               (int)(100*i_mtime_filter/i_mtime_total), i_mtime_filter/1000 );
1650 #endif
1651
1652     /* Slices used and PSNR */
1653     for( i=0; i<5; i++ )
1654     {
1655         static const int slice_order[] = { SLICE_TYPE_I, SLICE_TYPE_SI, SLICE_TYPE_P, SLICE_TYPE_SP, SLICE_TYPE_B };
1656         static const char *slice_name[] = { "P", "B", "I", "SP", "SI" };
1657         int i_slice = slice_order[i];
1658
1659         if( h->stat.i_slice_count[i_slice] > 0 )
1660         {
1661             const int i_count = h->stat.i_slice_count[i_slice];
1662             if( h->param.analyse.b_psnr )
1663             {
1664                 x264_log( h, X264_LOG_INFO,
1665                           "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",
1666                           slice_name[i_slice],
1667                           i_count,
1668                           (double)h->stat.i_slice_qp[i_slice] / i_count,
1669                           (double)h->stat.i_slice_size[i_slice] / i_count,
1670                           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,
1671                           h->stat.f_psnr_average[i_slice] / i_count,
1672                           x264_psnr( h->stat.i_sqe_global[i_slice], i_count * i_yuv_size ) );
1673             }
1674             else
1675             {
1676                 x264_log( h, X264_LOG_INFO,
1677                           "slice %s:%-5d Avg QP:%5.2f  size:%6.0f\n",
1678                           slice_name[i_slice],
1679                           i_count,
1680                           (double)h->stat.i_slice_qp[i_slice] / i_count,
1681                           (double)h->stat.i_slice_size[i_slice] / i_count );
1682             }
1683         }
1684     }
1685
1686     /* MB types used */
1687     if( h->stat.i_slice_count[SLICE_TYPE_I] > 0 )
1688     {
1689         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_I];
1690         const double i_count = h->stat.i_slice_count[SLICE_TYPE_I] * h->mb.i_mb_count / 100.0;
1691         x264_log( h, X264_LOG_INFO,
1692                   "mb I  I16..4: %4.1f%% %4.1f%% %4.1f%%\n",
1693                   i_mb_count[I_16x16]/ i_count,
1694                   i_mb_count[I_8x8]  / i_count,
1695                   i_mb_count[I_4x4]  / i_count );
1696     }
1697     if( h->stat.i_slice_count[SLICE_TYPE_P] > 0 )
1698     {
1699         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_P];
1700         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_P];
1701         const double i_count = h->stat.i_slice_count[SLICE_TYPE_P] * h->mb.i_mb_count / 100.0;
1702         x264_log( h, X264_LOG_INFO,
1703                   "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",
1704                   i_mb_count[I_16x16]/ i_count,
1705                   i_mb_count[I_8x8]  / i_count,
1706                   i_mb_count[I_4x4]  / i_count,
1707                   i_mb_size[PIXEL_16x16] / (i_count*4),
1708                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
1709                   i_mb_size[PIXEL_8x8] / (i_count*4),
1710                   (i_mb_size[PIXEL_8x4] + i_mb_size[PIXEL_4x8]) / (i_count*4),
1711                   i_mb_size[PIXEL_4x4] / (i_count*4),
1712                   i_mb_count[P_SKIP] / i_count );
1713     }
1714     if( h->stat.i_slice_count[SLICE_TYPE_B] > 0 )
1715     {
1716         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_B];
1717         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_B];
1718         const double i_count = h->stat.i_slice_count[SLICE_TYPE_B] * h->mb.i_mb_count / 100.0;
1719         x264_log( h, X264_LOG_INFO,
1720                   "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",
1721                   i_mb_count[I_16x16]  / i_count,
1722                   i_mb_count[I_8x8]    / i_count,
1723                   i_mb_count[I_4x4]    / i_count,
1724                   i_mb_size[PIXEL_16x16] / (i_count*4),
1725                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
1726                   i_mb_size[PIXEL_8x8] / (i_count*4),
1727                   i_mb_count[B_DIRECT] / i_count,
1728                   i_mb_count[B_SKIP]   / i_count );
1729     }
1730
1731     x264_ratecontrol_summary( h );
1732
1733     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 )
1734     {
1735         const int i_count = h->stat.i_slice_count[SLICE_TYPE_I] +
1736                             h->stat.i_slice_count[SLICE_TYPE_P] +
1737                             h->stat.i_slice_count[SLICE_TYPE_B];
1738         float fps = (float) h->param.i_fps_num / h->param.i_fps_den;
1739 #define SUM3(p) (p[SLICE_TYPE_I] + p[SLICE_TYPE_P] + p[SLICE_TYPE_B])
1740 #define SUM3b(p,o) (p[SLICE_TYPE_I][o] + p[SLICE_TYPE_P][o] + p[SLICE_TYPE_B][o])
1741         float f_bitrate = fps * SUM3(h->stat.i_slice_size) / i_count / 125;
1742
1743         if( h->param.analyse.b_transform_8x8 )
1744         {
1745             int64_t i_i8x8 = SUM3b( h->stat.i_mb_count, I_8x8 );
1746             int64_t i_intra = i_i8x8 + SUM3b( h->stat.i_mb_count, I_4x4 )
1747                                      + SUM3b( h->stat.i_mb_count, I_16x16 );
1748             x264_log( h, X264_LOG_INFO, "8x8 transform  intra:%.1f%%  inter:%.1f%%\n",
1749                       100. * i_i8x8 / i_intra,
1750                       100. * h->stat.i_mb_count_8x8dct[1] / h->stat.i_mb_count_8x8dct[0] );
1751         }
1752
1753         if( h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
1754             && h->stat.i_slice_count[SLICE_TYPE_B] )
1755         {
1756             x264_log( h, X264_LOG_INFO, "direct mvs  spatial:%.1f%%  temporal:%.1f%%\n",
1757                       h->stat.i_direct_frames[1] * 100. / h->stat.i_slice_count[SLICE_TYPE_B],
1758                       h->stat.i_direct_frames[0] * 100. / h->stat.i_slice_count[SLICE_TYPE_B] );
1759         }
1760
1761         if( h->param.i_frame_reference > 1 )
1762         {
1763             int i_slice;
1764             for( i_slice = 0; i_slice < 2; i_slice++ )
1765             {
1766                 char buf[200];
1767                 char *p = buf;
1768                 int64_t i_den = 0;
1769                 int i_max = 0;
1770                 for( i = 0; i < h->param.i_frame_reference; i++ )
1771                     if( h->stat.i_mb_count_ref[i_slice][i] )
1772                     {
1773                         i_den += h->stat.i_mb_count_ref[i_slice][i];
1774                         i_max = i;
1775                     }
1776                 if( i_max == 0 )
1777                     continue;
1778                 for( i = 0; i <= i_max; i++ )
1779                     p += sprintf( p, " %4.1f%%", 100. * h->stat.i_mb_count_ref[i_slice][i] / i_den );
1780                 x264_log( h, X264_LOG_INFO, "ref %c %s\n", i_slice==SLICE_TYPE_P ? 'P' : 'B', buf );
1781             }
1782         }
1783
1784         if( h->param.analyse.b_psnr )
1785             x264_log( h, X264_LOG_INFO,
1786                       "PSNR Mean Y:%6.3f U:%6.3f V:%6.3f Avg:%6.3f Global:%6.3f kb/s:%.2f\n",
1787                       SUM3( h->stat.f_psnr_mean_y ) / i_count,
1788                       SUM3( h->stat.f_psnr_mean_u ) / i_count,
1789                       SUM3( h->stat.f_psnr_mean_v ) / i_count,
1790                       SUM3( h->stat.f_psnr_average ) / i_count,
1791                       x264_psnr( SUM3( h->stat.i_sqe_global ), i_count * i_yuv_size ),
1792                       f_bitrate );
1793         else
1794             x264_log( h, X264_LOG_INFO, "kb/s:%.1f\n", f_bitrate );
1795     }
1796
1797     /* frames */
1798     for( i = 0; i < X264_BFRAME_MAX + 3; i++ )
1799     {
1800         if( h->frames.current[i] ) x264_frame_delete( h->frames.current[i] );
1801         if( h->frames.next[i] )    x264_frame_delete( h->frames.next[i] );
1802         if( h->frames.unused[i] )  x264_frame_delete( h->frames.unused[i] );
1803     }
1804     /* ref frames */
1805     for( i = 0; i < h->frames.i_max_dpb; i++ )
1806     {
1807         x264_frame_delete( h->frames.reference[i] );
1808     }
1809
1810     /* rc */
1811     x264_ratecontrol_delete( h );
1812
1813     /* param */
1814     if( h->param.rc.psz_stat_out )
1815         free( h->param.rc.psz_stat_out );
1816     if( h->param.rc.psz_stat_in )
1817         free( h->param.rc.psz_stat_in );
1818     if( h->param.rc.psz_rc_eq )
1819         free( h->param.rc.psz_rc_eq );
1820
1821     x264_cqm_delete( h );
1822     x264_macroblock_cache_end( h );
1823     x264_free( h->out.p_bitstream );
1824     for( i = 1; i < h->param.i_threads; i++ )
1825         x264_free( h->thread[i] );
1826     x264_free( h );
1827 }
1828