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     if( h->param.i_threads == 0 )
 335         h->param.i_threads = x264_cpu_num_processors();
 336     h->param.i_threads = x264_clip3( h->param.i_threads, 1, X264_SLICE_MAX );
 337     h->param.i_threads = X264_MIN( h->param.i_threads, (h->param.i_height + 15) / 16 );
 338 #ifndef HAVE_PTHREAD
 339     if( h->param.i_threads > 1 )
 340     {
 341         x264_log( h, X264_LOG_WARNING, "not compiled with pthread support!\n");
 342         x264_log( h, X264_LOG_WARNING, "multislicing anyway, but you won't see any speed gain.\n" );
 343     }
 344 #endif
 345
 346     if( h->param.rc.i_rc_method < 0 || h->param.rc.i_rc_method > 2 )
 347     {
 348         x264_log( h, X264_LOG_ERROR, "invalid RC method\n" );
 349         return -1;
 350     }
 351     h->param.rc.i_rf_constant = x264_clip3( h->param.rc.i_rf_constant, 0, 51 );
 352     h->param.rc.i_qp_constant = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
 353     if( h->param.rc.i_rc_method == X264_RC_CRF )
 354         h->param.rc.i_qp_constant = h->param.rc.i_rf_constant;
 355     if( (h->param.rc.i_rc_method == X264_RC_CQP || h->param.rc.i_rc_method == X264_RC_CRF)
 356         && h->param.rc.i_qp_constant == 0 )
 357     {
 358         h->mb.b_lossless = 1;
 359         h->param.i_cqm_preset = X264_CQM_FLAT;
 360         h->param.psz_cqm_file = NULL;
 361         h->param.rc.i_rc_method = X264_RC_CQP;
 362         h->param.rc.f_ip_factor = 1;
 363         h->param.rc.f_pb_factor = 1;
 364         h->param.analyse.b_transform_8x8 = 0;
 365         h->param.analyse.b_psnr = 0;
 366         h->param.analyse.b_ssim = 0;
 367         h->param.analyse.i_chroma_qp_offset = 0;
 368         h->param.analyse.i_trellis = 0;
 369         h->param.analyse.b_fast_pskip = 0;
 370         h->param.analyse.i_noise_reduction = 0;
 371         h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 1, 6 );
 372     }
 373
 374     if( ( h->param.i_width % 16 || h->param.i_height % 16 ) && !h->mb.b_lossless )
 375     {
 376         x264_log( h, X264_LOG_WARNING,
 377                   "width or height not divisible by 16 (%dx%d), compression will suffer.\n",
 378                   h->param.i_width, h->param.i_height );
 379     }
 380
 381     h->param.i_frame_reference = x264_clip3( h->param.i_frame_reference, 1, 16 );
 382     if( h->param.i_keyint_max <= 0 )
 383         h->param.i_keyint_max = 1;
 384     h->param.i_keyint_min = x264_clip3( h->param.i_keyint_min, 1, h->param.i_keyint_max/2+1 );
 385
 386     h->param.i_bframe = x264_clip3( h->param.i_bframe, 0, X264_BFRAME_MAX );
 387     h->param.i_bframe_bias = x264_clip3( h->param.i_bframe_bias, -90, 100 );
 388     h->param.b_bframe_pyramid = h->param.b_bframe_pyramid && h->param.i_bframe > 1;
 389     h->param.b_bframe_adaptive = h->param.b_bframe_adaptive && h->param.i_bframe > 0;
 390     h->param.analyse.b_weighted_bipred = h->param.analyse.b_weighted_bipred && h->param.i_bframe > 0;
 391     h->mb.b_direct_auto_write = h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
 392                                 && h->param.i_bframe
 393                                 && ( h->param.rc.b_stat_write || !h->param.rc.b_stat_read );
 394
 395     h->param.i_deblocking_filter_alphac0 = x264_clip3( h->param.i_deblocking_filter_alphac0, -6, 6 );
 396     h->param.i_deblocking_filter_beta    = x264_clip3( h->param.i_deblocking_filter_beta, -6, 6 );
 397
 398     h->param.i_cabac_init_idc = x264_clip3( h->param.i_cabac_init_idc, 0, 2 );
 399
 400     if( h->param.i_cqm_preset < X264_CQM_FLAT || h->param.i_cqm_preset > X264_CQM_CUSTOM )
 401         h->param.i_cqm_preset = X264_CQM_FLAT;
 402
 403     if( h->param.analyse.i_me_method < X264_ME_DIA ||
 404         h->param.analyse.i_me_method > X264_ME_ESA )
 405         h->param.analyse.i_me_method = X264_ME_HEX;
 406     if( h->param.analyse.i_me_range < 4 )
 407         h->param.analyse.i_me_range = 4;
 408     if( h->param.analyse.i_me_range > 16 && h->param.analyse.i_me_method <= X264_ME_HEX )
 409         h->param.analyse.i_me_range = 16;
 410     h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 1, 7 );
 411     h->param.analyse.b_bframe_rdo = h->param.analyse.b_bframe_rdo && h->param.analyse.i_subpel_refine >= 6;
 412     h->param.analyse.b_mixed_references = h->param.analyse.b_mixed_references && h->param.i_frame_reference > 1;
 413     h->param.analyse.inter &= X264_ANALYSE_PSUB16x16|X264_ANALYSE_PSUB8x8|X264_ANALYSE_BSUB16x16|
 414                               X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 415     h->param.analyse.intra &= X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 416     if( !(h->param.analyse.inter & X264_ANALYSE_PSUB16x16) )
 417         h->param.analyse.inter &= ~X264_ANALYSE_PSUB8x8;
 418     if( !h->param.analyse.b_transform_8x8 )
 419     {
 420         h->param.analyse.inter &= ~X264_ANALYSE_I8x8;
 421         h->param.analyse.intra &= ~X264_ANALYSE_I8x8;
 422     }
 423     h->param.analyse.i_chroma_qp_offset = x264_clip3(h->param.analyse.i_chroma_qp_offset, -12, 12);
 424     if( !h->param.b_cabac )
 425         h->param.analyse.i_trellis = 0;
 426     h->param.analyse.i_trellis = x264_clip3( h->param.analyse.i_trellis, 0, 2 );
 427     h->param.analyse.i_noise_reduction = x264_clip3( h->param.analyse.i_noise_reduction, 0, 1<<16 );
 428
 429     {
 430         const x264_level_t *l = x264_levels;
 431         while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
 432             l++;
 433         if( l->level_idc == 0 )
 434         {
 435             x264_log( h, X264_LOG_ERROR, "invalid level_idc: %d\n", h->param.i_level_idc );
 436             return -1;
 437         }
 438         if( h->param.analyse.i_mv_range <= 0 )
 439             h->param.analyse.i_mv_range = l->mv_range;
 440         else
 441             h->param.analyse.i_mv_range = x264_clip3(h->param.analyse.i_mv_range, 32, 2048);
 442     }
 443
 444     if( h->param.rc.f_qblur < 0 )
 445         h->param.rc.f_qblur = 0;
 446     if( h->param.rc.f_complexity_blur < 0 )
 447         h->param.rc.f_complexity_blur = 0;
 448
 449     h->param.i_sps_id &= 31;
 450
 451     if( h->param.i_log_level < X264_LOG_INFO )
 452     {
 453         h->param.analyse.b_psnr = 0;
 454         h->param.analyse.b_ssim = 0;
 455     }
 456
 457     /* ensure the booleans are 0 or 1 so they can be used in math */
 458 #define BOOLIFY(x) h->param.x = !!h->param.x
 459     BOOLIFY( b_cabac );
 460     BOOLIFY( b_deblocking_filter );
 461     BOOLIFY( analyse.b_transform_8x8 );
 462     BOOLIFY( analyse.b_bidir_me );
 463     BOOLIFY( analyse.b_chroma_me );
 464     BOOLIFY( analyse.b_fast_pskip );
 465     BOOLIFY( rc.b_stat_write );
 466     BOOLIFY( rc.b_stat_read );
 467 #undef BOOLIFY
 468
 469     return 0;
 470 }
 471
 472 static void mbcmp_init( x264_t *h )
 473 {
 474     memcpy( h->pixf.mbcmp,
 475             ( h->mb.b_lossless || h->param.analyse.i_subpel_refine <= 1 ) ? h->pixf.sad : h->pixf.satd,
 476             sizeof(h->pixf.mbcmp) );
 477 }
 478
 479 /****************************************************************************
 480  * x264_encoder_open:
 481  ****************************************************************************/
 482 x264_t *x264_encoder_open   ( x264_param_t *param )
 483 {
 484     x264_t *h = x264_malloc( sizeof( x264_t ) );
 485     int i;
 486
 487     memset( h, 0, sizeof( x264_t ) );
 488
 489     /* Create a copy of param */
 490     memcpy( &h->param, param, sizeof( x264_param_t ) );
 491
 492     if( x264_validate_parameters( h ) < 0 )
 493     {
 494         x264_free( h );
 495         return NULL;
 496     }
 497
 498     if( h->param.psz_cqm_file )
 499         if( x264_cqm_parse_file( h, h->param.psz_cqm_file ) < 0 )
 500         {
 501             x264_free( h );
 502             return NULL;
 503         }
 504
 505     if( h->param.rc.psz_stat_out )
 506         h->param.rc.psz_stat_out = strdup( h->param.rc.psz_stat_out );
 507     if( h->param.rc.psz_stat_in )
 508         h->param.rc.psz_stat_in = strdup( h->param.rc.psz_stat_in );
 509     if( h->param.rc.psz_rc_eq )
 510         h->param.rc.psz_rc_eq = strdup( h->param.rc.psz_rc_eq );
 511
 512     /* VUI */
 513     if( h->param.vui.i_sar_width > 0 && h->param.vui.i_sar_height > 0 )
 514     {
 515         int i_w = param->vui.i_sar_width;
 516         int i_h = param->vui.i_sar_height;
 517
 518         x264_reduce_fraction( &i_w, &i_h );
 519
 520         while( i_w > 65535 || i_h > 65535 )
 521         {
 522             i_w /= 2;
 523             i_h /= 2;
 524         }
 525
 526         h->param.vui.i_sar_width = 0;
 527         h->param.vui.i_sar_height = 0;
 528         if( i_w == 0 || i_h == 0 )
 529         {
 530             x264_log( h, X264_LOG_WARNING, "cannot create valid sample aspect ratio\n" );
 531         }
 532         else
 533         {
 534             x264_log( h, X264_LOG_INFO, "using SAR=%d/%d\n", i_w, i_h );
 535             h->param.vui.i_sar_width = i_w;
 536             h->param.vui.i_sar_height = i_h;
 537         }
 538     }
 539
 540     x264_reduce_fraction( &h->param.i_fps_num, &h->param.i_fps_den );
 541
 542     /* Init x264_t */
 543     h->out.i_nal = 0;
 544     h->out.i_bitstream = X264_MAX( 1000000, h->param.i_width * h->param.i_height * 1.7
 545         * ( h->param.rc.i_rc_method == X264_RC_ABR ? pow( 0.5, h->param.rc.i_qp_min )
 546           : pow( 0.5, h->param.rc.i_qp_constant ) * X264_MAX( 1, h->param.rc.f_ip_factor )));
 547     h->out.p_bitstream = x264_malloc( h->out.i_bitstream );
 548
 549     h->i_frame = 0;
 550     h->i_frame_num = 0;
 551     h->i_idr_pic_id = 0;
 552
 553     h->sps = &h->sps_array[0];
 554     x264_sps_init( h->sps, h->param.i_sps_id, &h->param );
 555
 556     h->pps = &h->pps_array[0];
 557     x264_pps_init( h->pps, h->param.i_sps_id, &h->param, h->sps);
 558
 559     x264_validate_levels( h );
 560
 561     x264_cqm_init( h );
 562
 563     h->mb.i_mb_count = h->sps->i_mb_width * h->sps->i_mb_height;
 564
 565     /* Init frames. */
 566     h->frames.i_delay = h->param.i_bframe;
 567     h->frames.i_max_ref0 = h->param.i_frame_reference;
 568     h->frames.i_max_ref1 = h->sps->vui.i_num_reorder_frames;
 569     h->frames.i_max_dpb  = h->sps->vui.i_max_dec_frame_buffering + 1;
 570     h->frames.b_have_lowres = !h->param.rc.b_stat_read
 571         && ( h->param.rc.i_rc_method == X264_RC_ABR
 572           || h->param.rc.i_rc_method == X264_RC_CRF
 573           || h->param.b_bframe_adaptive );
 574
 575     for( i = 0; i < X264_BFRAME_MAX + 3; i++ )
 576     {
 577         h->frames.current[i] = NULL;
 578         h->frames.next[i]    = NULL;
 579         h->frames.unused[i]  = NULL;
 580     }
 581     for( i = 0; i < 1 + h->frames.i_delay; i++ )
 582     {
 583         h->frames.unused[i] =  x264_frame_new( h );
 584         if( !h->frames.unused[i] )
 585             return NULL;
 586     }
 587     for( i = 0; i < h->frames.i_max_dpb; i++ )
 588     {
 589         h->frames.reference[i] = x264_frame_new( h );
 590         if( !h->frames.reference[i] )
 591             return NULL;
 592     }
 593     h->frames.reference[h->frames.i_max_dpb] = NULL;
 594     h->frames.i_last_idr = - h->param.i_keyint_max;
 595     h->frames.i_input    = 0;
 596     h->frames.last_nonb  = NULL;
 597
 598     h->i_ref0 = 0;
 599     h->i_ref1 = 0;
 600
 601     h->fdec = h->frames.reference[0];
 602
 603     if( x264_macroblock_cache_init( h ) < 0 )
 604         return NULL;
 605     x264_rdo_init( );
 606
 607     /* init CPU functions */
 608     x264_predict_16x16_init( h->param.cpu, h->predict_16x16 );
 609     x264_predict_8x8c_init( h->param.cpu, h->predict_8x8c );
 610     x264_predict_8x8_init( h->param.cpu, h->predict_8x8 );
 611     x264_predict_4x4_init( h->param.cpu, h->predict_4x4 );
 612
 613     x264_pixel_init( h->param.cpu, &h->pixf );
 614     x264_dct_init( h->param.cpu, &h->dctf );
 615     x264_mc_init( h->param.cpu, &h->mc );
 616     x264_csp_init( h->param.cpu, h->param.i_csp, &h->csp );
 617     x264_quant_init( h, h->param.cpu, &h->quantf );
 618     x264_deblock_init( h->param.cpu, &h->loopf );
 619
 620     mbcmp_init( h );
 621
 622     /* rate control */
 623     if( x264_ratecontrol_new( h ) < 0 )
 624         return NULL;
 625
 626     x264_log( h, X264_LOG_INFO, "using cpu capabilities %s%s%s%s%s%s\n",
 627              param->cpu&X264_CPU_MMX ? "MMX " : "",
 628              param->cpu&X264_CPU_MMXEXT ? "MMXEXT " : "",
 629              param->cpu&X264_CPU_SSE ? "SSE " : "",
 630              param->cpu&X264_CPU_SSE2 ? "SSE2 " : "",
 631              param->cpu&X264_CPU_3DNOW ? "3DNow! " : "",
 632              param->cpu&X264_CPU_ALTIVEC ? "Altivec " : "" );
 633
 634     h->thread[0] = h;
 635     h->i_thread_num = 0;
 636     for( i = 1; i < h->param.i_threads; i++ )
 637         h->thread[i] = x264_malloc( sizeof(x264_t) );
 638
 639 #ifdef DEBUG_DUMP_FRAME
 640     {
 641         /* create or truncate the reconstructed video file */
 642         FILE *f = fopen( "fdec.yuv", "w" );
 643         if( f )
 644             fclose( f );
 645         else
 646         {
 647             x264_log( h, X264_LOG_ERROR, "can't write to fdec.yuv\n" );
 648             x264_free( h );
 649             return NULL;
 650         }
 651     }
 652 #endif
 653
 654     return h;
 655 }
 656
 657 /****************************************************************************
 658  * x264_encoder_reconfig:
 659  ****************************************************************************/
 660 int x264_encoder_reconfig( x264_t *h, x264_param_t *param )
 661 {
 662 #define COPY(var) h->param.var = param->var
 663     COPY( i_frame_reference ); // but never uses more refs than initially specified
 664     COPY( i_bframe_bias );
 665     COPY( i_scenecut_threshold );
 666     COPY( b_deblocking_filter );
 667     COPY( i_deblocking_filter_alphac0 );
 668     COPY( i_deblocking_filter_beta );
 669     COPY( analyse.intra );
 670     COPY( analyse.inter );
 671     COPY( analyse.i_direct_mv_pred );
 672     COPY( analyse.i_me_method );
 673     COPY( analyse.i_me_range );
 674     COPY( analyse.i_noise_reduction );
 675     COPY( analyse.i_subpel_refine );
 676     COPY( analyse.i_trellis );
 677     COPY( analyse.b_bidir_me );
 678     COPY( analyse.b_bframe_rdo );
 679     COPY( analyse.b_chroma_me );
 680     COPY( analyse.b_dct_decimate );
 681     COPY( analyse.b_fast_pskip );
 682     COPY( analyse.b_mixed_references );
 683 #undef COPY
 684
 685     if( h->pps->b_transform_8x8_mode )
 686         h->param.analyse.b_transform_8x8 = param->analyse.b_transform_8x8;
 687
 688     mbcmp_init( h );
 689
 690     return x264_validate_parameters( h );
 691 }
 692
 693 /* internal usage */
 694 static void x264_nal_start( x264_t *h, int i_type, int i_ref_idc )
 695 {
 696     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 697
 698     nal->i_ref_idc = i_ref_idc;
 699     nal->i_type    = i_type;
 700
 701     bs_align_0( &h->out.bs );   /* not needed */
 702
 703     nal->i_payload= 0;
 704     nal->p_payload= &h->out.p_bitstream[bs_pos( &h->out.bs) / 8];
 705 }
 706 static void x264_nal_end( x264_t *h )
 707 {
 708     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 709
 710     bs_align_0( &h->out.bs );   /* not needed */
 711
 712     nal->i_payload = &h->out.p_bitstream[bs_pos( &h->out.bs)/8] - nal->p_payload;
 713
 714     h->out.i_nal++;
 715 }
 716
 717 /****************************************************************************
 718  * x264_encoder_headers:
 719  ****************************************************************************/
 720 int x264_encoder_headers( x264_t *h, x264_nal_t **pp_nal, int *pi_nal )
 721 {
 722     /* init bitstream context */
 723     h->out.i_nal = 0;
 724     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
 725
 726     /* Put SPS and PPS */
 727     if( h->i_frame == 0 )
 728     {
 729         /* identify ourself */
 730         x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
 731         x264_sei_version_write( h, &h->out.bs );
 732         x264_nal_end( h );
 733
 734         /* generate sequence parameters */
 735         x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
 736         x264_sps_write( &h->out.bs, h->sps );
 737         x264_nal_end( h );
 738
 739         /* generate picture parameters */
 740         x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
 741         x264_pps_write( &h->out.bs, h->pps );
 742         x264_nal_end( h );
 743     }
 744     /* now set output*/
 745     *pi_nal = h->out.i_nal;
 746     *pp_nal = &h->out.nal[0];
 747
 748     return 0;
 749 }
 750
 751
 752 static void x264_frame_put( x264_frame_t *list[X264_BFRAME_MAX], x264_frame_t *frame )
 753 {
 754     int i = 0;
 755     while( list[i] ) i++;
 756     list[i] = frame;
 757 }
 758
 759 static void x264_frame_push( x264_frame_t *list[X264_BFRAME_MAX], x264_frame_t *frame )
 760 {
 761     int i = 0;
 762     while( list[i] ) i++;
 763     while( i-- )
 764         list[i+1] = list[i];
 765     list[0] = frame;
 766 }
 767
 768 static x264_frame_t *x264_frame_get( x264_frame_t *list[X264_BFRAME_MAX+1] )
 769 {
 770     x264_frame_t *frame = list[0];
 771     int i;
 772     for( i = 0; list[i]; i++ )
 773         list[i] = list[i+1];
 774     return frame;
 775 }
 776
 777 static void x264_frame_sort( x264_frame_t *list[X264_BFRAME_MAX+1], int b_dts )
 778 {
 779     int i, b_ok;
 780     do {
 781         b_ok = 1;
 782         for( i = 0; list[i+1]; i++ )
 783         {
 784             int dtype = list[i]->i_type - list[i+1]->i_type;
 785             int dtime = list[i]->i_frame - list[i+1]->i_frame;
 786             int swap = b_dts ? dtype > 0 || ( dtype == 0 && dtime > 0 )
 787                              : dtime > 0;
 788             if( swap )
 789             {
 790                 XCHG( x264_frame_t*, list[i], list[i+1] );
 791                 b_ok = 0;
 792             }
 793         }
 794     } while( !b_ok );
 795 }
 796 #define x264_frame_sort_dts(list) x264_frame_sort(list, 1)
 797 #define x264_frame_sort_pts(list) x264_frame_sort(list, 0)
 798
 799 static inline void x264_reference_build_list( x264_t *h, int i_poc, int i_slice_type )
 800 {
 801     int i;
 802     int b_ok;
 803
 804     /* build ref list 0/1 */
 805     h->i_ref0 = 0;
 806     h->i_ref1 = 0;
 807     for( i = 1; i < h->frames.i_max_dpb; i++ )
 808     {
 809         if( h->frames.reference[i]->i_poc >= 0 )
 810         {
 811             if( h->frames.reference[i]->i_poc < i_poc )
 812             {
 813                 h->fref0[h->i_ref0++] = h->frames.reference[i];
 814             }
 815             else if( h->frames.reference[i]->i_poc > i_poc )
 816             {
 817                 h->fref1[h->i_ref1++] = h->frames.reference[i];
 818             }
 819         }
 820     }
 821
 822     /* Order ref0 from higher to lower poc */
 823     do
 824     {
 825         b_ok = 1;
 826         for( i = 0; i < h->i_ref0 - 1; i++ )
 827         {
 828             if( h->fref0[i]->i_poc < h->fref0[i+1]->i_poc )
 829             {
 830                 XCHG( x264_frame_t*, h->fref0[i], h->fref0[i+1] );
 831                 b_ok = 0;
 832                 break;
 833             }
 834         }
 835     } while( !b_ok );
 836     /* Order ref1 from lower to higher poc (bubble sort) for B-frame */
 837     do
 838     {
 839         b_ok = 1;
 840         for( i = 0; i < h->i_ref1 - 1; i++ )
 841         {
 842             if( h->fref1[i]->i_poc > h->fref1[i+1]->i_poc )
 843             {
 844                 XCHG( x264_frame_t*, h->fref1[i], h->fref1[i+1] );
 845                 b_ok = 0;
 846                 break;
 847             }
 848         }
 849     } while( !b_ok );
 850
 851     /* In the standard, a P-frame's ref list is sorted by frame_num.
 852      * We use POC, but check whether explicit reordering is needed */
 853     h->b_ref_reorder[0] =
 854     h->b_ref_reorder[1] = 0;
 855     if( i_slice_type == SLICE_TYPE_P )
 856     {
 857         for( i = 0; i < h->i_ref0 - 1; i++ )
 858             if( h->fref0[i]->i_frame_num < h->fref0[i+1]->i_frame_num )
 859             {
 860                 h->b_ref_reorder[0] = 1;
 861                 break;
 862             }
 863     }
 864
 865     h->i_ref1 = X264_MIN( h->i_ref1, h->frames.i_max_ref1 );
 866     h->i_ref0 = X264_MIN( h->i_ref0, h->frames.i_max_ref0 );
 867     h->i_ref0 = X264_MIN( h->i_ref0, h->param.i_frame_reference ); // if reconfig() has lowered the limit
 868     h->i_ref0 = X264_MIN( h->i_ref0, 16 - h->i_ref1 );
 869 }
 870
 871 static inline void x264_fdec_deblock( x264_t *h )
 872 {
 873     /* apply deblocking filter to the current decoded picture */
 874     if( !h->sh.i_disable_deblocking_filter_idc )
 875     {
 876         TIMER_START( i_mtime_filter );
 877         x264_frame_deblocking_filter( h, h->sh.i_type );
 878         TIMER_STOP( i_mtime_filter );
 879     }
 880 }
 881
 882 static inline void x264_reference_update( x264_t *h )
 883 {
 884     int i;
 885
 886     x264_fdec_deblock( h );
 887
 888     /* expand border */
 889     x264_frame_expand_border( h->fdec );
 890
 891     /* create filtered images */
 892     x264_frame_filter( h->param.cpu, h->fdec );
 893
 894     /* expand border of filtered images */
 895     x264_frame_expand_border_filtered( h->fdec );
 896
 897     /* move lowres copy of the image to the ref frame */
 898     for( i = 0; i < 4; i++)
 899         XCHG( uint8_t*, h->fdec->lowres[i], h->fenc->lowres[i] );
 900
 901     /* adaptive B decision needs a pointer, since it can't use the ref lists */
 902     if( h->sh.i_type != SLICE_TYPE_B )
 903         h->frames.last_nonb = h->fdec;
 904
 905     /* move frame in the buffer */
 906     h->fdec = h->frames.reference[h->frames.i_max_dpb-1];
 907     for( i = h->frames.i_max_dpb-1; i > 0; i-- )
 908     {
 909         h->frames.reference[i] = h->frames.reference[i-1];
 910     }
 911     h->frames.reference[0] = h->fdec;
 912 }
 913
 914 static inline void x264_reference_reset( x264_t *h )
 915 {
 916     int i;
 917
 918     /* reset ref pictures */
 919     for( i = 1; i < h->frames.i_max_dpb; i++ )
 920     {
 921         h->frames.reference[i]->i_poc = -1;
 922     }
 923     h->frames.reference[0]->i_poc = 0;
 924 }
 925
 926 static inline void x264_slice_init( x264_t *h, int i_nal_type, int i_slice_type, int i_global_qp )
 927 {
 928     /* ------------------------ Create slice header  ----------------------- */
 929     if( i_nal_type == NAL_SLICE_IDR )
 930     {
 931         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 );
 932
 933         /* increment id */
 934         h->i_idr_pic_id = ( h->i_idr_pic_id + 1 ) % 65536;
 935     }
 936     else
 937     {
 938         x264_slice_header_init( h, &h->sh, h->sps, h->pps, i_slice_type, -1, h->i_frame_num, i_global_qp );
 939
 940         /* always set the real higher num of ref frame used */
 941         h->sh.b_num_ref_idx_override = 1;
 942         h->sh.i_num_ref_idx_l0_active = h->i_ref0 <= 0 ? 1 : h->i_ref0;
 943         h->sh.i_num_ref_idx_l1_active = h->i_ref1 <= 0 ? 1 : h->i_ref1;
 944     }
 945
 946     h->fdec->i_frame_num = h->sh.i_frame_num;
 947
 948     if( h->sps->i_poc_type == 0 )
 949     {
 950         h->sh.i_poc_lsb = h->fdec->i_poc & ( (1 << h->sps->i_log2_max_poc_lsb) - 1 );
 951         h->sh.i_delta_poc_bottom = 0;   /* XXX won't work for field */
 952     }
 953     else if( h->sps->i_poc_type == 1 )
 954     {
 955         /* FIXME TODO FIXME */
 956     }
 957     else
 958     {
 959         /* Nothing to do ? */
 960     }
 961
 962     x264_macroblock_slice_init( h );
 963 }
 964
 965 static int x264_slice_write( x264_t *h )
 966 {
 967     int i_skip;
 968     int mb_xy;
 969     int i;
 970
 971     /* init stats */
 972     memset( &h->stat.frame, 0, sizeof(h->stat.frame) );
 973
 974     /* Slice */
 975     x264_nal_start( h, h->i_nal_type, h->i_nal_ref_idc );
 976
 977     /* Slice header */
 978     x264_slice_header_write( &h->out.bs, &h->sh, h->i_nal_ref_idc );
 979     if( h->param.b_cabac )
 980     {
 981         /* alignment needed */
 982         bs_align_1( &h->out.bs );
 983
 984         /* init cabac */
 985         x264_cabac_context_init( &h->cabac, h->sh.i_type, h->sh.i_qp, h->sh.i_cabac_init_idc );
 986         x264_cabac_encode_init ( &h->cabac, &h->out.bs );
 987     }
 988     h->mb.i_last_qp = h->sh.i_qp;
 989     h->mb.i_last_dqp = 0;
 990
 991     for( mb_xy = h->sh.i_first_mb, i_skip = 0; mb_xy < h->sh.i_last_mb; mb_xy++ )
 992     {
 993         const int i_mb_y = mb_xy / h->sps->i_mb_width;
 994         const int i_mb_x = mb_xy % h->sps->i_mb_width;
 995
 996         int mb_spos = bs_pos(&h->out.bs);
 997
 998         /* load cache */
 999         x264_macroblock_cache_load( h, i_mb_x, i_mb_y );
1000
1001         /* analyse parameters
1002          * Slice I: choose I_4x4 or I_16x16 mode
1003          * Slice P: choose between using P mode or intra (4x4 or 16x16)
1004          * */
1005         TIMER_START( i_mtime_analyse );
1006         x264_macroblock_analyse( h );
1007         TIMER_STOP( i_mtime_analyse );
1008
1009         /* encode this macrobock -> be careful it can change the mb type to P_SKIP if needed */
1010         TIMER_START( i_mtime_encode );
1011         x264_macroblock_encode( h );
1012         TIMER_STOP( i_mtime_encode );
1013
1014         TIMER_START( i_mtime_write );
1015         if( h->param.b_cabac )
1016         {
1017             if( mb_xy > h->sh.i_first_mb )
1018                 x264_cabac_encode_terminal( &h->cabac, 0 );
1019
1020             if( IS_SKIP( h->mb.i_type ) )
1021                 x264_cabac_mb_skip( h, 1 );
1022             else
1023             {
1024                 if( h->sh.i_type != SLICE_TYPE_I )
1025                     x264_cabac_mb_skip( h, 0 );
1026                 x264_macroblock_write_cabac( h, &h->cabac );
1027             }
1028         }
1029         else
1030         {
1031             if( IS_SKIP( h->mb.i_type ) )
1032                 i_skip++;
1033             else
1034             {
1035                 if( h->sh.i_type != SLICE_TYPE_I )
1036                 {
1037                     bs_write_ue( &h->out.bs, i_skip );  /* skip run */
1038                     i_skip = 0;
1039                 }
1040                 x264_macroblock_write_cavlc( h, &h->out.bs );
1041             }
1042         }
1043         TIMER_STOP( i_mtime_write );
1044
1045 #if VISUALIZE
1046         if( h->param.b_visualize )
1047             x264_visualize_mb( h );
1048 #endif
1049
1050         /* save cache */
1051         x264_macroblock_cache_save( h );
1052
1053         /* accumulate mb stats */
1054         h->stat.frame.i_mb_count[h->mb.i_type]++;
1055         if( !IS_SKIP(h->mb.i_type) && !IS_INTRA(h->mb.i_type) && !IS_DIRECT(h->mb.i_type) )
1056         {
1057             if( h->mb.i_partition != D_8x8 )
1058                 h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_partition ] ] += 4;
1059             else
1060                 for( i = 0; i < 4; i++ )
1061                     h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_sub_partition[i] ] ] ++;
1062             if( h->param.i_frame_reference > 1 )
1063             {
1064                 for( i = 0; i < 4; i++ )
1065                 {
1066                     int i_ref = h->mb.cache.ref[0][ x264_scan8[4*i] ];
1067                     if( i_ref >= 0 )
1068                         h->stat.frame.i_mb_count_ref[i_ref] ++;
1069                 }
1070             }
1071         }
1072         if( h->mb.i_cbp_luma && !IS_INTRA(h->mb.i_type) )
1073         {
1074             h->stat.frame.i_mb_count_8x8dct[0] ++;
1075             h->stat.frame.i_mb_count_8x8dct[1] += h->mb.b_transform_8x8;
1076         }
1077
1078         if( h->mb.b_variable_qp )
1079             x264_ratecontrol_mb(h, bs_pos(&h->out.bs) - mb_spos);
1080     }
1081
1082     if( h->param.b_cabac )
1083     {
1084         /* end of slice */
1085         x264_cabac_encode_terminal( &h->cabac, 1 );
1086     }
1087     else if( i_skip > 0 )
1088     {
1089         bs_write_ue( &h->out.bs, i_skip );  /* last skip run */
1090     }
1091
1092     if( h->param.b_cabac )
1093     {
1094         x264_cabac_encode_flush( &h->cabac );
1095
1096     }
1097     else
1098     {
1099         /* rbsp_slice_trailing_bits */
1100         bs_rbsp_trailing( &h->out.bs );
1101     }
1102
1103     x264_nal_end( h );
1104
1105     /* Compute misc bits */
1106     h->stat.frame.i_misc_bits = bs_pos( &h->out.bs )
1107                               + NALU_OVERHEAD * 8
1108                               - h->stat.frame.i_itex_bits
1109                               - h->stat.frame.i_ptex_bits
1110                               - h->stat.frame.i_hdr_bits;
1111
1112     return 0;
1113 }
1114
1115 static inline int x264_slices_write( x264_t *h )
1116 {
1117     int i_frame_size;
1118
1119 #if VISUALIZE
1120     if( h->param.b_visualize )
1121         x264_visualize_init( h );
1122 #endif
1123
1124     if( h->param.i_threads == 1 )
1125     {
1126         x264_ratecontrol_threads_start( h );
1127         x264_slice_write( h );
1128         i_frame_size = h->out.nal[h->out.i_nal-1].i_payload;
1129     }
1130     else
1131     {
1132         int i_nal = h->out.i_nal;
1133         int i_bs_size = h->out.i_bitstream / h->param.i_threads;
1134         int i;
1135         /* duplicate contexts */
1136         for( i = 0; i < h->param.i_threads; i++ )
1137         {
1138             x264_t *t = h->thread[i];
1139             if( i > 0 )
1140             {
1141                 memcpy( t, h, sizeof(x264_t) );
1142                 t->out.p_bitstream += i*i_bs_size;
1143                 bs_init( &t->out.bs, t->out.p_bitstream, i_bs_size );
1144                 t->i_thread_num = i;
1145             }
1146             t->sh.i_first_mb = (i    * h->sps->i_mb_height / h->param.i_threads) * h->sps->i_mb_width;
1147             t->sh.i_last_mb = ((i+1) * h->sps->i_mb_height / h->param.i_threads) * h->sps->i_mb_width;
1148             t->out.i_nal = i_nal + i;
1149         }
1150         x264_ratecontrol_threads_start( h );
1151
1152         /* dispatch */
1153 #ifdef HAVE_PTHREAD
1154         {
1155             pthread_t handles[X264_SLICE_MAX];
1156             for( i = 0; i < h->param.i_threads; i++ )
1157                 pthread_create( &handles[i], NULL, (void*)x264_slice_write, (void*)h->thread[i] );
1158             for( i = 0; i < h->param.i_threads; i++ )
1159                 pthread_join( handles[i], NULL );
1160         }
1161 #else
1162         for( i = 0; i < h->param.i_threads; i++ )
1163             x264_slice_write( h->thread[i] );
1164 #endif
1165
1166         /* merge contexts */
1167         i_frame_size = h->out.nal[i_nal].i_payload;
1168         for( i = 1; i < h->param.i_threads; i++ )
1169         {
1170             int j;
1171             x264_t *t = h->thread[i];
1172             h->out.nal[i_nal+i] = t->out.nal[i_nal+i];
1173             i_frame_size += t->out.nal[i_nal+i].i_payload;
1174             // all entries in stat.frame are ints
1175             for( j = 0; j < sizeof(h->stat.frame) / sizeof(int); j++ )
1176                 ((int*)&h->stat.frame)[j] += ((int*)&t->stat.frame)[j];
1177         }
1178         h->out.i_nal = i_nal + h->param.i_threads;
1179     }
1180
1181 #if VISUALIZE
1182     if( h->param.b_visualize )
1183     {
1184         x264_visualize_show( h );
1185         x264_visualize_close( h );
1186     }
1187 #endif
1188
1189     return i_frame_size;
1190 }
1191
1192 /****************************************************************************
1193  * x264_encoder_encode:
1194  *  XXX: i_poc   : is the poc of the current given picture
1195  *       i_frame : is the number of the frame being coded
1196  *  ex:  type frame poc
1197  *       I      0   2*0
1198  *       P      1   2*3
1199  *       B      2   2*1
1200  *       B      3   2*2
1201  *       P      4   2*6
1202  *       B      5   2*4
1203  *       B      6   2*5
1204  ****************************************************************************/
1205 int     x264_encoder_encode( x264_t *h,
1206                              x264_nal_t **pp_nal, int *pi_nal,
1207                              x264_picture_t *pic_in,
1208                              x264_picture_t *pic_out )
1209 {
1210     x264_frame_t   *frame_psnr = h->fdec; /* just to keep the current decoded frame for psnr calculation */
1211     int     i_nal_type;
1212     int     i_nal_ref_idc;
1213     int     i_slice_type;
1214     int     i_frame_size;
1215
1216     int i;
1217
1218     int   i_global_qp;
1219
1220     char psz_message[80];
1221
1222     /* no data out */
1223     *pi_nal = 0;
1224     *pp_nal = NULL;
1225
1226
1227     /* ------------------- Setup new frame from picture -------------------- */
1228     TIMER_START( i_mtime_encode_frame );
1229     if( pic_in != NULL )
1230     {
1231         /* 1: Copy the picture to a frame and move it to a buffer */
1232         x264_frame_t *fenc = x264_frame_get( h->frames.unused );
1233
1234         x264_frame_copy_picture( h, fenc, pic_in );
1235
1236         if( h->param.i_width % 16 || h->param.i_height % 16 )
1237             x264_frame_expand_border_mod16( h, fenc );
1238
1239         fenc->i_frame = h->frames.i_input++;
1240
1241         x264_frame_put( h->frames.next, fenc );
1242
1243         if( h->frames.b_have_lowres )
1244             x264_frame_init_lowres( h->param.cpu, fenc );
1245
1246         if( h->frames.i_input <= h->frames.i_delay )
1247         {
1248             /* Nothing yet to encode */
1249             /* waiting for filling bframe buffer */
1250             pic_out->i_type = X264_TYPE_AUTO;
1251             return 0;
1252         }
1253     }
1254
1255     if( h->frames.current[0] == NULL )
1256     {
1257         int bframes = 0;
1258         /* 2: Select frame types */
1259         if( h->frames.next[0] == NULL )
1260             return 0;
1261
1262         x264_slicetype_decide( h );
1263
1264         /* 3: move some B-frames and 1 non-B to encode queue */
1265         while( IS_X264_TYPE_B( h->frames.next[bframes]->i_type ) )
1266             bframes++;
1267         x264_frame_put( h->frames.current, x264_frame_get( &h->frames.next[bframes] ) );
1268         /* FIXME: when max B-frames > 3, BREF may no longer be centered after GOP closing */
1269         if( h->param.b_bframe_pyramid && bframes > 1 )
1270         {
1271             x264_frame_t *mid = x264_frame_get( &h->frames.next[bframes/2] );
1272             mid->i_type = X264_TYPE_BREF;
1273             x264_frame_put( h->frames.current, mid );
1274             bframes--;
1275         }
1276         while( bframes-- )
1277             x264_frame_put( h->frames.current, x264_frame_get( h->frames.next ) );
1278     }
1279     TIMER_STOP( i_mtime_encode_frame );
1280
1281     /* ------------------- Get frame to be encoded ------------------------- */
1282     /* 4: get picture to encode */
1283     h->fenc = x264_frame_get( h->frames.current );
1284     if( h->fenc == NULL )
1285     {
1286         /* Nothing yet to encode (ex: waiting for I/P with B frames) */
1287         /* waiting for filling bframe buffer */
1288         pic_out->i_type = X264_TYPE_AUTO;
1289         return 0;
1290     }
1291
1292 do_encode:
1293
1294     if( h->fenc->i_type == X264_TYPE_IDR )
1295     {
1296         h->frames.i_last_idr = h->fenc->i_frame;
1297     }
1298
1299     /* ------------------- Setup frame context ----------------------------- */
1300     /* 5: Init data dependant of frame type */
1301     TIMER_START( i_mtime_encode_frame );
1302     if( h->fenc->i_type == X264_TYPE_IDR )
1303     {
1304         /* reset ref pictures */
1305         x264_reference_reset( h );
1306
1307         i_nal_type    = NAL_SLICE_IDR;
1308         i_nal_ref_idc = NAL_PRIORITY_HIGHEST;
1309         i_slice_type = SLICE_TYPE_I;
1310     }
1311     else if( h->fenc->i_type == X264_TYPE_I )
1312     {
1313         i_nal_type    = NAL_SLICE;
1314         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
1315         i_slice_type = SLICE_TYPE_I;
1316     }
1317     else if( h->fenc->i_type == X264_TYPE_P )
1318     {
1319         i_nal_type    = NAL_SLICE;
1320         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
1321         i_slice_type = SLICE_TYPE_P;
1322     }
1323     else if( h->fenc->i_type == X264_TYPE_BREF )
1324     {
1325         i_nal_type    = NAL_SLICE;
1326         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* maybe add MMCO to forget it? -> low */
1327         i_slice_type = SLICE_TYPE_B;
1328     }
1329     else    /* B frame */
1330     {
1331         i_nal_type    = NAL_SLICE;
1332         i_nal_ref_idc = NAL_PRIORITY_DISPOSABLE;
1333         i_slice_type = SLICE_TYPE_B;
1334     }
1335
1336     h->fdec->i_poc =
1337     h->fenc->i_poc = 2 * (h->fenc->i_frame - h->frames.i_last_idr);
1338     h->fdec->i_type = h->fenc->i_type;
1339     h->fdec->i_frame = h->fenc->i_frame;
1340     h->fenc->b_kept_as_ref =
1341     h->fdec->b_kept_as_ref = i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE;
1342
1343
1344
1345     /* ------------------- Init                ----------------------------- */
1346     /* build ref list 0/1 */
1347     x264_reference_build_list( h, h->fdec->i_poc, i_slice_type );
1348
1349     /* Init the rate control */
1350     x264_ratecontrol_start( h, i_slice_type, h->fenc->i_qpplus1 );
1351     i_global_qp = x264_ratecontrol_qp( h );
1352
1353     pic_out->i_qpplus1 =
1354     h->fdec->i_qpplus1 = i_global_qp + 1;
1355
1356     if( i_slice_type == SLICE_TYPE_B )
1357         x264_macroblock_bipred_init( h );
1358
1359     /* ------------------------ Create slice header  ----------------------- */
1360     x264_slice_init( h, i_nal_type, i_slice_type, i_global_qp );
1361
1362     if( h->fenc->b_kept_as_ref )
1363         h->i_frame_num++;
1364
1365     /* ---------------------- Write the bitstream -------------------------- */
1366     /* Init bitstream context */
1367     h->out.i_nal = 0;
1368     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
1369
1370     if(h->param.b_aud){
1371         int pic_type;
1372
1373         if(i_slice_type == SLICE_TYPE_I)
1374             pic_type = 0;
1375         else if(i_slice_type == SLICE_TYPE_P)
1376             pic_type = 1;
1377         else if(i_slice_type == SLICE_TYPE_B)
1378             pic_type = 2;
1379         else
1380             pic_type = 7;
1381
1382         x264_nal_start(h, NAL_AUD, NAL_PRIORITY_DISPOSABLE);
1383         bs_write(&h->out.bs, 3, pic_type);
1384         bs_rbsp_trailing(&h->out.bs);
1385         x264_nal_end(h);
1386     }
1387
1388     h->i_nal_type = i_nal_type;
1389     h->i_nal_ref_idc = i_nal_ref_idc;
1390
1391     /* Write SPS and PPS */
1392     if( i_nal_type == NAL_SLICE_IDR && h->param.b_repeat_headers )
1393     {
1394         if( h->fenc->i_frame == 0 )
1395         {
1396             /* identify ourself */
1397             x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
1398             x264_sei_version_write( h, &h->out.bs );
1399             x264_nal_end( h );
1400         }
1401
1402         /* generate sequence parameters */
1403         x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
1404         x264_sps_write( &h->out.bs, h->sps );
1405         x264_nal_end( h );
1406
1407         /* generate picture parameters */
1408         x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
1409         x264_pps_write( &h->out.bs, h->pps );
1410         x264_nal_end( h );
1411     }
1412
1413     /* Write frame */
1414     i_frame_size = x264_slices_write( h );
1415
1416     /* restore CPU state (before using float again) */
1417     x264_cpu_restore( h->param.cpu );
1418
1419     if( i_slice_type == SLICE_TYPE_P && !h->param.rc.b_stat_read
1420         && h->param.i_scenecut_threshold >= 0 )
1421     {
1422         const int *mbs = h->stat.frame.i_mb_count;
1423         int i_mb_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1424         int i_mb_p = mbs[P_L0] + mbs[P_8x8];
1425         int i_mb_s = mbs[P_SKIP];
1426         int i_mb   = h->sps->i_mb_width * h->sps->i_mb_height;
1427         int64_t i_inter_cost = h->stat.frame.i_inter_cost;
1428         int64_t i_intra_cost = h->stat.frame.i_intra_cost;
1429
1430         float f_bias;
1431         int i_gop_size = h->fenc->i_frame - h->frames.i_last_idr;
1432         float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
1433         /* magic numbers pulled out of thin air */
1434         float f_thresh_min = f_thresh_max * h->param.i_keyint_min
1435                              / ( h->param.i_keyint_max * 4 );
1436         if( h->param.i_keyint_min == h->param.i_keyint_max )
1437              f_thresh_min= f_thresh_max;
1438
1439         /* macroblock_analyse() doesn't further analyse skipped mbs,
1440          * so we have to guess their cost */
1441         if( i_mb_s < i_mb )
1442             i_intra_cost = i_intra_cost * i_mb / (i_mb - i_mb_s);
1443
1444         if( i_gop_size < h->param.i_keyint_min / 4 )
1445             f_bias = f_thresh_min / 4;
1446         else if( i_gop_size <= h->param.i_keyint_min )
1447             f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
1448         else
1449         {
1450             f_bias = f_thresh_min
1451                      + ( f_thresh_max - f_thresh_min )
1452                        * ( i_gop_size - h->param.i_keyint_min )
1453                        / ( h->param.i_keyint_max - h->param.i_keyint_min );
1454         }
1455         f_bias = X264_MIN( f_bias, 1.0 );
1456
1457         /* Bad P will be reencoded as I */
1458         if( i_mb_s < i_mb &&
1459             i_inter_cost >= (1.0 - f_bias) * i_intra_cost )
1460         {
1461             int b;
1462
1463             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",
1464                       h->fenc->i_frame,
1465                       (double)i_intra_cost, (double)i_inter_cost,
1466                       (double)i_inter_cost / i_intra_cost,
1467                       f_bias, i_gop_size,
1468                       i_mb_i, i_mb_p, i_mb_s );
1469
1470             /* Restore frame num */
1471             h->i_frame_num--;
1472
1473             for( b = 0; h->frames.current[b] && IS_X264_TYPE_B( h->frames.current[b]->i_type ); b++ );
1474             if( b > 0 )
1475             {
1476                 /* If using B-frames, force GOP to be closed.
1477                  * Even if this frame is going to be I and not IDR, forcing a
1478                  * P-frame before the scenecut will probably help compression.
1479                  *
1480                  * We don't yet know exactly which frame is the scene cut, so
1481                  * we can't assign an I-frame. Instead, change the previous
1482                  * B-frame to P, and rearrange coding order. */
1483
1484                 if( h->param.b_bframe_adaptive || b > 1 )
1485                     h->fenc->i_type = X264_TYPE_AUTO;
1486                 x264_frame_sort_pts( h->frames.current );
1487                 x264_frame_push( h->frames.next, h->fenc );
1488                 h->fenc = h->frames.current[b-1];
1489                 h->frames.current[b-1] = NULL;
1490                 h->fenc->i_type = X264_TYPE_P;
1491                 x264_frame_sort_dts( h->frames.current );
1492             }
1493             /* Do IDR if needed */
1494             else if( i_gop_size >= h->param.i_keyint_min )
1495             {
1496                 x264_frame_t *tmp;
1497
1498                 /* Reset */
1499                 h->i_frame_num = 0;
1500
1501                 /* Reinit field of fenc */
1502                 h->fenc->i_type = X264_TYPE_IDR;
1503                 h->fenc->i_poc = 0;
1504
1505                 /* Put enqueued frames back in the pool */
1506                 while( (tmp = x264_frame_get( h->frames.current ) ) != NULL )
1507                     x264_frame_put( h->frames.next, tmp );
1508                 x264_frame_sort_pts( h->frames.next );
1509             }
1510             else
1511             {
1512                 h->fenc->i_type = X264_TYPE_I;
1513             }
1514             goto do_encode;
1515         }
1516     }
1517
1518     /* End bitstream, set output  */
1519     *pi_nal = h->out.i_nal;
1520     *pp_nal = h->out.nal;
1521
1522     /* Set output picture properties */
1523     if( i_slice_type == SLICE_TYPE_I )
1524         pic_out->i_type = i_nal_type == NAL_SLICE_IDR ? X264_TYPE_IDR : X264_TYPE_I;
1525     else if( i_slice_type == SLICE_TYPE_P )
1526         pic_out->i_type = X264_TYPE_P;
1527     else
1528         pic_out->i_type = X264_TYPE_B;
1529     pic_out->i_pts = h->fenc->i_pts;
1530
1531     pic_out->img.i_plane = h->fdec->i_plane;
1532     for(i = 0; i < 4; i++){
1533         pic_out->img.i_stride[i] = h->fdec->i_stride[i];
1534         pic_out->img.plane[i] = h->fdec->plane[i];
1535     }
1536
1537     /* ---------------------- Update encoder state ------------------------- */
1538
1539     /* update rc */
1540     x264_cpu_restore( h->param.cpu );
1541     x264_ratecontrol_end( h, i_frame_size * 8 );
1542
1543     /* handle references */
1544     if( i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE )
1545         x264_reference_update( h );
1546 #ifdef DEBUG_DUMP_FRAME
1547     else
1548         x264_fdec_deblock( h );
1549 #endif
1550     x264_frame_put( h->frames.unused, h->fenc );
1551
1552     /* increase frame count */
1553     h->i_frame++;
1554
1555     /* restore CPU state (before using float again) */
1556     x264_cpu_restore( h->param.cpu );
1557
1558     x264_noise_reduction_update( h );
1559
1560     TIMER_STOP( i_mtime_encode_frame );
1561
1562     /* ---------------------- Compute/Print statistics --------------------- */
1563     /* Slice stat */
1564     h->stat.i_slice_count[i_slice_type]++;
1565     h->stat.i_slice_size[i_slice_type] += i_frame_size + NALU_OVERHEAD;
1566     h->stat.i_slice_qp[i_slice_type] += i_global_qp;
1567
1568     for( i = 0; i < X264_MBTYPE_MAX; i++ )
1569         h->stat.i_mb_count[h->sh.i_type][i] += h->stat.frame.i_mb_count[i];
1570     for( i = 0; i < 2; i++ )
1571         h->stat.i_mb_count_8x8dct[i] += h->stat.frame.i_mb_count_8x8dct[i];
1572     if( h->sh.i_type != SLICE_TYPE_I )
1573     {
1574         for( i = 0; i < 7; i++ )
1575             h->stat.i_mb_count_size[h->sh.i_type][i] += h->stat.frame.i_mb_count_size[i];
1576         for( i = 0; i < 16; i++ )
1577             h->stat.i_mb_count_ref[h->sh.i_type][i] += h->stat.frame.i_mb_count_ref[i];
1578     }
1579     if( i_slice_type == SLICE_TYPE_B )
1580     {
1581         h->stat.i_direct_frames[ h->sh.b_direct_spatial_mv_pred ] ++;
1582         if( h->mb.b_direct_auto_write )
1583         {
1584             //FIXME somewhat arbitrary time constants
1585             if( h->stat.i_direct_score[0] + h->stat.i_direct_score[1] > h->mb.i_mb_count )
1586             {
1587                 for( i = 0; i < 2; i++ )
1588                     h->stat.i_direct_score[i] = h->stat.i_direct_score[i] * 9/10;
1589             }
1590             for( i = 0; i < 2; i++ )
1591                 h->stat.i_direct_score[i] += h->stat.frame.i_direct_score[i];
1592         }
1593     }
1594
1595     psz_message[0] = '\0';
1596     if( h->param.analyse.b_psnr )
1597     {
1598         int64_t sqe[3];
1599
1600         for( i=0; i<3; i++ )
1601         {
1602             sqe[i] = x264_pixel_ssd_wxh( &h->pixf,
1603                          frame_psnr->plane[i], frame_psnr->i_stride[i],
1604                          h->fenc->plane[i], h->fenc->i_stride[i],
1605                          h->param.i_width >> !!i, h->param.i_height >> !!i );
1606         }
1607         x264_cpu_restore( h->param.cpu );
1608
1609         h->stat.i_sqe_global[i_slice_type] += sqe[0] + sqe[1] + sqe[2];
1610         h->stat.f_psnr_average[i_slice_type] += x264_psnr( sqe[0] + sqe[1] + sqe[2], 3 * h->param.i_width * h->param.i_height / 2 );
1611         h->stat.f_psnr_mean_y[i_slice_type] += x264_psnr( sqe[0], h->param.i_width * h->param.i_height );
1612         h->stat.f_psnr_mean_u[i_slice_type] += x264_psnr( sqe[1], h->param.i_width * h->param.i_height / 4 );
1613         h->stat.f_psnr_mean_v[i_slice_type] += x264_psnr( sqe[2], h->param.i_width * h->param.i_height / 4 );
1614
1615         snprintf( psz_message, 80, " PSNR Y:%5.2f U:%5.2f V:%5.2f",
1616                   x264_psnr( sqe[0], h->param.i_width * h->param.i_height ),
1617                   x264_psnr( sqe[1], h->param.i_width * h->param.i_height / 4),
1618                   x264_psnr( sqe[2], h->param.i_width * h->param.i_height / 4) );
1619     }
1620
1621     if( h->param.analyse.b_ssim )
1622     {
1623         // offset by 2 pixels to avoid alignment of ssim blocks with dct blocks
1624         float ssim_y = x264_pixel_ssim_wxh( &h->pixf,
1625                          frame_psnr->plane[0] + 2+2*frame_psnr->i_stride[0], frame_psnr->i_stride[0],
1626                          h->fenc->plane[0] + 2+2*h->fenc->i_stride[0], h->fenc->i_stride[0],
1627                          h->param.i_width-2, h->param.i_height-2 );
1628         h->stat.f_ssim_mean_y[i_slice_type] += ssim_y;
1629         snprintf( psz_message + strlen(psz_message), 80 - strlen(psz_message),
1630                   " SSIM Y:%.5f", ssim_y );
1631     }
1632     psz_message[79] = '\0';
1633
1634     x264_log( h, X264_LOG_DEBUG,
1635                   "frame=%4d QP=%i NAL=%d Slice:%c Poc:%-3d I:%-4d P:%-4d SKIP:%-4d size=%d bytes%s\n",
1636               h->i_frame - 1,
1637               i_global_qp,
1638               i_nal_ref_idc,
1639               i_slice_type == SLICE_TYPE_I ? 'I' : (i_slice_type == SLICE_TYPE_P ? 'P' : 'B' ),
1640               frame_psnr->i_poc,
1641               h->stat.frame.i_mb_count_i,
1642               h->stat.frame.i_mb_count_p,
1643               h->stat.frame.i_mb_count_skip,
1644               i_frame_size,
1645               psz_message );
1646
1647
1648 #ifdef DEBUG_MB_TYPE
1649 {
1650     static const char mb_chars[] = { 'i', 'i', 'I', 'C', 'P', '8', 'S',
1651         'D', '<', 'X', 'B', 'X', '>', 'B', 'B', 'B', 'B', '8', 'S' };
1652     int mb_xy;
1653     for( mb_xy = 0; mb_xy < h->sps->i_mb_width * h->sps->i_mb_height; mb_xy++ )
1654     {
1655         if( h->mb.type[mb_xy] < X264_MBTYPE_MAX && h->mb.type[mb_xy] >= 0 )
1656             fprintf( stderr, "%c ", mb_chars[ h->mb.type[mb_xy] ] );
1657         else
1658             fprintf( stderr, "? " );
1659
1660         if( (mb_xy+1) % h->sps->i_mb_width == 0 )
1661             fprintf( stderr, "\n" );
1662     }
1663 }
1664 #endif
1665
1666 #ifdef DEBUG_DUMP_FRAME
1667     /* Dump reconstructed frame */
1668     x264_frame_dump( h, frame_psnr, "fdec.yuv" );
1669 #endif
1670     return 0;
1671 }
1672
1673 /****************************************************************************
1674  * x264_encoder_close:
1675  ****************************************************************************/
1676 void    x264_encoder_close  ( x264_t *h )
1677 {
1678 #ifdef DEBUG_BENCHMARK
1679     int64_t i_mtime_total = i_mtime_analyse + i_mtime_encode + i_mtime_write + i_mtime_filter + 1;
1680 #endif
1681     int64_t i_yuv_size = 3 * h->param.i_width * h->param.i_height / 2;
1682     int i;
1683
1684 #ifdef DEBUG_BENCHMARK
1685     x264_log( h, X264_LOG_INFO,
1686               "analyse=%d(%lldms) encode=%d(%lldms) write=%d(%lldms) filter=%d(%lldms)\n",
1687               (int)(100*i_mtime_analyse/i_mtime_total), i_mtime_analyse/1000,
1688               (int)(100*i_mtime_encode/i_mtime_total), i_mtime_encode/1000,
1689               (int)(100*i_mtime_write/i_mtime_total), i_mtime_write/1000,
1690               (int)(100*i_mtime_filter/i_mtime_total), i_mtime_filter/1000 );
1691 #endif
1692
1693     /* Slices used and PSNR */
1694     for( i=0; i<5; i++ )
1695     {
1696         static const int slice_order[] = { SLICE_TYPE_I, SLICE_TYPE_SI, SLICE_TYPE_P, SLICE_TYPE_SP, SLICE_TYPE_B };
1697         static const char *slice_name[] = { "P", "B", "I", "SP", "SI" };
1698         int i_slice = slice_order[i];
1699
1700         if( h->stat.i_slice_count[i_slice] > 0 )
1701         {
1702             const int i_count = h->stat.i_slice_count[i_slice];
1703             if( h->param.analyse.b_psnr )
1704             {
1705                 x264_log( h, X264_LOG_INFO,
1706                           "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",
1707                           slice_name[i_slice],
1708                           i_count,
1709                           (double)h->stat.i_slice_qp[i_slice] / i_count,
1710                           (double)h->stat.i_slice_size[i_slice] / i_count,
1711                           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,
1712                           h->stat.f_psnr_average[i_slice] / i_count,
1713                           x264_psnr( h->stat.i_sqe_global[i_slice], i_count * i_yuv_size ) );
1714             }
1715             else
1716             {
1717                 x264_log( h, X264_LOG_INFO,
1718                           "slice %s:%-5d Avg QP:%5.2f  size:%6.0f\n",
1719                           slice_name[i_slice],
1720                           i_count,
1721                           (double)h->stat.i_slice_qp[i_slice] / i_count,
1722                           (double)h->stat.i_slice_size[i_slice] / i_count );
1723             }
1724         }
1725     }
1726
1727     /* MB types used */
1728     if( h->stat.i_slice_count[SLICE_TYPE_I] > 0 )
1729     {
1730         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_I];
1731         const double i_count = h->stat.i_slice_count[SLICE_TYPE_I] * h->mb.i_mb_count / 100.0;
1732         x264_log( h, X264_LOG_INFO,
1733                   "mb I  I16..4: %4.1f%% %4.1f%% %4.1f%%\n",
1734                   i_mb_count[I_16x16]/ i_count,
1735                   i_mb_count[I_8x8]  / i_count,
1736                   i_mb_count[I_4x4]  / i_count );
1737     }
1738     if( h->stat.i_slice_count[SLICE_TYPE_P] > 0 )
1739     {
1740         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_P];
1741         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_P];
1742         const double i_count = h->stat.i_slice_count[SLICE_TYPE_P] * h->mb.i_mb_count / 100.0;
1743         x264_log( h, X264_LOG_INFO,
1744                   "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",
1745                   i_mb_count[I_16x16]/ i_count,
1746                   i_mb_count[I_8x8]  / i_count,
1747                   i_mb_count[I_4x4]  / i_count,
1748                   i_mb_size[PIXEL_16x16] / (i_count*4),
1749                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
1750                   i_mb_size[PIXEL_8x8] / (i_count*4),
1751                   (i_mb_size[PIXEL_8x4] + i_mb_size[PIXEL_4x8]) / (i_count*4),
1752                   i_mb_size[PIXEL_4x4] / (i_count*4),
1753                   i_mb_count[P_SKIP] / i_count );
1754     }
1755     if( h->stat.i_slice_count[SLICE_TYPE_B] > 0 )
1756     {
1757         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_B];
1758         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_B];
1759         const double i_count = h->stat.i_slice_count[SLICE_TYPE_B] * h->mb.i_mb_count / 100.0;
1760         x264_log( h, X264_LOG_INFO,
1761                   "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",
1762                   i_mb_count[I_16x16]  / i_count,
1763                   i_mb_count[I_8x8]    / i_count,
1764                   i_mb_count[I_4x4]    / i_count,
1765                   i_mb_size[PIXEL_16x16] / (i_count*4),
1766                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
1767                   i_mb_size[PIXEL_8x8] / (i_count*4),
1768                   i_mb_count[B_DIRECT] / i_count,
1769                   i_mb_count[B_SKIP]   / i_count );
1770     }
1771
1772     x264_ratecontrol_summary( h );
1773
1774     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 )
1775     {
1776         const int i_count = h->stat.i_slice_count[SLICE_TYPE_I] +
1777                             h->stat.i_slice_count[SLICE_TYPE_P] +
1778                             h->stat.i_slice_count[SLICE_TYPE_B];
1779         float fps = (float) h->param.i_fps_num / h->param.i_fps_den;
1780 #define SUM3(p) (p[SLICE_TYPE_I] + p[SLICE_TYPE_P] + p[SLICE_TYPE_B])
1781 #define SUM3b(p,o) (p[SLICE_TYPE_I][o] + p[SLICE_TYPE_P][o] + p[SLICE_TYPE_B][o])
1782         float f_bitrate = fps * SUM3(h->stat.i_slice_size) / i_count / 125;
1783
1784         if( h->pps->b_transform_8x8_mode )
1785         {
1786             int64_t i_i8x8 = SUM3b( h->stat.i_mb_count, I_8x8 );
1787             int64_t i_intra = i_i8x8 + SUM3b( h->stat.i_mb_count, I_4x4 )
1788                                      + SUM3b( h->stat.i_mb_count, I_16x16 );
1789             x264_log( h, X264_LOG_INFO, "8x8 transform  intra:%.1f%%  inter:%.1f%%\n",
1790                       100. * i_i8x8 / i_intra,
1791                       100. * h->stat.i_mb_count_8x8dct[1] / h->stat.i_mb_count_8x8dct[0] );
1792         }
1793
1794         if( h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
1795             && h->stat.i_slice_count[SLICE_TYPE_B] )
1796         {
1797             x264_log( h, X264_LOG_INFO, "direct mvs  spatial:%.1f%%  temporal:%.1f%%\n",
1798                       h->stat.i_direct_frames[1] * 100. / h->stat.i_slice_count[SLICE_TYPE_B],
1799                       h->stat.i_direct_frames[0] * 100. / h->stat.i_slice_count[SLICE_TYPE_B] );
1800         }
1801
1802         if( h->frames.i_max_ref0 > 1 )
1803         {
1804             int i_slice;
1805             for( i_slice = 0; i_slice < 2; i_slice++ )
1806             {
1807                 char buf[200];
1808                 char *p = buf;
1809                 int64_t i_den = 0;
1810                 int i_max = 0;
1811                 for( i = 0; i < h->frames.i_max_ref0; i++ )
1812                     if( h->stat.i_mb_count_ref[i_slice][i] )
1813                     {
1814                         i_den += h->stat.i_mb_count_ref[i_slice][i];
1815                         i_max = i;
1816                     }
1817                 if( i_max == 0 )
1818                     continue;
1819                 for( i = 0; i <= i_max; i++ )
1820                     p += sprintf( p, " %4.1f%%", 100. * h->stat.i_mb_count_ref[i_slice][i] / i_den );
1821                 x264_log( h, X264_LOG_INFO, "ref %c %s\n", i_slice==SLICE_TYPE_P ? 'P' : 'B', buf );
1822             }
1823         }
1824
1825         if( h->param.analyse.b_ssim )
1826         {
1827             x264_log( h, X264_LOG_INFO,
1828                       "SSIM Mean Y:%.7f\n",
1829                       SUM3( h->stat.f_ssim_mean_y ) / i_count );
1830         }
1831         if( h->param.analyse.b_psnr )
1832         {
1833             x264_log( h, X264_LOG_INFO,
1834                       "PSNR Mean Y:%6.3f U:%6.3f V:%6.3f Avg:%6.3f Global:%6.3f kb/s:%.2f\n",
1835                       SUM3( h->stat.f_psnr_mean_y ) / i_count,
1836                       SUM3( h->stat.f_psnr_mean_u ) / i_count,
1837                       SUM3( h->stat.f_psnr_mean_v ) / i_count,
1838                       SUM3( h->stat.f_psnr_average ) / i_count,
1839                       x264_psnr( SUM3( h->stat.i_sqe_global ), i_count * i_yuv_size ),
1840                       f_bitrate );
1841         }
1842         else
1843             x264_log( h, X264_LOG_INFO, "kb/s:%.1f\n", f_bitrate );
1844     }
1845
1846     /* frames */
1847     for( i = 0; i < X264_BFRAME_MAX + 3; i++ )
1848     {
1849         if( h->frames.current[i] ) x264_frame_delete( h->frames.current[i] );
1850         if( h->frames.next[i] )    x264_frame_delete( h->frames.next[i] );
1851         if( h->frames.unused[i] )  x264_frame_delete( h->frames.unused[i] );
1852     }
1853     /* ref frames */
1854     for( i = 0; i < h->frames.i_max_dpb; i++ )
1855     {
1856         x264_frame_delete( h->frames.reference[i] );
1857     }
1858
1859     /* rc */
1860     x264_ratecontrol_delete( h );
1861
1862     /* param */
1863     if( h->param.rc.psz_stat_out )
1864         free( h->param.rc.psz_stat_out );
1865     if( h->param.rc.psz_stat_in )
1866         free( h->param.rc.psz_stat_in );
1867     if( h->param.rc.psz_rc_eq )
1868         free( h->param.rc.psz_rc_eq );
1869
1870     x264_cqm_delete( h );
1871     x264_macroblock_cache_end( h );
1872     x264_free( h->out.p_bitstream );
1873     for( i = 1; i < h->param.i_threads; i++ )
1874         x264_free( h->thread[i] );
1875     x264_free( h );
1876 }