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 static void x264_encoder_frame_end( x264_t *h, x264_t *thread_current,
  65                                     x264_nal_t **pp_nal, int *pi_nal,
  66                                     x264_picture_t *pic_out );
  67
  68 /****************************************************************************
  69  *
  70  ******************************* x264 libs **********************************
  71  *
  72  ****************************************************************************/
  73 static float x264_psnr( int64_t i_sqe, int64_t i_size )
  74 {
  75     double f_mse = (double)i_sqe / ((double)65025.0 * (double)i_size);
  76     if( f_mse <= 0.0000000001 ) /* Max 100dB */
  77         return 100;
  78
  79     return (float)(-10.0 * log( f_mse ) / log( 10.0 ));
  80 }
  81
  82 #ifdef DEBUG_DUMP_FRAME
  83 static void x264_frame_dump( x264_t *h, x264_frame_t *fr, char *name )
  84 {
  85     FILE *f = fopen( name, "r+b" );
  86     int i, y;
  87     if( !f )
  88         return;
  89
  90     /* Write the frame in display order */
  91     fseek( f, fr->i_frame * h->param.i_height * h->param.i_width * 3 / 2, SEEK_SET );
  92
  93     for( i = 0; i < fr->i_plane; i++ )
  94     {
  95         for( y = 0; y < h->param.i_height / ( i == 0 ? 1 : 2 ); y++ )
  96         {
  97             fwrite( &fr->plane[i][y*fr->i_stride[i]], 1, h->param.i_width / ( i == 0 ? 1 : 2 ), f );
  98         }
  99     }
 100     fclose( f );
 101 }
 102 #endif
 103
 104
 105 /* Fill "default" values */
 106 static void x264_slice_header_init( x264_t *h, x264_slice_header_t *sh,
 107                                     x264_sps_t *sps, x264_pps_t *pps,
 108                                     int i_idr_pic_id, int i_frame, int i_qp )
 109 {
 110     x264_param_t *param = &h->param;
 111     int i;
 112
 113     /* First we fill all field */
 114     sh->sps = sps;
 115     sh->pps = pps;
 116
 117     sh->i_first_mb  = 0;
 118     sh->i_last_mb   = h->sps->i_mb_width * h->sps->i_mb_height;
 119     sh->i_pps_id    = pps->i_id;
 120
 121     sh->i_frame_num = i_frame;
 122
 123     sh->b_mbaff = h->param.b_interlaced;
 124     sh->b_field_pic = 0;    /* no field support for now */
 125     sh->b_bottom_field = 0; /* not yet used */
 126
 127     sh->i_idr_pic_id = i_idr_pic_id;
 128
 129     /* poc stuff, fixed later */
 130     sh->i_poc_lsb = 0;
 131     sh->i_delta_poc_bottom = 0;
 132     sh->i_delta_poc[0] = 0;
 133     sh->i_delta_poc[1] = 0;
 134
 135     sh->i_redundant_pic_cnt = 0;
 136
 137     if( !h->mb.b_direct_auto_read )
 138     {
 139         if( h->mb.b_direct_auto_write )
 140             sh->b_direct_spatial_mv_pred = ( h->stat.i_direct_score[1] > h->stat.i_direct_score[0] );
 141         else
 142             sh->b_direct_spatial_mv_pred = ( param->analyse.i_direct_mv_pred == X264_DIRECT_PRED_SPATIAL );
 143     }
 144     /* else b_direct_spatial_mv_pred was read from the 2pass statsfile */
 145
 146     sh->b_num_ref_idx_override = 0;
 147     sh->i_num_ref_idx_l0_active = 1;
 148     sh->i_num_ref_idx_l1_active = 1;
 149
 150     sh->b_ref_pic_list_reordering_l0 = h->b_ref_reorder[0];
 151     sh->b_ref_pic_list_reordering_l1 = h->b_ref_reorder[1];
 152
 153     /* If the ref list isn't in the default order, construct reordering header */
 154     /* List1 reordering isn't needed yet */
 155     if( sh->b_ref_pic_list_reordering_l0 )
 156     {
 157         int pred_frame_num = i_frame;
 158         for( i = 0; i < h->i_ref0; i++ )
 159         {
 160             int diff = h->fref0[i]->i_frame_num - pred_frame_num;
 161             if( diff == 0 )
 162                 x264_log( h, X264_LOG_ERROR, "diff frame num == 0\n" );
 163             sh->ref_pic_list_order[0][i].idc = ( diff > 0 );
 164             sh->ref_pic_list_order[0][i].arg = abs( diff ) - 1;
 165             pred_frame_num = h->fref0[i]->i_frame_num;
 166         }
 167     }
 168
 169     sh->i_cabac_init_idc = param->i_cabac_init_idc;
 170
 171     sh->i_qp = i_qp;
 172     sh->i_qp_delta = i_qp - pps->i_pic_init_qp;
 173     sh->b_sp_for_swidth = 0;
 174     sh->i_qs_delta = 0;
 175
 176     /* If effective qp <= 15, deblocking would have no effect anyway */
 177     if( param->b_deblocking_filter
 178         && ( h->mb.b_variable_qp
 179         || 15 < i_qp + 2 * X264_MAX(param->i_deblocking_filter_alphac0, param->i_deblocking_filter_beta) ) )
 180     {
 181         sh->i_disable_deblocking_filter_idc = 0;
 182     }
 183     else
 184     {
 185         sh->i_disable_deblocking_filter_idc = 1;
 186     }
 187     sh->i_alpha_c0_offset = param->i_deblocking_filter_alphac0 << 1;
 188     sh->i_beta_offset = param->i_deblocking_filter_beta << 1;
 189 }
 190
 191 static void x264_slice_header_write( bs_t *s, x264_slice_header_t *sh, int i_nal_ref_idc )
 192 {
 193     int i;
 194
 195     if( sh->b_mbaff )
 196     {
 197         assert( sh->i_first_mb % (2*sh->sps->i_mb_width) == 0 );
 198         bs_write_ue( s, sh->i_first_mb >> 1 );
 199     }
 200     else
 201         bs_write_ue( s, sh->i_first_mb );
 202
 203     bs_write_ue( s, sh->i_type + 5 );   /* same type things */
 204     bs_write_ue( s, sh->i_pps_id );
 205     bs_write( s, sh->sps->i_log2_max_frame_num, sh->i_frame_num );
 206
 207     if( !sh->sps->b_frame_mbs_only )
 208     {
 209         bs_write1( s, sh->b_field_pic );
 210         if ( sh->b_field_pic )
 211             bs_write1( s, sh->b_bottom_field );
 212     }
 213
 214     if( sh->i_idr_pic_id >= 0 ) /* NAL IDR */
 215     {
 216         bs_write_ue( s, sh->i_idr_pic_id );
 217     }
 218
 219     if( sh->sps->i_poc_type == 0 )
 220     {
 221         bs_write( s, sh->sps->i_log2_max_poc_lsb, sh->i_poc_lsb );
 222         if( sh->pps->b_pic_order && !sh->b_field_pic )
 223         {
 224             bs_write_se( s, sh->i_delta_poc_bottom );
 225         }
 226     }
 227     else if( sh->sps->i_poc_type == 1 && !sh->sps->b_delta_pic_order_always_zero )
 228     {
 229         bs_write_se( s, sh->i_delta_poc[0] );
 230         if( sh->pps->b_pic_order && !sh->b_field_pic )
 231         {
 232             bs_write_se( s, sh->i_delta_poc[1] );
 233         }
 234     }
 235
 236     if( sh->pps->b_redundant_pic_cnt )
 237     {
 238         bs_write_ue( s, sh->i_redundant_pic_cnt );
 239     }
 240
 241     if( sh->i_type == SLICE_TYPE_B )
 242     {
 243         bs_write1( s, sh->b_direct_spatial_mv_pred );
 244     }
 245     if( sh->i_type == SLICE_TYPE_P || sh->i_type == SLICE_TYPE_SP || sh->i_type == SLICE_TYPE_B )
 246     {
 247         bs_write1( s, sh->b_num_ref_idx_override );
 248         if( sh->b_num_ref_idx_override )
 249         {
 250             bs_write_ue( s, sh->i_num_ref_idx_l0_active - 1 );
 251             if( sh->i_type == SLICE_TYPE_B )
 252             {
 253                 bs_write_ue( s, sh->i_num_ref_idx_l1_active - 1 );
 254             }
 255         }
 256     }
 257
 258     /* ref pic list reordering */
 259     if( sh->i_type != SLICE_TYPE_I )
 260     {
 261         bs_write1( s, sh->b_ref_pic_list_reordering_l0 );
 262         if( sh->b_ref_pic_list_reordering_l0 )
 263         {
 264             for( i = 0; i < sh->i_num_ref_idx_l0_active; i++ )
 265             {
 266                 bs_write_ue( s, sh->ref_pic_list_order[0][i].idc );
 267                 bs_write_ue( s, sh->ref_pic_list_order[0][i].arg );
 268
 269             }
 270             bs_write_ue( s, 3 );
 271         }
 272     }
 273     if( sh->i_type == SLICE_TYPE_B )
 274     {
 275         bs_write1( s, sh->b_ref_pic_list_reordering_l1 );
 276         if( sh->b_ref_pic_list_reordering_l1 )
 277         {
 278             for( i = 0; i < sh->i_num_ref_idx_l1_active; i++ )
 279             {
 280                 bs_write_ue( s, sh->ref_pic_list_order[1][i].idc );
 281                 bs_write_ue( s, sh->ref_pic_list_order[1][i].arg );
 282             }
 283             bs_write_ue( s, 3 );
 284         }
 285     }
 286
 287     if( ( sh->pps->b_weighted_pred && ( sh->i_type == SLICE_TYPE_P || sh->i_type == SLICE_TYPE_SP ) ) ||
 288         ( sh->pps->b_weighted_bipred == 1 && sh->i_type == SLICE_TYPE_B ) )
 289     {
 290         /* FIXME */
 291     }
 292
 293     if( i_nal_ref_idc != 0 )
 294     {
 295         if( sh->i_idr_pic_id >= 0 )
 296         {
 297             bs_write1( s, 0 );  /* no output of prior pics flag */
 298             bs_write1( s, 0 );  /* long term reference flag */
 299         }
 300         else
 301         {
 302             bs_write1( s, 0 );  /* adaptive_ref_pic_marking_mode_flag */
 303         }
 304     }
 305
 306     if( sh->pps->b_cabac && sh->i_type != SLICE_TYPE_I )
 307     {
 308         bs_write_ue( s, sh->i_cabac_init_idc );
 309     }
 310     bs_write_se( s, sh->i_qp_delta );      /* slice qp delta */
 311
 312     if( sh->pps->b_deblocking_filter_control )
 313     {
 314         bs_write_ue( s, sh->i_disable_deblocking_filter_idc );
 315         if( sh->i_disable_deblocking_filter_idc != 1 )
 316         {
 317             bs_write_se( s, sh->i_alpha_c0_offset >> 1 );
 318             bs_write_se( s, sh->i_beta_offset >> 1 );
 319         }
 320     }
 321 }
 322
 323 /****************************************************************************
 324  *
 325  ****************************************************************************
 326  ****************************** External API*********************************
 327  ****************************************************************************
 328  *
 329  ****************************************************************************/
 330
 331 static int x264_validate_parameters( x264_t *h )
 332 {
 333     if( h->param.i_width <= 0 || h->param.i_height <= 0 )
 334     {
 335         x264_log( h, X264_LOG_ERROR, "invalid width x height (%dx%d)\n",
 336                   h->param.i_width, h->param.i_height );
 337         return -1;
 338     }
 339
 340     if( h->param.i_width % 2 || h->param.i_height % 2 )
 341     {
 342         x264_log( h, X264_LOG_ERROR, "width or height not divisible by 2 (%dx%d)\n",
 343                   h->param.i_width, h->param.i_height );
 344         return -1;
 345     }
 346     if( h->param.i_csp != X264_CSP_I420 )
 347     {
 348         x264_log( h, X264_LOG_ERROR, "invalid CSP (only I420 supported)\n" );
 349         return -1;
 350     }
 351
 352     if( h->param.i_threads == 0 )
 353         h->param.i_threads = x264_cpu_num_processors() * 3/2;
 354     h->param.i_threads = x264_clip3( h->param.i_threads, 1, X264_THREAD_MAX );
 355     h->param.i_threads = X264_MIN( h->param.i_threads, 1 + (h->param.i_height >> h->param.b_interlaced) / (X264_THREAD_HEIGHT + 16) ); // FIXME exact limit?
 356     if( h->param.i_threads > 1 )
 357     {
 358 #ifndef HAVE_PTHREAD
 359         x264_log( h, X264_LOG_WARNING, "not compiled with pthread support!\n");
 360         h->param.i_threads = 1;
 361 #else
 362         if( h->param.analyse.i_me_method == X264_ME_ESA )
 363         {
 364             x264_log( h, X264_LOG_WARNING, "threads are not yet compatible with ESA\n");
 365             h->param.analyse.i_me_method = X264_ME_UMH;
 366         }
 367         if( h->param.i_scenecut_threshold >= 0 )
 368             h->param.b_pre_scenecut = 1;
 369 #endif
 370     }
 371
 372     if( h->param.b_interlaced )
 373     {
 374         if( h->param.analyse.i_me_method == X264_ME_ESA )
 375         {
 376             x264_log( h, X264_LOG_WARNING, "interlace + me=esa is not implemented\n" );
 377             h->param.analyse.i_me_method = X264_ME_UMH;
 378         }
 379         if( h->param.analyse.i_direct_mv_pred > X264_DIRECT_PRED_SPATIAL )
 380         {
 381             x264_log( h, X264_LOG_WARNING, "interlace + direct=temporal is not implemented\n" );
 382             h->param.analyse.i_direct_mv_pred = X264_DIRECT_PRED_SPATIAL;
 383         }
 384     }
 385
 386     if( h->param.rc.i_rc_method < 0 || h->param.rc.i_rc_method > 2 )
 387     {
 388         x264_log( h, X264_LOG_ERROR, "invalid RC method\n" );
 389         return -1;
 390     }
 391     h->param.rc.f_rf_constant = x264_clip3f( h->param.rc.f_rf_constant, 0, 51 );
 392     h->param.rc.i_qp_constant = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
 393     if( h->param.rc.i_rc_method == X264_RC_CRF )
 394         h->param.rc.i_qp_constant = h->param.rc.f_rf_constant;
 395     if( (h->param.rc.i_rc_method == X264_RC_CQP || h->param.rc.i_rc_method == X264_RC_CRF)
 396         && h->param.rc.i_qp_constant == 0 )
 397     {
 398         h->mb.b_lossless = 1;
 399         h->param.i_cqm_preset = X264_CQM_FLAT;
 400         h->param.psz_cqm_file = NULL;
 401         h->param.rc.i_rc_method = X264_RC_CQP;
 402         h->param.rc.f_ip_factor = 1;
 403         h->param.rc.f_pb_factor = 1;
 404         h->param.analyse.b_transform_8x8 = 0;
 405         h->param.analyse.b_psnr = 0;
 406         h->param.analyse.b_ssim = 0;
 407         h->param.analyse.i_chroma_qp_offset = 0;
 408         h->param.analyse.i_trellis = 0;
 409         h->param.analyse.b_fast_pskip = 0;
 410         h->param.analyse.i_noise_reduction = 0;
 411         h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 1, 6 );
 412     }
 413
 414     if( ( h->param.i_width % 16 || h->param.i_height % 16 ) && !h->mb.b_lossless )
 415     {
 416         x264_log( h, X264_LOG_WARNING,
 417                   "width or height not divisible by 16 (%dx%d), compression will suffer.\n",
 418                   h->param.i_width, h->param.i_height );
 419     }
 420
 421     h->param.i_frame_reference = x264_clip3( h->param.i_frame_reference, 1, 16 );
 422     if( h->param.i_keyint_max <= 0 )
 423         h->param.i_keyint_max = 1;
 424     h->param.i_keyint_min = x264_clip3( h->param.i_keyint_min, 1, h->param.i_keyint_max/2+1 );
 425
 426     h->param.i_bframe = x264_clip3( h->param.i_bframe, 0, X264_BFRAME_MAX );
 427     h->param.i_bframe_bias = x264_clip3( h->param.i_bframe_bias, -90, 100 );
 428     h->param.b_bframe_pyramid = h->param.b_bframe_pyramid && h->param.i_bframe > 1;
 429     h->param.b_bframe_adaptive = h->param.b_bframe_adaptive && h->param.i_bframe > 0;
 430     h->param.analyse.b_weighted_bipred = h->param.analyse.b_weighted_bipred && h->param.i_bframe > 0;
 431     h->mb.b_direct_auto_write = h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
 432                                 && h->param.i_bframe
 433                                 && ( h->param.rc.b_stat_write || !h->param.rc.b_stat_read );
 434
 435     h->param.i_deblocking_filter_alphac0 = x264_clip3( h->param.i_deblocking_filter_alphac0, -6, 6 );
 436     h->param.i_deblocking_filter_beta    = x264_clip3( h->param.i_deblocking_filter_beta, -6, 6 );
 437     h->param.analyse.i_luma_deadzone[0] = x264_clip3( h->param.analyse.i_luma_deadzone[0], 0, 32 );
 438     h->param.analyse.i_luma_deadzone[1] = x264_clip3( h->param.analyse.i_luma_deadzone[1], 0, 32 );
 439     h->mb.i_luma_deadzone[0] = 32 - h->param.analyse.i_luma_deadzone[0];
 440     h->mb.i_luma_deadzone[1] = 32 - h->param.analyse.i_luma_deadzone[1];
 441
 442     h->param.i_cabac_init_idc = x264_clip3( h->param.i_cabac_init_idc, 0, 2 );
 443
 444     if( h->param.i_cqm_preset < X264_CQM_FLAT || h->param.i_cqm_preset > X264_CQM_CUSTOM )
 445         h->param.i_cqm_preset = X264_CQM_FLAT;
 446
 447     if( h->param.analyse.i_me_method < X264_ME_DIA ||
 448         h->param.analyse.i_me_method > X264_ME_ESA )
 449         h->param.analyse.i_me_method = X264_ME_HEX;
 450     if( h->param.analyse.i_me_range < 4 )
 451         h->param.analyse.i_me_range = 4;
 452     if( h->param.analyse.i_me_range > 16 && h->param.analyse.i_me_method <= X264_ME_HEX )
 453         h->param.analyse.i_me_range = 16;
 454     h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 1, 7 );
 455     h->param.analyse.b_bframe_rdo = h->param.analyse.b_bframe_rdo && h->param.analyse.i_subpel_refine >= 6;
 456     h->param.analyse.b_mixed_references = h->param.analyse.b_mixed_references && h->param.i_frame_reference > 1;
 457     h->param.analyse.inter &= X264_ANALYSE_PSUB16x16|X264_ANALYSE_PSUB8x8|X264_ANALYSE_BSUB16x16|
 458                               X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 459     h->param.analyse.intra &= X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 460     if( !(h->param.analyse.inter & X264_ANALYSE_PSUB16x16) )
 461         h->param.analyse.inter &= ~X264_ANALYSE_PSUB8x8;
 462     if( !h->param.analyse.b_transform_8x8 )
 463     {
 464         h->param.analyse.inter &= ~X264_ANALYSE_I8x8;
 465         h->param.analyse.intra &= ~X264_ANALYSE_I8x8;
 466     }
 467     h->param.analyse.i_chroma_qp_offset = x264_clip3(h->param.analyse.i_chroma_qp_offset, -12, 12);
 468     if( !h->param.b_cabac )
 469         h->param.analyse.i_trellis = 0;
 470     h->param.analyse.i_trellis = x264_clip3( h->param.analyse.i_trellis, 0, 2 );
 471     h->param.analyse.i_noise_reduction = x264_clip3( h->param.analyse.i_noise_reduction, 0, 1<<16 );
 472
 473     {
 474         const x264_level_t *l = x264_levels;
 475         while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
 476             l++;
 477         if( l->level_idc == 0 )
 478         {
 479             x264_log( h, X264_LOG_ERROR, "invalid level_idc: %d\n", h->param.i_level_idc );
 480             return -1;
 481         }
 482         if( h->param.analyse.i_mv_range <= 0 )
 483             h->param.analyse.i_mv_range = l->mv_range;
 484         else
 485             h->param.analyse.i_mv_range = x264_clip3(h->param.analyse.i_mv_range, 32, 2048);
 486         if( h->param.analyse.i_direct_8x8_inference < 0 )
 487             h->param.analyse.i_direct_8x8_inference = l->direct8x8;
 488     }
 489
 490     if( h->param.i_threads > 1 )
 491     {
 492         int r = h->param.analyse.i_mv_range_thread;
 493         int r2;
 494         if( r <= 0 )
 495         {
 496             // half of the available space is reserved and divided evenly among the threads,
 497             // the rest is allocated to whichever thread is far enough ahead to use it.
 498             // reserving more space increases quality for some videos, but costs more time
 499             // in thread synchronization.
 500             int max_range = (h->param.i_height + X264_THREAD_HEIGHT) / h->param.i_threads - X264_THREAD_HEIGHT;
 501             r = max_range / 2;
 502         }
 503         r = X264_MAX( r, h->param.analyse.i_me_range );
 504         r = X264_MIN( r, h->param.analyse.i_mv_range );
 505         // round up to use the whole mb row
 506         r2 = (r & ~15) + ((-X264_THREAD_HEIGHT) & 15);
 507         if( r2 < r )
 508             r2 += 16;
 509         x264_log( h, X264_LOG_DEBUG, "using mv_range_thread = %d\n", r2 );
 510         h->param.analyse.i_mv_range_thread = r2;
 511     }
 512
 513     if( h->param.rc.f_qblur < 0 )
 514         h->param.rc.f_qblur = 0;
 515     if( h->param.rc.f_complexity_blur < 0 )
 516         h->param.rc.f_complexity_blur = 0;
 517
 518     h->param.i_sps_id &= 31;
 519
 520     if( h->param.i_log_level < X264_LOG_INFO )
 521     {
 522         h->param.analyse.b_psnr = 0;
 523         h->param.analyse.b_ssim = 0;
 524     }
 525
 526     /* ensure the booleans are 0 or 1 so they can be used in math */
 527 #define BOOLIFY(x) h->param.x = !!h->param.x
 528     BOOLIFY( b_cabac );
 529     BOOLIFY( b_deblocking_filter );
 530     BOOLIFY( b_interlaced );
 531     BOOLIFY( analyse.b_transform_8x8 );
 532     BOOLIFY( analyse.i_direct_8x8_inference );
 533     BOOLIFY( analyse.b_bidir_me );
 534     BOOLIFY( analyse.b_chroma_me );
 535     BOOLIFY( analyse.b_fast_pskip );
 536     BOOLIFY( rc.b_stat_write );
 537     BOOLIFY( rc.b_stat_read );
 538 #undef BOOLIFY
 539
 540     return 0;
 541 }
 542
 543 static void mbcmp_init( x264_t *h )
 544 {
 545     memcpy( h->pixf.mbcmp,
 546             ( h->mb.b_lossless || h->param.analyse.i_subpel_refine <= 1 ) ? h->pixf.sad : h->pixf.satd,
 547             sizeof(h->pixf.mbcmp) );
 548 }
 549
 550 /****************************************************************************
 551  * x264_encoder_open:
 552  ****************************************************************************/
 553 x264_t *x264_encoder_open   ( x264_param_t *param )
 554 {
 555     x264_t *h = x264_malloc( sizeof( x264_t ) );
 556     int i;
 557
 558     memset( h, 0, sizeof( x264_t ) );
 559
 560     /* Create a copy of param */
 561     memcpy( &h->param, param, sizeof( x264_param_t ) );
 562
 563     if( x264_validate_parameters( h ) < 0 )
 564     {
 565         x264_free( h );
 566         return NULL;
 567     }
 568
 569     if( h->param.psz_cqm_file )
 570         if( x264_cqm_parse_file( h, h->param.psz_cqm_file ) < 0 )
 571         {
 572             x264_free( h );
 573             return NULL;
 574         }
 575
 576     if( h->param.rc.psz_stat_out )
 577         h->param.rc.psz_stat_out = strdup( h->param.rc.psz_stat_out );
 578     if( h->param.rc.psz_stat_in )
 579         h->param.rc.psz_stat_in = strdup( h->param.rc.psz_stat_in );
 580     if( h->param.rc.psz_rc_eq )
 581         h->param.rc.psz_rc_eq = strdup( h->param.rc.psz_rc_eq );
 582
 583     /* VUI */
 584     if( h->param.vui.i_sar_width > 0 && h->param.vui.i_sar_height > 0 )
 585     {
 586         int i_w = param->vui.i_sar_width;
 587         int i_h = param->vui.i_sar_height;
 588
 589         x264_reduce_fraction( &i_w, &i_h );
 590
 591         while( i_w > 65535 || i_h > 65535 )
 592         {
 593             i_w /= 2;
 594             i_h /= 2;
 595         }
 596
 597         h->param.vui.i_sar_width = 0;
 598         h->param.vui.i_sar_height = 0;
 599         if( i_w == 0 || i_h == 0 )
 600         {
 601             x264_log( h, X264_LOG_WARNING, "cannot create valid sample aspect ratio\n" );
 602         }
 603         else
 604         {
 605             x264_log( h, X264_LOG_INFO, "using SAR=%d/%d\n", i_w, i_h );
 606             h->param.vui.i_sar_width = i_w;
 607             h->param.vui.i_sar_height = i_h;
 608         }
 609     }
 610
 611     x264_reduce_fraction( &h->param.i_fps_num, &h->param.i_fps_den );
 612
 613     /* Init x264_t */
 614     h->i_frame = 0;
 615     h->i_frame_num = 0;
 616     h->i_idr_pic_id = 0;
 617
 618     h->sps = &h->sps_array[0];
 619     x264_sps_init( h->sps, h->param.i_sps_id, &h->param );
 620
 621     h->pps = &h->pps_array[0];
 622     x264_pps_init( h->pps, h->param.i_sps_id, &h->param, h->sps);
 623
 624     x264_validate_levels( h );
 625
 626     x264_cqm_init( h );
 627
 628     h->mb.i_mb_count = h->sps->i_mb_width * h->sps->i_mb_height;
 629
 630     /* Init frames. */
 631     h->frames.i_delay = h->param.i_bframe + h->param.i_threads - 1;
 632     h->frames.i_max_ref0 = h->param.i_frame_reference;
 633     h->frames.i_max_ref1 = h->sps->vui.i_num_reorder_frames;
 634     h->frames.i_max_dpb  = h->sps->vui.i_max_dec_frame_buffering;
 635     h->frames.b_have_lowres = !h->param.rc.b_stat_read
 636         && ( h->param.rc.i_rc_method == X264_RC_ABR
 637           || h->param.rc.i_rc_method == X264_RC_CRF
 638           || h->param.b_bframe_adaptive );
 639
 640     h->frames.i_last_idr = - h->param.i_keyint_max;
 641     h->frames.i_input    = 0;
 642     h->frames.last_nonb  = NULL;
 643
 644     h->i_ref0 = 0;
 645     h->i_ref1 = 0;
 646
 647     x264_rdo_init( );
 648
 649     /* init CPU functions */
 650     x264_predict_16x16_init( h->param.cpu, h->predict_16x16 );
 651     x264_predict_8x8c_init( h->param.cpu, h->predict_8x8c );
 652     x264_predict_8x8_init( h->param.cpu, h->predict_8x8 );
 653     x264_predict_4x4_init( h->param.cpu, h->predict_4x4 );
 654
 655     x264_pixel_init( h->param.cpu, &h->pixf );
 656     x264_dct_init( h->param.cpu, &h->dctf );
 657     x264_zigzag_init( h->param.cpu, &h->zigzagf, h->param.b_interlaced );
 658     x264_mc_init( h->param.cpu, &h->mc );
 659     x264_csp_init( h->param.cpu, h->param.i_csp, &h->csp );
 660     x264_quant_init( h, h->param.cpu, &h->quantf );
 661     x264_deblock_init( h->param.cpu, &h->loopf );
 662     x264_dct_init_weights();
 663
 664     mbcmp_init( h );
 665
 666     x264_log( h, X264_LOG_INFO, "using cpu capabilities %s%s%s%s%s%s\n",
 667              param->cpu&X264_CPU_MMX ? "MMX " : "",
 668              param->cpu&X264_CPU_MMXEXT ? "MMXEXT " : "",
 669              param->cpu&X264_CPU_SSE ? "SSE " : "",
 670              param->cpu&X264_CPU_SSE2 ? "SSE2 " : "",
 671              param->cpu&X264_CPU_3DNOW ? "3DNow! " : "",
 672              param->cpu&X264_CPU_ALTIVEC ? "Altivec " : "" );
 673
 674     h->out.i_nal = 0;
 675     h->out.i_bitstream = X264_MAX( 1000000, h->param.i_width * h->param.i_height * 1.7
 676         * ( h->param.rc.i_rc_method == X264_RC_ABR ? pow( 0.5, h->param.rc.i_qp_min )
 677           : pow( 0.5, h->param.rc.i_qp_constant ) * X264_MAX( 1, h->param.rc.f_ip_factor )));
 678
 679     h->thread[0] = h;
 680     h->i_thread_num = 0;
 681     for( i = 1; i < h->param.i_threads; i++ )
 682         h->thread[i] = x264_malloc( sizeof(x264_t) );
 683
 684     for( i = 0; i < h->param.i_threads; i++ )
 685     {
 686         if( i > 0 )
 687             *h->thread[i] = *h;
 688         h->thread[i]->fdec = x264_frame_pop_unused( h );
 689         h->thread[i]->out.p_bitstream = x264_malloc( h->out.i_bitstream );
 690         if( x264_macroblock_cache_init( h->thread[i] ) < 0 )
 691             return NULL;
 692     }
 693
 694     if( x264_ratecontrol_new( h ) < 0 )
 695         return NULL;
 696
 697 #ifdef DEBUG_DUMP_FRAME
 698     {
 699         /* create or truncate the reconstructed video file */
 700         FILE *f = fopen( "fdec.yuv", "w" );
 701         if( f )
 702             fclose( f );
 703         else
 704         {
 705             x264_log( h, X264_LOG_ERROR, "can't write to fdec.yuv\n" );
 706             x264_free( h );
 707             return NULL;
 708         }
 709     }
 710 #endif
 711
 712     return h;
 713 }
 714
 715 /****************************************************************************
 716  * x264_encoder_reconfig:
 717  ****************************************************************************/
 718 int x264_encoder_reconfig( x264_t *h, x264_param_t *param )
 719 {
 720 #define COPY(var) h->param.var = param->var
 721     COPY( i_frame_reference ); // but never uses more refs than initially specified
 722     COPY( i_bframe_bias );
 723     COPY( i_scenecut_threshold );
 724     COPY( b_deblocking_filter );
 725     COPY( i_deblocking_filter_alphac0 );
 726     COPY( i_deblocking_filter_beta );
 727     COPY( analyse.intra );
 728     COPY( analyse.inter );
 729     COPY( analyse.i_direct_mv_pred );
 730     COPY( analyse.i_me_method );
 731     COPY( analyse.i_me_range );
 732     COPY( analyse.i_noise_reduction );
 733     COPY( analyse.i_subpel_refine );
 734     COPY( analyse.i_trellis );
 735     COPY( analyse.b_bidir_me );
 736     COPY( analyse.b_bframe_rdo );
 737     COPY( analyse.b_chroma_me );
 738     COPY( analyse.b_dct_decimate );
 739     COPY( analyse.b_fast_pskip );
 740     COPY( analyse.b_mixed_references );
 741 #undef COPY
 742
 743     if( h->pps->b_transform_8x8_mode )
 744         h->param.analyse.b_transform_8x8 = param->analyse.b_transform_8x8;
 745
 746     mbcmp_init( h );
 747
 748     return x264_validate_parameters( h );
 749 }
 750
 751 /* internal usage */
 752 static void x264_nal_start( x264_t *h, int i_type, int i_ref_idc )
 753 {
 754     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 755
 756     nal->i_ref_idc = i_ref_idc;
 757     nal->i_type    = i_type;
 758
 759     bs_align_0( &h->out.bs );   /* not needed */
 760
 761     nal->i_payload= 0;
 762     nal->p_payload= &h->out.p_bitstream[bs_pos( &h->out.bs) / 8];
 763 }
 764 static void x264_nal_end( x264_t *h )
 765 {
 766     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 767
 768     bs_align_0( &h->out.bs );   /* not needed */
 769
 770     nal->i_payload = &h->out.p_bitstream[bs_pos( &h->out.bs)/8] - nal->p_payload;
 771
 772     h->out.i_nal++;
 773 }
 774
 775 /****************************************************************************
 776  * x264_encoder_headers:
 777  ****************************************************************************/
 778 int x264_encoder_headers( x264_t *h, x264_nal_t **pp_nal, int *pi_nal )
 779 {
 780     /* init bitstream context */
 781     h->out.i_nal = 0;
 782     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
 783
 784     /* Put SPS and PPS */
 785     if( h->i_frame == 0 )
 786     {
 787         /* identify ourself */
 788         x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
 789         x264_sei_version_write( h, &h->out.bs );
 790         x264_nal_end( h );
 791
 792         /* generate sequence parameters */
 793         x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
 794         x264_sps_write( &h->out.bs, h->sps );
 795         x264_nal_end( h );
 796
 797         /* generate picture parameters */
 798         x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
 799         x264_pps_write( &h->out.bs, h->pps );
 800         x264_nal_end( h );
 801     }
 802     /* now set output*/
 803     *pi_nal = h->out.i_nal;
 804     *pp_nal = &h->out.nal[0];
 805     h->out.i_nal = 0;
 806
 807     return 0;
 808 }
 809
 810 static inline void x264_reference_build_list( x264_t *h, int i_poc )
 811 {
 812     int i;
 813     int b_ok;
 814
 815     /* build ref list 0/1 */
 816     h->i_ref0 = 0;
 817     h->i_ref1 = 0;
 818     for( i = 0; h->frames.reference[i]; i++ )
 819     {
 820         if( h->frames.reference[i]->i_poc < i_poc )
 821         {
 822             h->fref0[h->i_ref0++] = h->frames.reference[i];
 823         }
 824         else if( h->frames.reference[i]->i_poc > i_poc )
 825         {
 826             h->fref1[h->i_ref1++] = h->frames.reference[i];
 827         }
 828     }
 829
 830     /* Order ref0 from higher to lower poc */
 831     do
 832     {
 833         b_ok = 1;
 834         for( i = 0; i < h->i_ref0 - 1; i++ )
 835         {
 836             if( h->fref0[i]->i_poc < h->fref0[i+1]->i_poc )
 837             {
 838                 XCHG( x264_frame_t*, h->fref0[i], h->fref0[i+1] );
 839                 b_ok = 0;
 840                 break;
 841             }
 842         }
 843     } while( !b_ok );
 844     /* Order ref1 from lower to higher poc (bubble sort) for B-frame */
 845     do
 846     {
 847         b_ok = 1;
 848         for( i = 0; i < h->i_ref1 - 1; i++ )
 849         {
 850             if( h->fref1[i]->i_poc > h->fref1[i+1]->i_poc )
 851             {
 852                 XCHG( x264_frame_t*, h->fref1[i], h->fref1[i+1] );
 853                 b_ok = 0;
 854                 break;
 855             }
 856         }
 857     } while( !b_ok );
 858
 859     /* In the standard, a P-frame's ref list is sorted by frame_num.
 860      * We use POC, but check whether explicit reordering is needed */
 861     h->b_ref_reorder[0] =
 862     h->b_ref_reorder[1] = 0;
 863     if( h->sh.i_type == SLICE_TYPE_P )
 864     {
 865         for( i = 0; i < h->i_ref0 - 1; i++ )
 866             if( h->fref0[i]->i_frame_num < h->fref0[i+1]->i_frame_num )
 867             {
 868                 h->b_ref_reorder[0] = 1;
 869                 break;
 870             }
 871     }
 872
 873     h->i_ref1 = X264_MIN( h->i_ref1, h->frames.i_max_ref1 );
 874     h->i_ref0 = X264_MIN( h->i_ref0, h->frames.i_max_ref0 );
 875     h->i_ref0 = X264_MIN( h->i_ref0, h->param.i_frame_reference ); // if reconfig() has lowered the limit
 876     h->i_ref0 = X264_MIN( h->i_ref0, 16 - h->i_ref1 );
 877     h->mb.pic.i_fref[0] = h->i_ref0;
 878     h->mb.pic.i_fref[1] = h->i_ref1;
 879 }
 880
 881 static void x264_fdec_filter_row( x264_t *h, int mb_y )
 882 {
 883     /* mb_y is the mb to be encoded next, not the mb to be filtered here */
 884     int b_hpel = h->fdec->b_kept_as_ref;
 885     int b_deblock = !h->sh.i_disable_deblocking_filter_idc;
 886     int b_end = mb_y == h->sps->i_mb_height;
 887     int min_y = mb_y - (1 << h->sh.b_mbaff);
 888 #ifndef DEBUG_DUMP_FRAME
 889     b_deblock &= b_hpel;
 890 #endif
 891     if( mb_y & h->sh.b_mbaff )
 892         return;
 893     if( min_y < 0 )
 894         return;
 895
 896     if( !b_end )
 897     {
 898         int i, j;
 899         for( j=0; j<=h->sh.b_mbaff; j++ )
 900             for( i=0; i<3; i++ )
 901             {
 902                 memcpy( h->mb.intra_border_backup[j][i],
 903                         h->fdec->plane[i] + ((mb_y*16 >> !!i) + j - 1 - h->sh.b_mbaff) * h->fdec->i_stride[i],
 904                         h->sps->i_mb_width*16 >> !!i );
 905             }
 906     }
 907
 908     if( b_deblock )
 909     {
 910         int max_y = b_end ? h->sps->i_mb_height : mb_y;
 911         int y;
 912         for( y = min_y; y < max_y; y += (1 << h->sh.b_mbaff) )
 913             x264_frame_deblock_row( h, y );
 914     }
 915
 916     if( b_hpel )
 917     {
 918         x264_frame_expand_border( h, h->fdec, min_y, b_end );
 919         x264_frame_filter( h->param.cpu, h->fdec, h->sh.b_mbaff, min_y, b_end );
 920         x264_frame_expand_border_filtered( h, h->fdec, min_y, b_end );
 921     }
 922
 923     if( h->param.i_threads > 1 && h->fdec->b_kept_as_ref )
 924     {
 925         x264_frame_cond_broadcast( h->fdec, mb_y*16 + (b_end ? 10000 : -(X264_THREAD_HEIGHT << h->sh.b_mbaff)) );
 926     }
 927 }
 928
 929 static inline void x264_reference_update( x264_t *h )
 930 {
 931     int i;
 932
 933     if( h->fdec->i_frame >= 0 )
 934         h->i_frame++;
 935
 936     if( !h->fdec->b_kept_as_ref )
 937     {
 938         if( h->param.i_threads > 1 )
 939         {
 940             x264_frame_push_unused( h, h->fdec );
 941             h->fdec = x264_frame_pop_unused( h );
 942         }
 943         return;
 944     }
 945
 946     /* move lowres copy of the image to the ref frame */
 947     for( i = 0; i < 4; i++)
 948     {
 949         XCHG( uint8_t*, h->fdec->lowres[i], h->fenc->lowres[i] );
 950         XCHG( uint8_t*, h->fdec->buffer_lowres[i], h->fenc->buffer_lowres[i] );
 951     }
 952
 953     /* adaptive B decision needs a pointer, since it can't use the ref lists */
 954     if( h->sh.i_type != SLICE_TYPE_B )
 955         h->frames.last_nonb = h->fdec;
 956
 957     /* move frame in the buffer */
 958     x264_frame_push( h->frames.reference, h->fdec );
 959     if( h->frames.reference[h->frames.i_max_dpb] )
 960         x264_frame_push_unused( h, x264_frame_shift( h->frames.reference ) );
 961     h->fdec = x264_frame_pop_unused( h );
 962 }
 963
 964 static inline void x264_reference_reset( x264_t *h )
 965 {
 966     while( h->frames.reference[0] )
 967         x264_frame_push_unused( h, x264_frame_pop( h->frames.reference ) );
 968     h->fdec->i_poc =
 969     h->fenc->i_poc = 0;
 970 }
 971
 972 static inline void x264_slice_init( x264_t *h, int i_nal_type, int i_global_qp )
 973 {
 974     /* ------------------------ Create slice header  ----------------------- */
 975     if( i_nal_type == NAL_SLICE_IDR )
 976     {
 977         x264_slice_header_init( h, &h->sh, h->sps, h->pps, h->i_idr_pic_id, h->i_frame_num, i_global_qp );
 978
 979         /* increment id */
 980         h->i_idr_pic_id = ( h->i_idr_pic_id + 1 ) % 65536;
 981     }
 982     else
 983     {
 984         x264_slice_header_init( h, &h->sh, h->sps, h->pps, -1, h->i_frame_num, i_global_qp );
 985
 986         /* always set the real higher num of ref frame used */
 987         h->sh.b_num_ref_idx_override = 1;
 988         h->sh.i_num_ref_idx_l0_active = h->i_ref0 <= 0 ? 1 : h->i_ref0;
 989         h->sh.i_num_ref_idx_l1_active = h->i_ref1 <= 0 ? 1 : h->i_ref1;
 990     }
 991
 992     h->fdec->i_frame_num = h->sh.i_frame_num;
 993
 994     if( h->sps->i_poc_type == 0 )
 995     {
 996         h->sh.i_poc_lsb = h->fdec->i_poc & ( (1 << h->sps->i_log2_max_poc_lsb) - 1 );
 997         h->sh.i_delta_poc_bottom = 0;   /* XXX won't work for field */
 998     }
 999     else if( h->sps->i_poc_type == 1 )
1000     {
1001         /* FIXME TODO FIXME */
1002     }
1003     else
1004     {
1005         /* Nothing to do ? */
1006     }
1007
1008     x264_macroblock_slice_init( h );
1009 }
1010
1011 static int x264_slice_write( x264_t *h )
1012 {
1013     int i_skip;
1014     int mb_xy;
1015     int i;
1016
1017     /* init stats */
1018     memset( &h->stat.frame, 0, sizeof(h->stat.frame) );
1019
1020     /* Slice */
1021     x264_nal_start( h, h->i_nal_type, h->i_nal_ref_idc );
1022
1023     /* Slice header */
1024     x264_slice_header_write( &h->out.bs, &h->sh, h->i_nal_ref_idc );
1025     if( h->param.b_cabac )
1026     {
1027         /* alignment needed */
1028         bs_align_1( &h->out.bs );
1029
1030         /* init cabac */
1031         x264_cabac_context_init( &h->cabac, h->sh.i_type, h->sh.i_qp, h->sh.i_cabac_init_idc );
1032         x264_cabac_encode_init ( &h->cabac, &h->out.bs );
1033     }
1034     h->mb.i_last_qp = h->sh.i_qp;
1035     h->mb.i_last_dqp = 0;
1036
1037     for( mb_xy = h->sh.i_first_mb, i_skip = 0; mb_xy < h->sh.i_last_mb; )
1038     {
1039         const int i_mb_y = mb_xy / h->sps->i_mb_width;
1040         const int i_mb_x = mb_xy % h->sps->i_mb_width;
1041         int mb_spos = bs_pos(&h->out.bs);
1042
1043         if( i_mb_x == 0 )
1044             x264_fdec_filter_row( h, i_mb_y );
1045
1046         /* load cache */
1047         x264_macroblock_cache_load( h, i_mb_x, i_mb_y );
1048
1049         /* analyse parameters
1050          * Slice I: choose I_4x4 or I_16x16 mode
1051          * Slice P: choose between using P mode or intra (4x4 or 16x16)
1052          * */
1053         TIMER_START( i_mtime_analyse );
1054         x264_macroblock_analyse( h );
1055         TIMER_STOP( i_mtime_analyse );
1056
1057         /* encode this macroblock -> be careful it can change the mb type to P_SKIP if needed */
1058         TIMER_START( i_mtime_encode );
1059         x264_macroblock_encode( h );
1060         TIMER_STOP( i_mtime_encode );
1061
1062         TIMER_START( i_mtime_write );
1063         if( h->param.b_cabac )
1064         {
1065             if( mb_xy > h->sh.i_first_mb && !(h->sh.b_mbaff && (i_mb_y&1)) )
1066                 x264_cabac_encode_terminal( &h->cabac, 0 );
1067
1068             if( IS_SKIP( h->mb.i_type ) )
1069                 x264_cabac_mb_skip( h, 1 );
1070             else
1071             {
1072                 if( h->sh.i_type != SLICE_TYPE_I )
1073                     x264_cabac_mb_skip( h, 0 );
1074                 x264_macroblock_write_cabac( h, &h->cabac );
1075             }
1076         }
1077         else
1078         {
1079             if( IS_SKIP( h->mb.i_type ) )
1080                 i_skip++;
1081             else
1082             {
1083                 if( h->sh.i_type != SLICE_TYPE_I )
1084                 {
1085                     bs_write_ue( &h->out.bs, i_skip );  /* skip run */
1086                     i_skip = 0;
1087                 }
1088                 x264_macroblock_write_cavlc( h, &h->out.bs );
1089             }
1090         }
1091         TIMER_STOP( i_mtime_write );
1092
1093 #if VISUALIZE
1094         if( h->param.b_visualize )
1095             x264_visualize_mb( h );
1096 #endif
1097
1098         /* save cache */
1099         x264_macroblock_cache_save( h );
1100
1101         /* accumulate mb stats */
1102         h->stat.frame.i_mb_count[h->mb.i_type]++;
1103         if( !IS_SKIP(h->mb.i_type) && !IS_INTRA(h->mb.i_type) && !IS_DIRECT(h->mb.i_type) )
1104         {
1105             if( h->mb.i_partition != D_8x8 )
1106                 h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_partition ] ] += 4;
1107             else
1108                 for( i = 0; i < 4; i++ )
1109                     h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_sub_partition[i] ] ] ++;
1110             if( h->param.i_frame_reference > 1 )
1111             {
1112                 for( i = 0; i < 4; i++ )
1113                 {
1114                     int i_ref = h->mb.cache.ref[0][ x264_scan8[4*i] ];
1115                     if( i_ref >= 0 )
1116                         h->stat.frame.i_mb_count_ref[i_ref] ++;
1117                 }
1118             }
1119         }
1120         if( h->mb.i_cbp_luma && !IS_INTRA(h->mb.i_type) )
1121         {
1122             h->stat.frame.i_mb_count_8x8dct[0] ++;
1123             h->stat.frame.i_mb_count_8x8dct[1] += h->mb.b_transform_8x8;
1124         }
1125
1126         if( h->mb.b_variable_qp )
1127             x264_ratecontrol_mb(h, bs_pos(&h->out.bs) - mb_spos);
1128
1129         if( h->sh.b_mbaff )
1130         {
1131             if( (i_mb_y&1) && i_mb_x == h->sps->i_mb_width - 1 )
1132                 mb_xy++;
1133             else if( i_mb_y&1 )
1134                 mb_xy += 1 - h->sps->i_mb_width;
1135             else
1136                 mb_xy += h->sps->i_mb_width;
1137         }
1138         else
1139             mb_xy++;
1140     }
1141
1142     if( h->param.b_cabac )
1143     {
1144         /* end of slice */
1145         x264_cabac_encode_terminal( &h->cabac, 1 );
1146     }
1147     else if( i_skip > 0 )
1148     {
1149         bs_write_ue( &h->out.bs, i_skip );  /* last skip run */
1150     }
1151
1152     if( h->param.b_cabac )
1153     {
1154         x264_cabac_encode_flush( &h->cabac );
1155
1156     }
1157     else
1158     {
1159         /* rbsp_slice_trailing_bits */
1160         bs_rbsp_trailing( &h->out.bs );
1161     }
1162
1163     x264_nal_end( h );
1164
1165     /* Compute misc bits */
1166     h->stat.frame.i_misc_bits = bs_pos( &h->out.bs )
1167                               + NALU_OVERHEAD * 8
1168                               - h->stat.frame.i_itex_bits
1169                               - h->stat.frame.i_ptex_bits
1170                               - h->stat.frame.i_hdr_bits;
1171
1172     return 0;
1173 }
1174
1175 static void x264_thread_sync_context( x264_t *dst, x264_t *src )
1176 {
1177     x264_frame_t **f;
1178     if( dst == src )
1179         return;
1180
1181     // reference counting
1182     for( f = src->frames.reference; *f; f++ )
1183         (*f)->i_reference_count++;
1184     for( f = dst->frames.reference; *f; f++ )
1185         x264_frame_push_unused( src, *f );
1186     src->fdec->i_reference_count++;
1187     x264_frame_push_unused( src, dst->fdec );
1188
1189     // copy everything except the per-thread pointers and the constants.
1190     memcpy( &dst->i_frame, &src->i_frame, offsetof(x264_t, mb.type) - offsetof(x264_t, i_frame) );
1191     memcpy( &dst->mb.i_type, &src->mb.i_type, offsetof(x264_t, rc) - offsetof(x264_t, mb.i_type) );
1192     dst->stat = src->stat;
1193 }
1194
1195 static void x264_thread_sync_stat( x264_t *dst, x264_t *src )
1196 {
1197     if( dst == src )
1198         return;
1199     memcpy( &dst->stat.i_slice_count, &src->stat.i_slice_count, sizeof(dst->stat) - sizeof(dst->stat.frame) );
1200 }
1201
1202 static int x264_slices_write( x264_t *h )
1203 {
1204     int i_frame_size;
1205
1206 #if VISUALIZE
1207     if( h->param.b_visualize )
1208         x264_visualize_init( h );
1209 #endif
1210
1211     x264_slice_write( h );
1212     i_frame_size = h->out.nal[h->out.i_nal-1].i_payload;
1213     x264_fdec_filter_row( h, h->sps->i_mb_height );
1214
1215 #if VISUALIZE
1216     if( h->param.b_visualize )
1217     {
1218         x264_visualize_show( h );
1219         x264_visualize_close( h );
1220     }
1221 #endif
1222
1223     h->out.i_frame_size = i_frame_size;
1224     return 0;
1225 }
1226
1227 /****************************************************************************
1228  * x264_encoder_encode:
1229  *  XXX: i_poc   : is the poc of the current given picture
1230  *       i_frame : is the number of the frame being coded
1231  *  ex:  type frame poc
1232  *       I      0   2*0
1233  *       P      1   2*3
1234  *       B      2   2*1
1235  *       B      3   2*2
1236  *       P      4   2*6
1237  *       B      5   2*4
1238  *       B      6   2*5
1239  ****************************************************************************/
1240 int     x264_encoder_encode( x264_t *h,
1241                              x264_nal_t **pp_nal, int *pi_nal,
1242                              x264_picture_t *pic_in,
1243                              x264_picture_t *pic_out )
1244 {
1245     x264_t *thread_current, *thread_prev, *thread_oldest;
1246     int     i_nal_type;
1247     int     i_nal_ref_idc;
1248
1249     int   i_global_qp;
1250
1251     if( h->param.i_threads > 1)
1252     {
1253         int i = ++h->i_thread_phase;
1254         int t = h->param.i_threads;
1255         thread_current = h->thread[ i%t ];
1256         thread_prev    = h->thread[ (i-1)%t ];
1257         thread_oldest  = h->thread[ (i+1)%t ];
1258         x264_thread_sync_context( thread_current, thread_prev );
1259         x264_thread_sync_ratecontrol( thread_current, thread_prev, thread_oldest );
1260         h = thread_current;
1261 //      fprintf(stderr, "current: %p  prev: %p  oldest: %p \n", thread_current, thread_prev, thread_oldest);
1262     }
1263     else
1264     {
1265         thread_current =
1266         thread_prev    =
1267         thread_oldest  = h;
1268     }
1269
1270     // ok to call this before encoding any frames, since the initial values of fdec have b_kept_as_ref=0
1271     x264_reference_update( h );
1272     h->fdec->i_lines_completed = -1;
1273
1274     /* no data out */
1275     *pi_nal = 0;
1276     *pp_nal = NULL;
1277
1278     /* ------------------- Setup new frame from picture -------------------- */
1279     TIMER_START( i_mtime_encode_frame );
1280     if( pic_in != NULL )
1281     {
1282         /* 1: Copy the picture to a frame and move it to a buffer */
1283         x264_frame_t *fenc = x264_frame_pop_unused( h );
1284
1285         x264_frame_copy_picture( h, fenc, pic_in );
1286
1287         if( h->param.i_width != 16 * h->sps->i_mb_width ||
1288             h->param.i_height != 16 * h->sps->i_mb_height )
1289             x264_frame_expand_border_mod16( h, fenc );
1290
1291         fenc->i_frame = h->frames.i_input++;
1292
1293         x264_frame_push( h->frames.next, fenc );
1294
1295         if( h->frames.b_have_lowres )
1296             x264_frame_init_lowres( h->param.cpu, fenc );
1297
1298         if( h->frames.i_input <= h->frames.i_delay + 1 - h->param.i_threads )
1299         {
1300             /* Nothing yet to encode */
1301             /* waiting for filling bframe buffer */
1302             pic_out->i_type = X264_TYPE_AUTO;
1303             return 0;
1304         }
1305     }
1306
1307     if( h->frames.current[0] == NULL )
1308     {
1309         int bframes = 0;
1310         /* 2: Select frame types */
1311         if( h->frames.next[0] == NULL )
1312         {
1313             x264_encoder_frame_end( thread_oldest, thread_current, pp_nal, pi_nal, pic_out );
1314             return 0;
1315         }
1316
1317         x264_slicetype_decide( h );
1318
1319         /* 3: move some B-frames and 1 non-B to encode queue */
1320         while( IS_X264_TYPE_B( h->frames.next[bframes]->i_type ) )
1321             bframes++;
1322         x264_frame_push( h->frames.current, x264_frame_shift( &h->frames.next[bframes] ) );
1323         /* FIXME: when max B-frames > 3, BREF may no longer be centered after GOP closing */
1324         if( h->param.b_bframe_pyramid && bframes > 1 )
1325         {
1326             x264_frame_t *mid = x264_frame_shift( &h->frames.next[bframes/2] );
1327             mid->i_type = X264_TYPE_BREF;
1328             x264_frame_push( h->frames.current, mid );
1329             bframes--;
1330         }
1331         while( bframes-- )
1332             x264_frame_push( h->frames.current, x264_frame_shift( h->frames.next ) );
1333     }
1334     TIMER_STOP( i_mtime_encode_frame );
1335
1336     /* ------------------- Get frame to be encoded ------------------------- */
1337     /* 4: get picture to encode */
1338     h->fenc = x264_frame_shift( h->frames.current );
1339     if( h->fenc == NULL )
1340     {
1341         /* Nothing yet to encode (ex: waiting for I/P with B frames) */
1342         /* waiting for filling bframe buffer */
1343         pic_out->i_type = X264_TYPE_AUTO;
1344         return 0;
1345     }
1346
1347 do_encode:
1348
1349     if( h->fenc->i_type == X264_TYPE_IDR )
1350     {
1351         h->frames.i_last_idr = h->fenc->i_frame;
1352     }
1353
1354     /* ------------------- Setup frame context ----------------------------- */
1355     /* 5: Init data dependent of frame type */
1356     TIMER_START( i_mtime_encode_frame );
1357     if( h->fenc->i_type == X264_TYPE_IDR )
1358     {
1359         /* reset ref pictures */
1360         x264_reference_reset( h );
1361
1362         i_nal_type    = NAL_SLICE_IDR;
1363         i_nal_ref_idc = NAL_PRIORITY_HIGHEST;
1364         h->sh.i_type = SLICE_TYPE_I;
1365     }
1366     else if( h->fenc->i_type == X264_TYPE_I )
1367     {
1368         i_nal_type    = NAL_SLICE;
1369         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
1370         h->sh.i_type = SLICE_TYPE_I;
1371     }
1372     else if( h->fenc->i_type == X264_TYPE_P )
1373     {
1374         i_nal_type    = NAL_SLICE;
1375         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
1376         h->sh.i_type = SLICE_TYPE_P;
1377     }
1378     else if( h->fenc->i_type == X264_TYPE_BREF )
1379     {
1380         i_nal_type    = NAL_SLICE;
1381         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* maybe add MMCO to forget it? -> low */
1382         h->sh.i_type = SLICE_TYPE_B;
1383     }
1384     else    /* B frame */
1385     {
1386         i_nal_type    = NAL_SLICE;
1387         i_nal_ref_idc = NAL_PRIORITY_DISPOSABLE;
1388         h->sh.i_type = SLICE_TYPE_B;
1389     }
1390
1391     h->fdec->i_poc =
1392     h->fenc->i_poc = 2 * (h->fenc->i_frame - h->frames.i_last_idr);
1393     h->fdec->i_type = h->fenc->i_type;
1394     h->fdec->i_frame = h->fenc->i_frame;
1395     h->fenc->b_kept_as_ref =
1396     h->fdec->b_kept_as_ref = i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE && h->param.i_keyint_max > 1;
1397
1398
1399
1400     /* ------------------- Init                ----------------------------- */
1401     /* build ref list 0/1 */
1402     x264_reference_build_list( h, h->fdec->i_poc );
1403
1404     /* Init the rate control */
1405     x264_ratecontrol_start( h, h->fenc->i_qpplus1 );
1406     i_global_qp = x264_ratecontrol_qp( h );
1407
1408     pic_out->i_qpplus1 =
1409     h->fdec->i_qpplus1 = i_global_qp + 1;
1410
1411     if( h->sh.i_type == SLICE_TYPE_B )
1412         x264_macroblock_bipred_init( h );
1413
1414     /* ------------------------ Create slice header  ----------------------- */
1415     x264_slice_init( h, i_nal_type, i_global_qp );
1416
1417     if( i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE )
1418         h->i_frame_num++;
1419
1420     /* ---------------------- Write the bitstream -------------------------- */
1421     /* Init bitstream context */
1422     h->out.i_nal = 0;
1423     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
1424
1425     if(h->param.b_aud){
1426         int pic_type;
1427
1428         if(h->sh.i_type == SLICE_TYPE_I)
1429             pic_type = 0;
1430         else if(h->sh.i_type == SLICE_TYPE_P)
1431             pic_type = 1;
1432         else if(h->sh.i_type == SLICE_TYPE_B)
1433             pic_type = 2;
1434         else
1435             pic_type = 7;
1436
1437         x264_nal_start(h, NAL_AUD, NAL_PRIORITY_DISPOSABLE);
1438         bs_write(&h->out.bs, 3, pic_type);
1439         bs_rbsp_trailing(&h->out.bs);
1440         x264_nal_end(h);
1441     }
1442
1443     h->i_nal_type = i_nal_type;
1444     h->i_nal_ref_idc = i_nal_ref_idc;
1445
1446     /* Write SPS and PPS */
1447     if( i_nal_type == NAL_SLICE_IDR && h->param.b_repeat_headers )
1448     {
1449         if( h->fenc->i_frame == 0 )
1450         {
1451             /* identify ourself */
1452             x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
1453             x264_sei_version_write( h, &h->out.bs );
1454             x264_nal_end( h );
1455         }
1456
1457         /* generate sequence parameters */
1458         x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
1459         x264_sps_write( &h->out.bs, h->sps );
1460         x264_nal_end( h );
1461
1462         /* generate picture parameters */
1463         x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
1464         x264_pps_write( &h->out.bs, h->pps );
1465         x264_nal_end( h );
1466     }
1467
1468     /* Write frame */
1469     if( h->param.i_threads > 1 )
1470     {
1471         pthread_create( &h->thread_handle, NULL, (void*)x264_slices_write, h );
1472         h->b_thread_active = 1;
1473     }
1474     else
1475         x264_slices_write( h );
1476
1477     /* restore CPU state (before using float again) */
1478     x264_cpu_restore( h->param.cpu );
1479
1480     if( h->sh.i_type == SLICE_TYPE_P && !h->param.rc.b_stat_read
1481         && h->param.i_scenecut_threshold >= 0
1482         && !h->param.b_pre_scenecut )
1483     {
1484         const int *mbs = h->stat.frame.i_mb_count;
1485         int i_mb_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1486         int i_mb_p = mbs[P_L0] + mbs[P_8x8];
1487         int i_mb_s = mbs[P_SKIP];
1488         int i_mb   = h->sps->i_mb_width * h->sps->i_mb_height;
1489         int64_t i_inter_cost = h->stat.frame.i_inter_cost;
1490         int64_t i_intra_cost = h->stat.frame.i_intra_cost;
1491
1492         float f_bias;
1493         int i_gop_size = h->fenc->i_frame - h->frames.i_last_idr;
1494         float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
1495         /* magic numbers pulled out of thin air */
1496         float f_thresh_min = f_thresh_max * h->param.i_keyint_min
1497                              / ( h->param.i_keyint_max * 4 );
1498         if( h->param.i_keyint_min == h->param.i_keyint_max )
1499              f_thresh_min= f_thresh_max;
1500
1501         /* macroblock_analyse() doesn't further analyse skipped mbs,
1502          * so we have to guess their cost */
1503         if( h->stat.frame.i_mbs_analysed > 0 )
1504             i_intra_cost = i_intra_cost * i_mb / h->stat.frame.i_mbs_analysed;
1505
1506         if( i_gop_size < h->param.i_keyint_min / 4 )
1507             f_bias = f_thresh_min / 4;
1508         else if( i_gop_size <= h->param.i_keyint_min )
1509             f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
1510         else
1511         {
1512             f_bias = f_thresh_min
1513                      + ( f_thresh_max - f_thresh_min )
1514                        * ( i_gop_size - h->param.i_keyint_min )
1515                        / ( h->param.i_keyint_max - h->param.i_keyint_min );
1516         }
1517         f_bias = X264_MIN( f_bias, 1.0 );
1518
1519         /* Bad P will be reencoded as I */
1520         if( h->stat.frame.i_mbs_analysed > 0 &&
1521             i_inter_cost >= (1.0 - f_bias) * i_intra_cost )
1522         {
1523             int b;
1524
1525             x264_log( h, X264_LOG_DEBUG, "scene cut at %d Icost:%.0f Pcost:%.0f ratio:%.4f bias:%.4f gop:%d (imb:%d pmb:%d smb:%d)\n",
1526                       h->fenc->i_frame,
1527                       (double)i_intra_cost, (double)i_inter_cost,
1528                       1. - (double)i_inter_cost / i_intra_cost,
1529                       f_bias, i_gop_size,
1530                       i_mb_i, i_mb_p, i_mb_s );
1531
1532             /* Restore frame num */
1533             h->i_frame_num--;
1534
1535             for( b = 0; h->frames.current[b] && IS_X264_TYPE_B( h->frames.current[b]->i_type ); b++ );
1536             if( b > 0 )
1537             {
1538                 /* If using B-frames, force GOP to be closed.
1539                  * Even if this frame is going to be I and not IDR, forcing a
1540                  * P-frame before the scenecut will probably help compression.
1541                  *
1542                  * We don't yet know exactly which frame is the scene cut, so
1543                  * we can't assign an I-frame. Instead, change the previous
1544                  * B-frame to P, and rearrange coding order. */
1545
1546                 if( h->param.b_bframe_adaptive || b > 1 )
1547                     h->fenc->i_type = X264_TYPE_AUTO;
1548                 x264_frame_sort_pts( h->frames.current );
1549                 x264_frame_unshift( h->frames.next, h->fenc );
1550                 h->fenc = h->frames.current[b-1];
1551                 h->frames.current[b-1] = NULL;
1552                 h->fenc->i_type = X264_TYPE_P;
1553                 x264_frame_sort_dts( h->frames.current );
1554             }
1555             /* Do IDR if needed */
1556             else if( i_gop_size >= h->param.i_keyint_min )
1557             {
1558                 /* Reset */
1559                 h->i_frame_num = 0;
1560
1561                 /* Reinit field of fenc */
1562                 h->fenc->i_type = X264_TYPE_IDR;
1563                 h->fenc->i_poc = 0;
1564
1565                 /* Put enqueued frames back in the pool */
1566                 while( h->frames.current[0] )
1567                     x264_frame_push( h->frames.next, x264_frame_shift( h->frames.current ) );
1568                 x264_frame_sort_pts( h->frames.next );
1569             }
1570             else
1571             {
1572                 h->fenc->i_type = X264_TYPE_I;
1573             }
1574             goto do_encode;
1575         }
1576     }
1577
1578     x264_encoder_frame_end( thread_oldest, thread_current, pp_nal, pi_nal, pic_out );
1579     return 0;
1580 }
1581
1582 static void x264_encoder_frame_end( x264_t *h, x264_t *thread_current,
1583                                     x264_nal_t **pp_nal, int *pi_nal,
1584                                     x264_picture_t *pic_out )
1585 {
1586     int i;
1587     char psz_message[80];
1588
1589     if( h->b_thread_active )
1590     {
1591         pthread_join( h->thread_handle, NULL );
1592         h->b_thread_active = 0;
1593     }
1594     if( !h->out.i_nal )
1595     {
1596         pic_out->i_type = X264_TYPE_AUTO;
1597         return;
1598     }
1599
1600     x264_frame_push_unused( thread_current, h->fenc );
1601
1602     /* End bitstream, set output  */
1603     *pi_nal = h->out.i_nal;
1604     *pp_nal = h->out.nal;
1605     h->out.i_nal = 0;
1606
1607     /* Set output picture properties */
1608     if( h->sh.i_type == SLICE_TYPE_I )
1609         pic_out->i_type = h->i_nal_type == NAL_SLICE_IDR ? X264_TYPE_IDR : X264_TYPE_I;
1610     else if( h->sh.i_type == SLICE_TYPE_P )
1611         pic_out->i_type = X264_TYPE_P;
1612     else
1613         pic_out->i_type = X264_TYPE_B;
1614     pic_out->i_pts = h->fenc->i_pts;
1615
1616     pic_out->img.i_plane = h->fdec->i_plane;
1617     for(i = 0; i < 4; i++){
1618         pic_out->img.i_stride[i] = h->fdec->i_stride[i];
1619         pic_out->img.plane[i] = h->fdec->plane[i];
1620     }
1621
1622     /* ---------------------- Update encoder state ------------------------- */
1623
1624     /* update rc */
1625     x264_cpu_restore( h->param.cpu );
1626     x264_ratecontrol_end( h, h->out.i_frame_size * 8 );
1627
1628     /* restore CPU state (before using float again) */
1629     x264_cpu_restore( h->param.cpu );
1630
1631     x264_noise_reduction_update( h );
1632
1633     TIMER_STOP( i_mtime_encode_frame );
1634
1635     /* ---------------------- Compute/Print statistics --------------------- */
1636     x264_thread_sync_stat( h, h->thread[0] );
1637
1638     /* Slice stat */
1639     h->stat.i_slice_count[h->sh.i_type]++;
1640     h->stat.i_slice_size[h->sh.i_type] += h->out.i_frame_size + NALU_OVERHEAD;
1641     h->stat.i_slice_qp[h->sh.i_type] += h->fdec->i_qpplus1 - 1;
1642
1643     for( i = 0; i < X264_MBTYPE_MAX; i++ )
1644         h->stat.i_mb_count[h->sh.i_type][i] += h->stat.frame.i_mb_count[i];
1645     for( i = 0; i < 2; i++ )
1646         h->stat.i_mb_count_8x8dct[i] += h->stat.frame.i_mb_count_8x8dct[i];
1647     if( h->sh.i_type != SLICE_TYPE_I )
1648     {
1649         for( i = 0; i < 7; i++ )
1650             h->stat.i_mb_count_size[h->sh.i_type][i] += h->stat.frame.i_mb_count_size[i];
1651         for( i = 0; i < 32; i++ )
1652             h->stat.i_mb_count_ref[h->sh.i_type][i] += h->stat.frame.i_mb_count_ref[i];
1653     }
1654     if( h->sh.i_type == SLICE_TYPE_B )
1655     {
1656         h->stat.i_direct_frames[ h->sh.b_direct_spatial_mv_pred ] ++;
1657         if( h->mb.b_direct_auto_write )
1658         {
1659             //FIXME somewhat arbitrary time constants
1660             if( h->stat.i_direct_score[0] + h->stat.i_direct_score[1] > h->mb.i_mb_count )
1661             {
1662                 for( i = 0; i < 2; i++ )
1663                     h->stat.i_direct_score[i] = h->stat.i_direct_score[i] * 9/10;
1664             }
1665             for( i = 0; i < 2; i++ )
1666                 h->stat.i_direct_score[i] += h->stat.frame.i_direct_score[i];
1667         }
1668     }
1669
1670     psz_message[0] = '\0';
1671     if( h->param.analyse.b_psnr )
1672     {
1673         int64_t sqe[3];
1674
1675         for( i=0; i<3; i++ )
1676         {
1677             sqe[i] = x264_pixel_ssd_wxh( &h->pixf,
1678                          h->fdec->plane[i], h->fdec->i_stride[i],
1679                          h->fenc->plane[i], h->fenc->i_stride[i],
1680                          h->param.i_width >> !!i, h->param.i_height >> !!i );
1681         }
1682         x264_cpu_restore( h->param.cpu );
1683
1684         h->stat.i_sqe_global[h->sh.i_type] += sqe[0] + sqe[1] + sqe[2];
1685         h->stat.f_psnr_average[h->sh.i_type] += x264_psnr( sqe[0] + sqe[1] + sqe[2], 3 * h->param.i_width * h->param.i_height / 2 );
1686         h->stat.f_psnr_mean_y[h->sh.i_type] += x264_psnr( sqe[0], h->param.i_width * h->param.i_height );
1687         h->stat.f_psnr_mean_u[h->sh.i_type] += x264_psnr( sqe[1], h->param.i_width * h->param.i_height / 4 );
1688         h->stat.f_psnr_mean_v[h->sh.i_type] += x264_psnr( sqe[2], h->param.i_width * h->param.i_height / 4 );
1689
1690         snprintf( psz_message, 80, " PSNR Y:%5.2f U:%5.2f V:%5.2f",
1691                   x264_psnr( sqe[0], h->param.i_width * h->param.i_height ),
1692                   x264_psnr( sqe[1], h->param.i_width * h->param.i_height / 4),
1693                   x264_psnr( sqe[2], h->param.i_width * h->param.i_height / 4) );
1694     }
1695
1696     if( h->param.analyse.b_ssim )
1697     {
1698         // offset by 2 pixels to avoid alignment of ssim blocks with dct blocks
1699         float ssim_y = x264_pixel_ssim_wxh( &h->pixf,
1700                          h->fdec->plane[0] + 2+2*h->fdec->i_stride[0], h->fdec->i_stride[0],
1701                          h->fenc->plane[0] + 2+2*h->fenc->i_stride[0], h->fenc->i_stride[0],
1702                          h->param.i_width-2, h->param.i_height-2 );
1703         h->stat.f_ssim_mean_y[h->sh.i_type] += ssim_y;
1704         snprintf( psz_message + strlen(psz_message), 80 - strlen(psz_message),
1705                   " SSIM Y:%.5f", ssim_y );
1706     }
1707     psz_message[79] = '\0';
1708
1709     x264_log( h, X264_LOG_DEBUG,
1710                   "frame=%4d QP=%i NAL=%d Slice:%c Poc:%-3d I:%-4d P:%-4d SKIP:%-4d size=%d bytes%s\n",
1711               h->i_frame,
1712               h->fdec->i_qpplus1 - 1,
1713               h->i_nal_ref_idc,
1714               h->sh.i_type == SLICE_TYPE_I ? 'I' : (h->sh.i_type == SLICE_TYPE_P ? 'P' : 'B' ),
1715               h->fdec->i_poc,
1716               h->stat.frame.i_mb_count_i,
1717               h->stat.frame.i_mb_count_p,
1718               h->stat.frame.i_mb_count_skip,
1719               h->out.i_frame_size,
1720               psz_message );
1721
1722     // keep stats all in one place
1723     x264_thread_sync_stat( h->thread[0], h );
1724     // for the use of the next frame
1725     x264_thread_sync_stat( thread_current, h );
1726
1727 #ifdef DEBUG_MB_TYPE
1728 {
1729     static const char mb_chars[] = { 'i', 'i', 'I', 'C', 'P', '8', 'S',
1730         'D', '<', 'X', 'B', 'X', '>', 'B', 'B', 'B', 'B', '8', 'S' };
1731     int mb_xy;
1732     for( mb_xy = 0; mb_xy < h->sps->i_mb_width * h->sps->i_mb_height; mb_xy++ )
1733     {
1734         if( h->mb.type[mb_xy] < X264_MBTYPE_MAX && h->mb.type[mb_xy] >= 0 )
1735             fprintf( stderr, "%c ", mb_chars[ h->mb.type[mb_xy] ] );
1736         else
1737             fprintf( stderr, "? " );
1738
1739         if( (mb_xy+1) % h->sps->i_mb_width == 0 )
1740             fprintf( stderr, "\n" );
1741     }
1742 }
1743 #endif
1744
1745 #ifdef DEBUG_DUMP_FRAME
1746     /* Dump reconstructed frame */
1747     x264_frame_dump( h, h->fdec, "fdec.yuv" );
1748 #endif
1749 }
1750
1751 /****************************************************************************
1752  * x264_encoder_close:
1753  ****************************************************************************/
1754 void    x264_encoder_close  ( x264_t *h )
1755 {
1756 #ifdef DEBUG_BENCHMARK
1757     int64_t i_mtime_total = i_mtime_analyse + i_mtime_encode + i_mtime_write + i_mtime_filter + 1;
1758 #endif
1759     int64_t i_yuv_size = 3 * h->param.i_width * h->param.i_height / 2;
1760     int i;
1761
1762     for( i=0; i<h->param.i_threads; i++ )
1763     {
1764         // don't strictly have to wait for the other threads, but it's simpler than cancelling them
1765         if( h->thread[i]->b_thread_active )
1766             pthread_join( h->thread[i]->thread_handle, NULL );
1767     }
1768
1769 #ifdef DEBUG_BENCHMARK
1770     x264_log( h, X264_LOG_INFO,
1771               "analyse=%d(%lldms) encode=%d(%lldms) write=%d(%lldms) filter=%d(%lldms)\n",
1772               (int)(100*i_mtime_analyse/i_mtime_total), i_mtime_analyse/1000,
1773               (int)(100*i_mtime_encode/i_mtime_total), i_mtime_encode/1000,
1774               (int)(100*i_mtime_write/i_mtime_total), i_mtime_write/1000,
1775               (int)(100*i_mtime_filter/i_mtime_total), i_mtime_filter/1000 );
1776 #endif
1777
1778     /* Slices used and PSNR */
1779     for( i=0; i<5; i++ )
1780     {
1781         static const int slice_order[] = { SLICE_TYPE_I, SLICE_TYPE_SI, SLICE_TYPE_P, SLICE_TYPE_SP, SLICE_TYPE_B };
1782         static const char *slice_name[] = { "P", "B", "I", "SP", "SI" };
1783         int i_slice = slice_order[i];
1784
1785         if( h->stat.i_slice_count[i_slice] > 0 )
1786         {
1787             const int i_count = h->stat.i_slice_count[i_slice];
1788             if( h->param.analyse.b_psnr )
1789             {
1790                 x264_log( h, X264_LOG_INFO,
1791                           "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",
1792                           slice_name[i_slice],
1793                           i_count,
1794                           (double)h->stat.i_slice_qp[i_slice] / i_count,
1795                           (double)h->stat.i_slice_size[i_slice] / i_count,
1796                           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,
1797                           h->stat.f_psnr_average[i_slice] / i_count,
1798                           x264_psnr( h->stat.i_sqe_global[i_slice], i_count * i_yuv_size ) );
1799             }
1800             else
1801             {
1802                 x264_log( h, X264_LOG_INFO,
1803                           "slice %s:%-5d Avg QP:%5.2f  size:%6.0f\n",
1804                           slice_name[i_slice],
1805                           i_count,
1806                           (double)h->stat.i_slice_qp[i_slice] / i_count,
1807                           (double)h->stat.i_slice_size[i_slice] / i_count );
1808             }
1809         }
1810     }
1811
1812     /* MB types used */
1813     if( h->stat.i_slice_count[SLICE_TYPE_I] > 0 )
1814     {
1815         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_I];
1816         const double i_count = h->stat.i_slice_count[SLICE_TYPE_I] * h->mb.i_mb_count / 100.0;
1817         x264_log( h, X264_LOG_INFO,
1818                   "mb I  I16..4: %4.1f%% %4.1f%% %4.1f%%\n",
1819                   i_mb_count[I_16x16]/ i_count,
1820                   i_mb_count[I_8x8]  / i_count,
1821                   i_mb_count[I_4x4]  / i_count );
1822     }
1823     if( h->stat.i_slice_count[SLICE_TYPE_P] > 0 )
1824     {
1825         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_P];
1826         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_P];
1827         const double i_count = h->stat.i_slice_count[SLICE_TYPE_P] * h->mb.i_mb_count / 100.0;
1828         x264_log( h, X264_LOG_INFO,
1829                   "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",
1830                   i_mb_count[I_16x16]/ i_count,
1831                   i_mb_count[I_8x8]  / i_count,
1832                   i_mb_count[I_4x4]  / i_count,
1833                   i_mb_size[PIXEL_16x16] / (i_count*4),
1834                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
1835                   i_mb_size[PIXEL_8x8] / (i_count*4),
1836                   (i_mb_size[PIXEL_8x4] + i_mb_size[PIXEL_4x8]) / (i_count*4),
1837                   i_mb_size[PIXEL_4x4] / (i_count*4),
1838                   i_mb_count[P_SKIP] / i_count );
1839     }
1840     if( h->stat.i_slice_count[SLICE_TYPE_B] > 0 )
1841     {
1842         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_B];
1843         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_B];
1844         const double i_count = h->stat.i_slice_count[SLICE_TYPE_B] * h->mb.i_mb_count / 100.0;
1845         x264_log( h, X264_LOG_INFO,
1846                   "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",
1847                   i_mb_count[I_16x16]  / i_count,
1848                   i_mb_count[I_8x8]    / i_count,
1849                   i_mb_count[I_4x4]    / i_count,
1850                   i_mb_size[PIXEL_16x16] / (i_count*4),
1851                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
1852                   i_mb_size[PIXEL_8x8] / (i_count*4),
1853                   i_mb_count[B_DIRECT] / i_count,
1854                   i_mb_count[B_SKIP]   / i_count );
1855     }
1856
1857     x264_ratecontrol_summary( h );
1858
1859     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 )
1860     {
1861         const int i_count = h->stat.i_slice_count[SLICE_TYPE_I] +
1862                             h->stat.i_slice_count[SLICE_TYPE_P] +
1863                             h->stat.i_slice_count[SLICE_TYPE_B];
1864         float fps = (float) h->param.i_fps_num / h->param.i_fps_den;
1865 #define SUM3(p) (p[SLICE_TYPE_I] + p[SLICE_TYPE_P] + p[SLICE_TYPE_B])
1866 #define SUM3b(p,o) (p[SLICE_TYPE_I][o] + p[SLICE_TYPE_P][o] + p[SLICE_TYPE_B][o])
1867         float f_bitrate = fps * SUM3(h->stat.i_slice_size) / i_count / 125;
1868
1869         if( h->pps->b_transform_8x8_mode )
1870         {
1871             int64_t i_i8x8 = SUM3b( h->stat.i_mb_count, I_8x8 );
1872             int64_t i_intra = i_i8x8 + SUM3b( h->stat.i_mb_count, I_4x4 )
1873                                      + SUM3b( h->stat.i_mb_count, I_16x16 );
1874             x264_log( h, X264_LOG_INFO, "8x8 transform  intra:%.1f%%  inter:%.1f%%\n",
1875                       100. * i_i8x8 / i_intra,
1876                       100. * h->stat.i_mb_count_8x8dct[1] / h->stat.i_mb_count_8x8dct[0] );
1877         }
1878
1879         if( h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
1880             && h->stat.i_slice_count[SLICE_TYPE_B] )
1881         {
1882             x264_log( h, X264_LOG_INFO, "direct mvs  spatial:%.1f%%  temporal:%.1f%%\n",
1883                       h->stat.i_direct_frames[1] * 100. / h->stat.i_slice_count[SLICE_TYPE_B],
1884                       h->stat.i_direct_frames[0] * 100. / h->stat.i_slice_count[SLICE_TYPE_B] );
1885         }
1886
1887         if( h->frames.i_max_ref0 > 1 )
1888         {
1889             int i_slice;
1890             for( i_slice = 0; i_slice < 2; i_slice++ )
1891             {
1892                 char buf[200];
1893                 char *p = buf;
1894                 int64_t i_den = 0;
1895                 int i_max = 0;
1896                 for( i = 0; i < h->frames.i_max_ref0 << h->param.b_interlaced; i++ )
1897                     if( h->stat.i_mb_count_ref[i_slice][i] )
1898                     {
1899                         i_den += h->stat.i_mb_count_ref[i_slice][i];
1900                         i_max = i;
1901                     }
1902                 if( i_max == 0 )
1903                     continue;
1904                 for( i = 0; i <= i_max; i++ )
1905                     p += sprintf( p, " %4.1f%%", 100. * h->stat.i_mb_count_ref[i_slice][i] / i_den );
1906                 x264_log( h, X264_LOG_INFO, "ref %c %s\n", i_slice==SLICE_TYPE_P ? 'P' : 'B', buf );
1907             }
1908         }
1909
1910         if( h->param.analyse.b_ssim )
1911         {
1912             x264_log( h, X264_LOG_INFO,
1913                       "SSIM Mean Y:%.7f\n",
1914                       SUM3( h->stat.f_ssim_mean_y ) / i_count );
1915         }
1916         if( h->param.analyse.b_psnr )
1917         {
1918             x264_log( h, X264_LOG_INFO,
1919                       "PSNR Mean Y:%6.3f U:%6.3f V:%6.3f Avg:%6.3f Global:%6.3f kb/s:%.2f\n",
1920                       SUM3( h->stat.f_psnr_mean_y ) / i_count,
1921                       SUM3( h->stat.f_psnr_mean_u ) / i_count,
1922                       SUM3( h->stat.f_psnr_mean_v ) / i_count,
1923                       SUM3( h->stat.f_psnr_average ) / i_count,
1924                       x264_psnr( SUM3( h->stat.i_sqe_global ), i_count * i_yuv_size ),
1925                       f_bitrate );
1926         }
1927         else
1928             x264_log( h, X264_LOG_INFO, "kb/s:%.1f\n", f_bitrate );
1929     }
1930
1931     /* frames */
1932     for( i = 0; h->frames.current[i]; i++ )
1933         x264_frame_delete( h->frames.current[i] );
1934     for( i = 0; h->frames.next[i]; i++ )
1935         x264_frame_delete( h->frames.next[i] );
1936     for( i = 0; h->frames.unused[i]; i++ )
1937         x264_frame_delete( h->frames.unused[i] );
1938     for( i = 0; h->frames.reference[i]; i++ )
1939         x264_frame_delete( h->frames.reference[i] );
1940
1941     /* rc */
1942     x264_ratecontrol_delete( h );
1943
1944     /* param */
1945     if( h->param.rc.psz_stat_out )
1946         free( h->param.rc.psz_stat_out );
1947     if( h->param.rc.psz_stat_in )
1948         free( h->param.rc.psz_stat_in );
1949     if( h->param.rc.psz_rc_eq )
1950         free( h->param.rc.psz_rc_eq );
1951
1952     x264_cqm_delete( h );
1953     for( i = h->param.i_threads - 1; i >= 0; i-- )
1954     {
1955         x264_macroblock_cache_end( h->thread[i] );
1956         x264_free( h->thread[i]->out.p_bitstream );
1957         x264_free( h->thread[i] );
1958     }
1959 }