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 <math.h>
  25
  26 #include "common/common.h"
  27 #include "common/cpu.h"
  28
  29 #include "set.h"
  30 #include "analyse.h"
  31 #include "ratecontrol.h"
  32 #include "macroblock.h"
  33
  34 #if VISUALIZE
  35 #include "common/visualize.h"
  36 #endif
  37
  38 //#define DEBUG_MB_TYPE
  39 //#define DEBUG_DUMP_FRAME
  40 //#define DEBUG_BENCHMARK
  41
  42 #ifdef DEBUG_BENCHMARK
  43 static int64_t i_mtime_encode_frame = 0;
  44 static int64_t i_mtime_analyse = 0;
  45 static int64_t i_mtime_encode = 0;
  46 static int64_t i_mtime_write = 0;
  47 static int64_t i_mtime_filter = 0;
  48 #define TIMER_START( d ) \
  49     { \
  50         int64_t d##start = x264_mdate();
  51
  52 #define TIMER_STOP( d ) \
  53         d += x264_mdate() - d##start;\
  54     }
  55 #else
  56 #define TIMER_START( d )
  57 #define TIMER_STOP( d )
  58 #endif
  59
  60 #define NALU_OVERHEAD 5 // startcode + NAL type costs 5 bytes per frame
  61
  62 static void x264_encoder_frame_end( x264_t *h, x264_t *thread_current,
  63                                     x264_nal_t **pp_nal, int *pi_nal,
  64                                     x264_picture_t *pic_out );
  65
  66 /****************************************************************************
  67  *
  68  ******************************* x264 libs **********************************
  69  *
  70  ****************************************************************************/
  71 static float x264_psnr( int64_t i_sqe, int64_t i_size )
  72 {
  73     double f_mse = (double)i_sqe / ((double)65025.0 * (double)i_size);
  74     if( f_mse <= 0.0000000001 ) /* Max 100dB */
  75         return 100;
  76
  77     return (float)(-10.0 * log( f_mse ) / log( 10.0 ));
  78 }
  79
  80 #ifdef DEBUG_DUMP_FRAME
  81 static void x264_frame_dump( x264_t *h, x264_frame_t *fr, char *name )
  82 {
  83     FILE *f = fopen( name, "r+b" );
  84     int i, y;
  85     if( !f )
  86         return;
  87
  88     /* Write the frame in display order */
  89     fseek( f, fr->i_frame * h->param.i_height * h->param.i_width * 3 / 2, SEEK_SET );
  90
  91     for( i = 0; i < fr->i_plane; i++ )
  92     {
  93         for( y = 0; y < h->param.i_height / ( i == 0 ? 1 : 2 ); y++ )
  94         {
  95             fwrite( &fr->plane[i][y*fr->i_stride[i]], 1, h->param.i_width / ( i == 0 ? 1 : 2 ), f );
  96         }
  97     }
  98     fclose( f );
  99 }
 100 #endif
 101
 102
 103 /* Fill "default" values */
 104 static void x264_slice_header_init( x264_t *h, x264_slice_header_t *sh,
 105                                     x264_sps_t *sps, x264_pps_t *pps,
 106                                     int i_idr_pic_id, int i_frame, int i_qp )
 107 {
 108     x264_param_t *param = &h->param;
 109     int i;
 110
 111     /* First we fill all field */
 112     sh->sps = sps;
 113     sh->pps = pps;
 114
 115     sh->i_first_mb  = 0;
 116     sh->i_last_mb   = h->sps->i_mb_width * h->sps->i_mb_height;
 117     sh->i_pps_id    = pps->i_id;
 118
 119     sh->i_frame_num = i_frame;
 120
 121     sh->b_mbaff = h->param.b_interlaced;
 122     sh->b_field_pic = 0;    /* no field support for now */
 123     sh->b_bottom_field = 0; /* not yet used */
 124
 125     sh->i_idr_pic_id = i_idr_pic_id;
 126
 127     /* poc stuff, fixed later */
 128     sh->i_poc_lsb = 0;
 129     sh->i_delta_poc_bottom = 0;
 130     sh->i_delta_poc[0] = 0;
 131     sh->i_delta_poc[1] = 0;
 132
 133     sh->i_redundant_pic_cnt = 0;
 134
 135     if( !h->mb.b_direct_auto_read )
 136     {
 137         if( h->mb.b_direct_auto_write )
 138             sh->b_direct_spatial_mv_pred = ( h->stat.i_direct_score[1] > h->stat.i_direct_score[0] );
 139         else
 140             sh->b_direct_spatial_mv_pred = ( param->analyse.i_direct_mv_pred == X264_DIRECT_PRED_SPATIAL );
 141     }
 142     /* else b_direct_spatial_mv_pred was read from the 2pass statsfile */
 143
 144     sh->b_num_ref_idx_override = 0;
 145     sh->i_num_ref_idx_l0_active = 1;
 146     sh->i_num_ref_idx_l1_active = 1;
 147
 148     sh->b_ref_pic_list_reordering_l0 = h->b_ref_reorder[0];
 149     sh->b_ref_pic_list_reordering_l1 = h->b_ref_reorder[1];
 150
 151     /* If the ref list isn't in the default order, construct reordering header */
 152     /* List1 reordering isn't needed yet */
 153     if( sh->b_ref_pic_list_reordering_l0 )
 154     {
 155         int pred_frame_num = i_frame;
 156         for( i = 0; i < h->i_ref0; i++ )
 157         {
 158             int diff = h->fref0[i]->i_frame_num - pred_frame_num;
 159             if( diff == 0 )
 160                 x264_log( h, X264_LOG_ERROR, "diff frame num == 0\n" );
 161             sh->ref_pic_list_order[0][i].idc = ( diff > 0 );
 162             sh->ref_pic_list_order[0][i].arg = abs( diff ) - 1;
 163             pred_frame_num = h->fref0[i]->i_frame_num;
 164         }
 165     }
 166
 167     sh->i_cabac_init_idc = param->i_cabac_init_idc;
 168
 169     sh->i_qp = i_qp;
 170     sh->i_qp_delta = i_qp - pps->i_pic_init_qp;
 171     sh->b_sp_for_swidth = 0;
 172     sh->i_qs_delta = 0;
 173
 174     /* If effective qp <= 15, deblocking would have no effect anyway */
 175     if( param->b_deblocking_filter
 176         && ( h->mb.b_variable_qp
 177         || 15 < i_qp + 2 * X264_MAX(param->i_deblocking_filter_alphac0, param->i_deblocking_filter_beta) ) )
 178     {
 179         sh->i_disable_deblocking_filter_idc = 0;
 180     }
 181     else
 182     {
 183         sh->i_disable_deblocking_filter_idc = 1;
 184     }
 185     sh->i_alpha_c0_offset = param->i_deblocking_filter_alphac0 << 1;
 186     sh->i_beta_offset = param->i_deblocking_filter_beta << 1;
 187 }
 188
 189 static void x264_slice_header_write( bs_t *s, x264_slice_header_t *sh, int i_nal_ref_idc )
 190 {
 191     int i;
 192
 193     if( sh->b_mbaff )
 194     {
 195         assert( sh->i_first_mb % (2*sh->sps->i_mb_width) == 0 );
 196         bs_write_ue( s, sh->i_first_mb >> 1 );
 197     }
 198     else
 199         bs_write_ue( s, sh->i_first_mb );
 200
 201     bs_write_ue( s, sh->i_type + 5 );   /* same type things */
 202     bs_write_ue( s, sh->i_pps_id );
 203     bs_write( s, sh->sps->i_log2_max_frame_num, sh->i_frame_num );
 204
 205     if( !sh->sps->b_frame_mbs_only )
 206     {
 207         bs_write1( s, sh->b_field_pic );
 208         if ( sh->b_field_pic )
 209             bs_write1( s, sh->b_bottom_field );
 210     }
 211
 212     if( sh->i_idr_pic_id >= 0 ) /* NAL IDR */
 213     {
 214         bs_write_ue( s, sh->i_idr_pic_id );
 215     }
 216
 217     if( sh->sps->i_poc_type == 0 )
 218     {
 219         bs_write( s, sh->sps->i_log2_max_poc_lsb, sh->i_poc_lsb );
 220         if( sh->pps->b_pic_order && !sh->b_field_pic )
 221         {
 222             bs_write_se( s, sh->i_delta_poc_bottom );
 223         }
 224     }
 225     else if( sh->sps->i_poc_type == 1 && !sh->sps->b_delta_pic_order_always_zero )
 226     {
 227         bs_write_se( s, sh->i_delta_poc[0] );
 228         if( sh->pps->b_pic_order && !sh->b_field_pic )
 229         {
 230             bs_write_se( s, sh->i_delta_poc[1] );
 231         }
 232     }
 233
 234     if( sh->pps->b_redundant_pic_cnt )
 235     {
 236         bs_write_ue( s, sh->i_redundant_pic_cnt );
 237     }
 238
 239     if( sh->i_type == SLICE_TYPE_B )
 240     {
 241         bs_write1( s, sh->b_direct_spatial_mv_pred );
 242     }
 243     if( sh->i_type == SLICE_TYPE_P || sh->i_type == SLICE_TYPE_SP || sh->i_type == SLICE_TYPE_B )
 244     {
 245         bs_write1( s, sh->b_num_ref_idx_override );
 246         if( sh->b_num_ref_idx_override )
 247         {
 248             bs_write_ue( s, sh->i_num_ref_idx_l0_active - 1 );
 249             if( sh->i_type == SLICE_TYPE_B )
 250             {
 251                 bs_write_ue( s, sh->i_num_ref_idx_l1_active - 1 );
 252             }
 253         }
 254     }
 255
 256     /* ref pic list reordering */
 257     if( sh->i_type != SLICE_TYPE_I )
 258     {
 259         bs_write1( s, sh->b_ref_pic_list_reordering_l0 );
 260         if( sh->b_ref_pic_list_reordering_l0 )
 261         {
 262             for( i = 0; i < sh->i_num_ref_idx_l0_active; i++ )
 263             {
 264                 bs_write_ue( s, sh->ref_pic_list_order[0][i].idc );
 265                 bs_write_ue( s, sh->ref_pic_list_order[0][i].arg );
 266
 267             }
 268             bs_write_ue( s, 3 );
 269         }
 270     }
 271     if( sh->i_type == SLICE_TYPE_B )
 272     {
 273         bs_write1( s, sh->b_ref_pic_list_reordering_l1 );
 274         if( sh->b_ref_pic_list_reordering_l1 )
 275         {
 276             for( i = 0; i < sh->i_num_ref_idx_l1_active; i++ )
 277             {
 278                 bs_write_ue( s, sh->ref_pic_list_order[1][i].idc );
 279                 bs_write_ue( s, sh->ref_pic_list_order[1][i].arg );
 280             }
 281             bs_write_ue( s, 3 );
 282         }
 283     }
 284
 285     if( ( sh->pps->b_weighted_pred && ( sh->i_type == SLICE_TYPE_P || sh->i_type == SLICE_TYPE_SP ) ) ||
 286         ( sh->pps->b_weighted_bipred == 1 && sh->i_type == SLICE_TYPE_B ) )
 287     {
 288         /* FIXME */
 289     }
 290
 291     if( i_nal_ref_idc != 0 )
 292     {
 293         if( sh->i_idr_pic_id >= 0 )
 294         {
 295             bs_write1( s, 0 );  /* no output of prior pics flag */
 296             bs_write1( s, 0 );  /* long term reference flag */
 297         }
 298         else
 299         {
 300             bs_write1( s, 0 );  /* adaptive_ref_pic_marking_mode_flag */
 301         }
 302     }
 303
 304     if( sh->pps->b_cabac && sh->i_type != SLICE_TYPE_I )
 305     {
 306         bs_write_ue( s, sh->i_cabac_init_idc );
 307     }
 308     bs_write_se( s, sh->i_qp_delta );      /* slice qp delta */
 309
 310     if( sh->pps->b_deblocking_filter_control )
 311     {
 312         bs_write_ue( s, sh->i_disable_deblocking_filter_idc );
 313         if( sh->i_disable_deblocking_filter_idc != 1 )
 314         {
 315             bs_write_se( s, sh->i_alpha_c0_offset >> 1 );
 316             bs_write_se( s, sh->i_beta_offset >> 1 );
 317         }
 318     }
 319 }
 320
 321 /****************************************************************************
 322  *
 323  ****************************************************************************
 324  ****************************** External API*********************************
 325  ****************************************************************************
 326  *
 327  ****************************************************************************/
 328
 329 static int x264_validate_parameters( x264_t *h )
 330 {
 331     if( h->param.i_width <= 0 || h->param.i_height <= 0 )
 332     {
 333         x264_log( h, X264_LOG_ERROR, "invalid width x height (%dx%d)\n",
 334                   h->param.i_width, h->param.i_height );
 335         return -1;
 336     }
 337
 338     if( h->param.i_width % 2 || h->param.i_height % 2 )
 339     {
 340         x264_log( h, X264_LOG_ERROR, "width or height not divisible by 2 (%dx%d)\n",
 341                   h->param.i_width, h->param.i_height );
 342         return -1;
 343     }
 344     if( h->param.i_csp != X264_CSP_I420 )
 345     {
 346         x264_log( h, X264_LOG_ERROR, "invalid CSP (only I420 supported)\n" );
 347         return -1;
 348     }
 349
 350     if( h->param.i_threads == 0 )
 351         h->param.i_threads = x264_cpu_num_processors() * 3/2;
 352     h->param.i_threads = x264_clip3( h->param.i_threads, 1, X264_THREAD_MAX );
 353     if( h->param.i_threads > 1 )
 354     {
 355 #ifndef HAVE_PTHREAD
 356         x264_log( h, X264_LOG_WARNING, "not compiled with pthread support!\n");
 357         h->param.i_threads = 1;
 358 #else
 359         if( h->param.i_scenecut_threshold >= 0 )
 360             h->param.b_pre_scenecut = 1;
 361 #endif
 362     }
 363
 364     if( h->param.b_interlaced )
 365     {
 366         if( h->param.analyse.i_me_method >= X264_ME_ESA )
 367         {
 368             x264_log( h, X264_LOG_WARNING, "interlace + me=esa is not implemented\n" );
 369             h->param.analyse.i_me_method = X264_ME_UMH;
 370         }
 371         if( h->param.analyse.i_direct_mv_pred > X264_DIRECT_PRED_SPATIAL )
 372         {
 373             x264_log( h, X264_LOG_WARNING, "interlace + direct=temporal is not implemented\n" );
 374             h->param.analyse.i_direct_mv_pred = X264_DIRECT_PRED_SPATIAL;
 375         }
 376     }
 377
 378     if( h->param.rc.i_rc_method < 0 || h->param.rc.i_rc_method > 2 )
 379     {
 380         x264_log( h, X264_LOG_ERROR, "no ratecontrol method specified\n" );
 381         return -1;
 382     }
 383     h->param.rc.f_rf_constant = x264_clip3f( h->param.rc.f_rf_constant, 0, 51 );
 384     h->param.rc.i_qp_constant = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
 385     if( h->param.rc.i_rc_method == X264_RC_CRF )
 386         h->param.rc.i_qp_constant = h->param.rc.f_rf_constant;
 387     if( (h->param.rc.i_rc_method == X264_RC_CQP || h->param.rc.i_rc_method == X264_RC_CRF)
 388         && h->param.rc.i_qp_constant == 0 )
 389     {
 390         h->mb.b_lossless = 1;
 391         h->param.i_cqm_preset = X264_CQM_FLAT;
 392         h->param.psz_cqm_file = NULL;
 393         h->param.rc.i_rc_method = X264_RC_CQP;
 394         h->param.rc.f_ip_factor = 1;
 395         h->param.rc.f_pb_factor = 1;
 396         h->param.analyse.b_transform_8x8 = 0;
 397         h->param.analyse.b_psnr = 0;
 398         h->param.analyse.b_ssim = 0;
 399         h->param.analyse.i_chroma_qp_offset = 0;
 400         h->param.analyse.i_trellis = 0;
 401         h->param.analyse.b_fast_pskip = 0;
 402         h->param.analyse.i_noise_reduction = 0;
 403         h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 1, 6 );
 404     }
 405     if( h->param.rc.i_rc_method == X264_RC_CQP )
 406     {
 407         float qp_p = h->param.rc.i_qp_constant;
 408         float qp_i = qp_p - 6*log(h->param.rc.f_ip_factor)/log(2);
 409         float qp_b = qp_p + 6*log(h->param.rc.f_pb_factor)/log(2);
 410         h->param.rc.i_qp_min = x264_clip3( (int)(X264_MIN3( qp_p, qp_i, qp_b )), 0, 51 );
 411         h->param.rc.i_qp_max = x264_clip3( (int)(X264_MAX3( qp_p, qp_i, qp_b ) + .999), 0, 51 );
 412     }
 413
 414     if( ( h->param.i_width % 16 || h->param.i_height % 16 )
 415         && h->param.i_height != 1080 && !h->mb.b_lossless )
 416     {
 417         // There's nothing special about 1080 in that the warning still applies to it,
 418         // but chances are the user can't help it if his content is already 1080p,
 419         // so there's no point in warning in that case.
 420         x264_log( h, X264_LOG_WARNING,
 421                   "width or height not divisible by 16 (%dx%d), compression will suffer.\n",
 422                   h->param.i_width, h->param.i_height );
 423     }
 424
 425     h->param.i_frame_reference = x264_clip3( h->param.i_frame_reference, 1, 16 );
 426     if( h->param.i_keyint_max <= 0 )
 427         h->param.i_keyint_max = 1;
 428     h->param.i_keyint_min = x264_clip3( h->param.i_keyint_min, 1, h->param.i_keyint_max/2+1 );
 429
 430     h->param.i_bframe = x264_clip3( h->param.i_bframe, 0, X264_BFRAME_MAX );
 431     h->param.i_bframe_bias = x264_clip3( h->param.i_bframe_bias, -90, 100 );
 432     h->param.b_bframe_pyramid = h->param.b_bframe_pyramid && h->param.i_bframe > 1;
 433     h->param.b_bframe_adaptive = h->param.b_bframe_adaptive && h->param.i_bframe > 0;
 434     h->param.analyse.b_weighted_bipred = h->param.analyse.b_weighted_bipred && h->param.i_bframe > 0;
 435     h->mb.b_direct_auto_write = h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
 436                                 && h->param.i_bframe
 437                                 && ( h->param.rc.b_stat_write || !h->param.rc.b_stat_read );
 438     if( h->param.i_scenecut_threshold < 0 )
 439         h->param.b_pre_scenecut = 0;
 440
 441     h->param.i_deblocking_filter_alphac0 = x264_clip3( h->param.i_deblocking_filter_alphac0, -6, 6 );
 442     h->param.i_deblocking_filter_beta    = x264_clip3( h->param.i_deblocking_filter_beta, -6, 6 );
 443     h->param.analyse.i_luma_deadzone[0] = x264_clip3( h->param.analyse.i_luma_deadzone[0], 0, 32 );
 444     h->param.analyse.i_luma_deadzone[1] = x264_clip3( h->param.analyse.i_luma_deadzone[1], 0, 32 );
 445
 446     h->param.i_cabac_init_idc = x264_clip3( h->param.i_cabac_init_idc, 0, 2 );
 447
 448     if( h->param.i_cqm_preset < X264_CQM_FLAT || h->param.i_cqm_preset > X264_CQM_CUSTOM )
 449         h->param.i_cqm_preset = X264_CQM_FLAT;
 450
 451     if( h->param.analyse.i_me_method < X264_ME_DIA ||
 452         h->param.analyse.i_me_method > X264_ME_TESA )
 453         h->param.analyse.i_me_method = X264_ME_HEX;
 454     if( h->param.analyse.i_me_range < 4 )
 455         h->param.analyse.i_me_range = 4;
 456     if( h->param.analyse.i_me_range > 16 && h->param.analyse.i_me_method <= X264_ME_HEX )
 457         h->param.analyse.i_me_range = 16;
 458     if( h->param.analyse.i_me_method == X264_ME_TESA &&
 459         (h->mb.b_lossless || h->param.analyse.i_subpel_refine <= 1) )
 460         h->param.analyse.i_me_method = X264_ME_ESA;
 461     h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 1, 7 );
 462     h->param.analyse.b_bframe_rdo = h->param.analyse.b_bframe_rdo && h->param.analyse.i_subpel_refine >= 6;
 463     h->param.analyse.b_mixed_references = h->param.analyse.b_mixed_references && h->param.i_frame_reference > 1;
 464     h->param.analyse.inter &= X264_ANALYSE_PSUB16x16|X264_ANALYSE_PSUB8x8|X264_ANALYSE_BSUB16x16|
 465                               X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 466     h->param.analyse.intra &= X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 467     if( !(h->param.analyse.inter & X264_ANALYSE_PSUB16x16) )
 468         h->param.analyse.inter &= ~X264_ANALYSE_PSUB8x8;
 469     if( !h->param.analyse.b_transform_8x8 )
 470     {
 471         h->param.analyse.inter &= ~X264_ANALYSE_I8x8;
 472         h->param.analyse.intra &= ~X264_ANALYSE_I8x8;
 473     }
 474     h->param.analyse.i_chroma_qp_offset = x264_clip3(h->param.analyse.i_chroma_qp_offset, -12, 12);
 475     if( !h->param.b_cabac )
 476         h->param.analyse.i_trellis = 0;
 477     h->param.analyse.i_trellis = x264_clip3( h->param.analyse.i_trellis, 0, 2 );
 478     h->param.analyse.i_noise_reduction = x264_clip3( h->param.analyse.i_noise_reduction, 0, 1<<16 );
 479
 480     {
 481         const x264_level_t *l = x264_levels;
 482         while( l->level_idc != 0 && l->level_idc != h->param.i_level_idc )
 483             l++;
 484         if( l->level_idc == 0 )
 485         {
 486             x264_log( h, X264_LOG_ERROR, "invalid level_idc: %d\n", h->param.i_level_idc );
 487             return -1;
 488         }
 489         if( h->param.analyse.i_mv_range <= 0 )
 490             h->param.analyse.i_mv_range = l->mv_range;
 491         else
 492             h->param.analyse.i_mv_range = x264_clip3(h->param.analyse.i_mv_range, 32, 512);
 493         if( h->param.analyse.i_direct_8x8_inference < 0 )
 494             h->param.analyse.i_direct_8x8_inference = l->direct8x8;
 495     }
 496
 497     if( h->param.i_threads > 1 )
 498     {
 499         int r = h->param.analyse.i_mv_range_thread;
 500         int r2;
 501         if( r <= 0 )
 502         {
 503             // half of the available space is reserved and divided evenly among the threads,
 504             // the rest is allocated to whichever thread is far enough ahead to use it.
 505             // reserving more space increases quality for some videos, but costs more time
 506             // in thread synchronization.
 507             int max_range = (h->param.i_height + X264_THREAD_HEIGHT) / h->param.i_threads - X264_THREAD_HEIGHT;
 508             r = max_range / 2;
 509         }
 510         r = X264_MAX( r, h->param.analyse.i_me_range );
 511         r = X264_MIN( r, h->param.analyse.i_mv_range );
 512         // round up to use the whole mb row
 513         r2 = (r & ~15) + ((-X264_THREAD_HEIGHT) & 15);
 514         if( r2 < r )
 515             r2 += 16;
 516         x264_log( h, X264_LOG_DEBUG, "using mv_range_thread = %d\n", r2 );
 517         h->param.analyse.i_mv_range_thread = r2;
 518     }
 519
 520     if( h->param.rc.f_qblur < 0 )
 521         h->param.rc.f_qblur = 0;
 522     if( h->param.rc.f_complexity_blur < 0 )
 523         h->param.rc.f_complexity_blur = 0;
 524
 525     h->param.i_sps_id &= 31;
 526
 527     if( h->param.i_log_level < X264_LOG_INFO )
 528     {
 529         h->param.analyse.b_psnr = 0;
 530         h->param.analyse.b_ssim = 0;
 531     }
 532
 533     /* ensure the booleans are 0 or 1 so they can be used in math */
 534 #define BOOLIFY(x) h->param.x = !!h->param.x
 535     BOOLIFY( b_cabac );
 536     BOOLIFY( b_deblocking_filter );
 537     BOOLIFY( b_interlaced );
 538     BOOLIFY( analyse.b_transform_8x8 );
 539     BOOLIFY( analyse.i_direct_8x8_inference );
 540     BOOLIFY( analyse.b_bidir_me );
 541     BOOLIFY( analyse.b_chroma_me );
 542     BOOLIFY( analyse.b_fast_pskip );
 543     BOOLIFY( rc.b_stat_write );
 544     BOOLIFY( rc.b_stat_read );
 545 #undef BOOLIFY
 546
 547     return 0;
 548 }
 549
 550 static void mbcmp_init( x264_t *h )
 551 {
 552     int satd = !h->mb.b_lossless && h->param.analyse.i_subpel_refine > 1;
 553     memcpy( h->pixf.mbcmp, satd ? h->pixf.satd : h->pixf.sad, sizeof(h->pixf.mbcmp) );
 554     satd &= h->param.analyse.i_me_method == X264_ME_TESA;
 555     memcpy( h->pixf.fpelcmp, satd ? h->pixf.satd : h->pixf.sad, sizeof(h->pixf.fpelcmp) );
 556     memcpy( h->pixf.fpelcmp_x3, satd ? h->pixf.satd_x3 : h->pixf.sad_x3, sizeof(h->pixf.fpelcmp_x3) );
 557     memcpy( h->pixf.fpelcmp_x4, satd ? h->pixf.satd_x4 : h->pixf.sad_x4, sizeof(h->pixf.fpelcmp_x4) );
 558 }
 559
 560 /****************************************************************************
 561  * x264_encoder_open:
 562  ****************************************************************************/
 563 x264_t *x264_encoder_open   ( x264_param_t *param )
 564 {
 565     x264_t *h = x264_malloc( sizeof( x264_t ) );
 566     int i;
 567
 568     memset( h, 0, sizeof( x264_t ) );
 569
 570     /* Create a copy of param */
 571     memcpy( &h->param, param, sizeof( x264_param_t ) );
 572
 573     if( x264_validate_parameters( h ) < 0 )
 574     {
 575         x264_free( h );
 576         return NULL;
 577     }
 578
 579     if( h->param.psz_cqm_file )
 580         if( x264_cqm_parse_file( h, h->param.psz_cqm_file ) < 0 )
 581         {
 582             x264_free( h );
 583             return NULL;
 584         }
 585
 586     if( h->param.rc.psz_stat_out )
 587         h->param.rc.psz_stat_out = strdup( h->param.rc.psz_stat_out );
 588     if( h->param.rc.psz_stat_in )
 589         h->param.rc.psz_stat_in = strdup( h->param.rc.psz_stat_in );
 590     if( h->param.rc.psz_rc_eq )
 591         h->param.rc.psz_rc_eq = strdup( h->param.rc.psz_rc_eq );
 592
 593     /* VUI */
 594     if( h->param.vui.i_sar_width > 0 && h->param.vui.i_sar_height > 0 )
 595     {
 596         int i_w = param->vui.i_sar_width;
 597         int i_h = param->vui.i_sar_height;
 598
 599         x264_reduce_fraction( &i_w, &i_h );
 600
 601         while( i_w > 65535 || i_h > 65535 )
 602         {
 603             i_w /= 2;
 604             i_h /= 2;
 605         }
 606
 607         h->param.vui.i_sar_width = 0;
 608         h->param.vui.i_sar_height = 0;
 609         if( i_w == 0 || i_h == 0 )
 610         {
 611             x264_log( h, X264_LOG_WARNING, "cannot create valid sample aspect ratio\n" );
 612         }
 613         else
 614         {
 615             x264_log( h, X264_LOG_INFO, "using SAR=%d/%d\n", i_w, i_h );
 616             h->param.vui.i_sar_width = i_w;
 617             h->param.vui.i_sar_height = i_h;
 618         }
 619     }
 620
 621     x264_reduce_fraction( &h->param.i_fps_num, &h->param.i_fps_den );
 622
 623     /* Init x264_t */
 624     h->i_frame = 0;
 625     h->i_frame_num = 0;
 626     h->i_idr_pic_id = 0;
 627
 628     h->sps = &h->sps_array[0];
 629     x264_sps_init( h->sps, h->param.i_sps_id, &h->param );
 630
 631     h->pps = &h->pps_array[0];
 632     x264_pps_init( h->pps, h->param.i_sps_id, &h->param, h->sps);
 633
 634     x264_validate_levels( h );
 635
 636     if( x264_cqm_init( h ) < 0 )
 637     {
 638         x264_free( h );
 639         return NULL;
 640     }
 641
 642     h->mb.i_mb_count = h->sps->i_mb_width * h->sps->i_mb_height;
 643
 644     /* Init frames. */
 645     h->frames.i_delay = h->param.i_bframe + h->param.i_threads - 1;
 646     h->frames.i_max_ref0 = h->param.i_frame_reference;
 647     h->frames.i_max_ref1 = h->sps->vui.i_num_reorder_frames;
 648     h->frames.i_max_dpb  = h->sps->vui.i_max_dec_frame_buffering;
 649     h->frames.b_have_lowres = !h->param.rc.b_stat_read
 650         && ( h->param.rc.i_rc_method == X264_RC_ABR
 651           || h->param.rc.i_rc_method == X264_RC_CRF
 652           || h->param.b_bframe_adaptive
 653           || h->param.b_pre_scenecut );
 654
 655     h->frames.i_last_idr = - h->param.i_keyint_max;
 656     h->frames.i_input    = 0;
 657     h->frames.last_nonb  = NULL;
 658
 659     h->i_ref0 = 0;
 660     h->i_ref1 = 0;
 661
 662     x264_rdo_init( );
 663
 664     /* init CPU functions */
 665     x264_predict_16x16_init( h->param.cpu, h->predict_16x16 );
 666     x264_predict_8x8c_init( h->param.cpu, h->predict_8x8c );
 667     x264_predict_8x8_init( h->param.cpu, h->predict_8x8 );
 668     x264_predict_4x4_init( h->param.cpu, h->predict_4x4 );
 669
 670     x264_pixel_init( h->param.cpu, &h->pixf );
 671     x264_dct_init( h->param.cpu, &h->dctf );
 672     x264_zigzag_init( h->param.cpu, &h->zigzagf, h->param.b_interlaced );
 673     x264_mc_init( h->param.cpu, &h->mc );
 674     x264_csp_init( h->param.cpu, h->param.i_csp, &h->csp );
 675     x264_quant_init( h, h->param.cpu, &h->quantf );
 676     x264_deblock_init( h->param.cpu, &h->loopf );
 677     x264_dct_init_weights();
 678
 679     mbcmp_init( h );
 680
 681     x264_log( h, X264_LOG_INFO, "using cpu capabilities: %s%s%s%s%s%s%s%s%s%s\n",
 682              param->cpu&X264_CPU_MMX ? "MMX " : "",
 683              param->cpu&X264_CPU_MMXEXT ? "MMXEXT " : "",
 684              param->cpu&X264_CPU_SSE ? "SSE " : "",
 685              param->cpu&X264_CPU_SSE2 ? "SSE2 " : "",
 686              param->cpu&X264_CPU_SSE3 ? "SSE3 " : "",
 687              param->cpu&X264_CPU_SSSE3 ? "SSSE3 " : "",
 688              param->cpu&X264_CPU_3DNOW ? "3DNow! " : "",
 689              param->cpu&X264_CPU_ALTIVEC ? "Altivec " : "",
 690              param->cpu&X264_CPU_CACHELINE_SPLIT ?
 691                  param->cpu&X264_CPU_CACHELINE_32 ? "Cache32 " :
 692                  param->cpu&X264_CPU_CACHELINE_64 ? "Cache64 " : "Cache? " : "",
 693              param->cpu ? "" : "none!" );
 694
 695     h->out.i_nal = 0;
 696     h->out.i_bitstream = X264_MAX( 1000000, h->param.i_width * h->param.i_height * 4
 697         * ( h->param.rc.i_rc_method == X264_RC_ABR ? pow( 0.95, h->param.rc.i_qp_min )
 698           : pow( 0.95, h->param.rc.i_qp_constant ) * X264_MAX( 1, h->param.rc.f_ip_factor )));
 699
 700     h->thread[0] = h;
 701     h->i_thread_num = 0;
 702     for( i = 1; i < h->param.i_threads; i++ )
 703         h->thread[i] = x264_malloc( sizeof(x264_t) );
 704
 705     for( i = 0; i < h->param.i_threads; i++ )
 706     {
 707         if( i > 0 )
 708             *h->thread[i] = *h;
 709         h->thread[i]->fdec = x264_frame_pop_unused( h );
 710         h->thread[i]->out.p_bitstream = x264_malloc( h->out.i_bitstream );
 711         if( x264_macroblock_cache_init( h->thread[i] ) < 0 )
 712             return NULL;
 713     }
 714
 715     if( x264_ratecontrol_new( h ) < 0 )
 716         return NULL;
 717
 718 #ifdef DEBUG_DUMP_FRAME
 719     {
 720         /* create or truncate the reconstructed video file */
 721         FILE *f = fopen( "fdec.yuv", "w" );
 722         if( f )
 723             fclose( f );
 724         else
 725         {
 726             x264_log( h, X264_LOG_ERROR, "can't write to fdec.yuv\n" );
 727             x264_free( h );
 728             return NULL;
 729         }
 730     }
 731 #endif
 732
 733     return h;
 734 }
 735
 736 /****************************************************************************
 737  * x264_encoder_reconfig:
 738  ****************************************************************************/
 739 int x264_encoder_reconfig( x264_t *h, x264_param_t *param )
 740 {
 741 #define COPY(var) h->param.var = param->var
 742     COPY( i_frame_reference ); // but never uses more refs than initially specified
 743     COPY( i_bframe_bias );
 744     if( h->param.i_scenecut_threshold >= 0 && param->i_scenecut_threshold >= 0 )
 745         COPY( i_scenecut_threshold ); // can't turn it on or off, only vary the threshold
 746     COPY( b_deblocking_filter );
 747     COPY( i_deblocking_filter_alphac0 );
 748     COPY( i_deblocking_filter_beta );
 749     COPY( analyse.intra );
 750     COPY( analyse.inter );
 751     COPY( analyse.i_direct_mv_pred );
 752     COPY( analyse.i_me_method );
 753     COPY( analyse.i_me_range );
 754     COPY( analyse.i_noise_reduction );
 755     COPY( analyse.i_subpel_refine );
 756     COPY( analyse.i_trellis );
 757     COPY( analyse.b_bidir_me );
 758     COPY( analyse.b_bframe_rdo );
 759     COPY( analyse.b_chroma_me );
 760     COPY( analyse.b_dct_decimate );
 761     COPY( analyse.b_fast_pskip );
 762     COPY( analyse.b_mixed_references );
 763     // can only twiddle these if they were enabled to begin with:
 764     if( h->pps->b_transform_8x8_mode )
 765         COPY( analyse.b_transform_8x8 );
 766     if( h->frames.i_max_ref1 > 1 )
 767         COPY( b_bframe_pyramid );
 768 #undef COPY
 769
 770     mbcmp_init( h );
 771
 772     return x264_validate_parameters( h );
 773 }
 774
 775 /* internal usage */
 776 static void x264_nal_start( x264_t *h, int i_type, int i_ref_idc )
 777 {
 778     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 779
 780     nal->i_ref_idc = i_ref_idc;
 781     nal->i_type    = i_type;
 782
 783     nal->i_payload= 0;
 784     nal->p_payload= &h->out.p_bitstream[bs_pos( &h->out.bs ) / 8];
 785 }
 786 static void x264_nal_end( x264_t *h )
 787 {
 788     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 789     nal->i_payload = &h->out.p_bitstream[bs_pos( &h->out.bs ) / 8] - nal->p_payload;
 790     h->out.i_nal++;
 791 }
 792
 793 /****************************************************************************
 794  * x264_encoder_headers:
 795  ****************************************************************************/
 796 int x264_encoder_headers( x264_t *h, x264_nal_t **pp_nal, int *pi_nal )
 797 {
 798     /* init bitstream context */
 799     h->out.i_nal = 0;
 800     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
 801
 802     /* Put SPS and PPS */
 803     if( h->i_frame == 0 )
 804     {
 805         /* identify ourself */
 806         x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
 807         x264_sei_version_write( h, &h->out.bs );
 808         x264_nal_end( h );
 809
 810         /* generate sequence parameters */
 811         x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
 812         x264_sps_write( &h->out.bs, h->sps );
 813         x264_nal_end( h );
 814
 815         /* generate picture parameters */
 816         x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
 817         x264_pps_write( &h->out.bs, h->pps );
 818         x264_nal_end( h );
 819     }
 820     /* now set output*/
 821     *pi_nal = h->out.i_nal;
 822     *pp_nal = &h->out.nal[0];
 823     h->out.i_nal = 0;
 824
 825     return 0;
 826 }
 827
 828 static inline void x264_reference_build_list( x264_t *h, int i_poc )
 829 {
 830     int i;
 831     int b_ok;
 832
 833     /* build ref list 0/1 */
 834     h->i_ref0 = 0;
 835     h->i_ref1 = 0;
 836     for( i = 0; h->frames.reference[i]; i++ )
 837     {
 838         if( h->frames.reference[i]->i_poc < i_poc )
 839         {
 840             h->fref0[h->i_ref0++] = h->frames.reference[i];
 841         }
 842         else if( h->frames.reference[i]->i_poc > i_poc )
 843         {
 844             h->fref1[h->i_ref1++] = h->frames.reference[i];
 845         }
 846     }
 847
 848     /* Order ref0 from higher to lower poc */
 849     do
 850     {
 851         b_ok = 1;
 852         for( i = 0; i < h->i_ref0 - 1; i++ )
 853         {
 854             if( h->fref0[i]->i_poc < h->fref0[i+1]->i_poc )
 855             {
 856                 XCHG( x264_frame_t*, h->fref0[i], h->fref0[i+1] );
 857                 b_ok = 0;
 858                 break;
 859             }
 860         }
 861     } while( !b_ok );
 862     /* Order ref1 from lower to higher poc (bubble sort) for B-frame */
 863     do
 864     {
 865         b_ok = 1;
 866         for( i = 0; i < h->i_ref1 - 1; i++ )
 867         {
 868             if( h->fref1[i]->i_poc > h->fref1[i+1]->i_poc )
 869             {
 870                 XCHG( x264_frame_t*, h->fref1[i], h->fref1[i+1] );
 871                 b_ok = 0;
 872                 break;
 873             }
 874         }
 875     } while( !b_ok );
 876
 877     /* In the standard, a P-frame's ref list is sorted by frame_num.
 878      * We use POC, but check whether explicit reordering is needed */
 879     h->b_ref_reorder[0] =
 880     h->b_ref_reorder[1] = 0;
 881     if( h->sh.i_type == SLICE_TYPE_P )
 882     {
 883         for( i = 0; i < h->i_ref0 - 1; i++ )
 884             if( h->fref0[i]->i_frame_num < h->fref0[i+1]->i_frame_num )
 885             {
 886                 h->b_ref_reorder[0] = 1;
 887                 break;
 888             }
 889     }
 890
 891     h->i_ref1 = X264_MIN( h->i_ref1, h->frames.i_max_ref1 );
 892     h->i_ref0 = X264_MIN( h->i_ref0, h->frames.i_max_ref0 );
 893     h->i_ref0 = X264_MIN( h->i_ref0, h->param.i_frame_reference ); // if reconfig() has lowered the limit
 894     assert( h->i_ref0 + h->i_ref1 <= 16 );
 895     h->mb.pic.i_fref[0] = h->i_ref0;
 896     h->mb.pic.i_fref[1] = h->i_ref1;
 897 }
 898
 899 static void x264_fdec_filter_row( x264_t *h, int mb_y )
 900 {
 901     /* mb_y is the mb to be encoded next, not the mb to be filtered here */
 902     int b_hpel = h->fdec->b_kept_as_ref;
 903     int b_deblock = !h->sh.i_disable_deblocking_filter_idc;
 904     int b_end = mb_y == h->sps->i_mb_height;
 905     int min_y = mb_y - (1 << h->sh.b_mbaff);
 906 #ifndef DEBUG_DUMP_FRAME
 907     b_deblock &= b_hpel;
 908 #endif
 909     if( mb_y & h->sh.b_mbaff )
 910         return;
 911     if( min_y < 0 )
 912         return;
 913
 914     if( !b_end )
 915     {
 916         int i, j;
 917         for( j=0; j<=h->sh.b_mbaff; j++ )
 918             for( i=0; i<3; i++ )
 919             {
 920                 memcpy( h->mb.intra_border_backup[j][i],
 921                         h->fdec->plane[i] + ((mb_y*16 >> !!i) + j - 1 - h->sh.b_mbaff) * h->fdec->i_stride[i],
 922                         h->sps->i_mb_width*16 >> !!i );
 923             }
 924     }
 925
 926     if( b_deblock )
 927     {
 928         int max_y = b_end ? h->sps->i_mb_height : mb_y;
 929         int y;
 930         for( y = min_y; y < max_y; y += (1 << h->sh.b_mbaff) )
 931             x264_frame_deblock_row( h, y );
 932     }
 933
 934     if( b_hpel )
 935     {
 936         x264_frame_expand_border( h, h->fdec, min_y, b_end );
 937         x264_frame_filter( h, h->fdec, min_y, b_end );
 938         x264_frame_expand_border_filtered( h, h->fdec, min_y, b_end );
 939     }
 940
 941     if( h->param.i_threads > 1 && h->fdec->b_kept_as_ref )
 942     {
 943         x264_frame_cond_broadcast( h->fdec, mb_y*16 + (b_end ? 10000 : -(X264_THREAD_HEIGHT << h->sh.b_mbaff)) );
 944     }
 945 }
 946
 947 static inline void x264_reference_update( x264_t *h )
 948 {
 949     int i;
 950
 951     if( h->fdec->i_frame >= 0 )
 952         h->i_frame++;
 953
 954     if( !h->fdec->b_kept_as_ref )
 955     {
 956         if( h->param.i_threads > 1 )
 957         {
 958             x264_frame_push_unused( h, h->fdec );
 959             h->fdec = x264_frame_pop_unused( h );
 960         }
 961         return;
 962     }
 963
 964     /* move lowres copy of the image to the ref frame */
 965     for( i = 0; i < 4; i++)
 966     {
 967         XCHG( uint8_t*, h->fdec->lowres[i], h->fenc->lowres[i] );
 968         XCHG( uint8_t*, h->fdec->buffer_lowres[i], h->fenc->buffer_lowres[i] );
 969     }
 970
 971     /* adaptive B decision needs a pointer, since it can't use the ref lists */
 972     if( h->sh.i_type != SLICE_TYPE_B )
 973         h->frames.last_nonb = h->fdec;
 974
 975     /* move frame in the buffer */
 976     x264_frame_push( h->frames.reference, h->fdec );
 977     if( h->frames.reference[h->frames.i_max_dpb] )
 978         x264_frame_push_unused( h, x264_frame_shift( h->frames.reference ) );
 979     h->fdec = x264_frame_pop_unused( h );
 980 }
 981
 982 static inline void x264_reference_reset( x264_t *h )
 983 {
 984     while( h->frames.reference[0] )
 985         x264_frame_push_unused( h, x264_frame_pop( h->frames.reference ) );
 986     h->fdec->i_poc =
 987     h->fenc->i_poc = 0;
 988 }
 989
 990 static inline void x264_slice_init( x264_t *h, int i_nal_type, int i_global_qp )
 991 {
 992     /* ------------------------ Create slice header  ----------------------- */
 993     if( i_nal_type == NAL_SLICE_IDR )
 994     {
 995         x264_slice_header_init( h, &h->sh, h->sps, h->pps, h->i_idr_pic_id, h->i_frame_num, i_global_qp );
 996
 997         /* increment id */
 998         h->i_idr_pic_id = ( h->i_idr_pic_id + 1 ) % 65536;
 999     }
1000     else
1001     {
1002         x264_slice_header_init( h, &h->sh, h->sps, h->pps, -1, h->i_frame_num, i_global_qp );
1003
1004         /* always set the real higher num of ref frame used */
1005         h->sh.b_num_ref_idx_override = 1;
1006         h->sh.i_num_ref_idx_l0_active = h->i_ref0 <= 0 ? 1 : h->i_ref0;
1007         h->sh.i_num_ref_idx_l1_active = h->i_ref1 <= 0 ? 1 : h->i_ref1;
1008     }
1009
1010     h->fdec->i_frame_num = h->sh.i_frame_num;
1011
1012     if( h->sps->i_poc_type == 0 )
1013     {
1014         h->sh.i_poc_lsb = h->fdec->i_poc & ( (1 << h->sps->i_log2_max_poc_lsb) - 1 );
1015         h->sh.i_delta_poc_bottom = 0;   /* XXX won't work for field */
1016     }
1017     else if( h->sps->i_poc_type == 1 )
1018     {
1019         /* FIXME TODO FIXME */
1020     }
1021     else
1022     {
1023         /* Nothing to do ? */
1024     }
1025
1026     x264_macroblock_slice_init( h );
1027 }
1028
1029 static void x264_slice_write( x264_t *h )
1030 {
1031     int i_skip;
1032     int mb_xy;
1033     int i;
1034
1035     /* init stats */
1036     memset( &h->stat.frame, 0, sizeof(h->stat.frame) );
1037
1038     /* Slice */
1039     x264_nal_start( h, h->i_nal_type, h->i_nal_ref_idc );
1040
1041     /* Slice header */
1042     x264_slice_header_write( &h->out.bs, &h->sh, h->i_nal_ref_idc );
1043     if( h->param.b_cabac )
1044     {
1045         /* alignment needed */
1046         bs_align_1( &h->out.bs );
1047
1048         /* init cabac */
1049         x264_cabac_context_init( &h->cabac, h->sh.i_type, h->sh.i_qp, h->sh.i_cabac_init_idc );
1050         x264_cabac_encode_init ( &h->cabac, h->out.bs.p, h->out.bs.p_end );
1051     }
1052     h->mb.i_last_qp = h->sh.i_qp;
1053     h->mb.i_last_dqp = 0;
1054
1055     for( mb_xy = h->sh.i_first_mb, i_skip = 0; mb_xy < h->sh.i_last_mb; )
1056     {
1057         const int i_mb_y = mb_xy / h->sps->i_mb_width;
1058         const int i_mb_x = mb_xy % h->sps->i_mb_width;
1059         int mb_spos = bs_pos(&h->out.bs) + x264_cabac_pos(&h->cabac);
1060
1061         if( i_mb_x == 0 )
1062             x264_fdec_filter_row( h, i_mb_y );
1063
1064         /* load cache */
1065         x264_macroblock_cache_load( h, i_mb_x, i_mb_y );
1066
1067         /* analyse parameters
1068          * Slice I: choose I_4x4 or I_16x16 mode
1069          * Slice P: choose between using P mode or intra (4x4 or 16x16)
1070          * */
1071         TIMER_START( i_mtime_analyse );
1072         x264_macroblock_analyse( h );
1073         TIMER_STOP( i_mtime_analyse );
1074
1075         /* encode this macroblock -> be careful it can change the mb type to P_SKIP if needed */
1076         TIMER_START( i_mtime_encode );
1077         x264_macroblock_encode( h );
1078         TIMER_STOP( i_mtime_encode );
1079
1080         TIMER_START( i_mtime_write );
1081         if( h->param.b_cabac )
1082         {
1083             if( mb_xy > h->sh.i_first_mb && !(h->sh.b_mbaff && (i_mb_y&1)) )
1084                 x264_cabac_encode_terminal( &h->cabac );
1085
1086             if( IS_SKIP( h->mb.i_type ) )
1087                 x264_cabac_mb_skip( h, 1 );
1088             else
1089             {
1090                 if( h->sh.i_type != SLICE_TYPE_I )
1091                     x264_cabac_mb_skip( h, 0 );
1092                 x264_macroblock_write_cabac( h, &h->cabac );
1093             }
1094         }
1095         else
1096         {
1097             if( IS_SKIP( h->mb.i_type ) )
1098                 i_skip++;
1099             else
1100             {
1101                 if( h->sh.i_type != SLICE_TYPE_I )
1102                 {
1103                     bs_write_ue( &h->out.bs, i_skip );  /* skip run */
1104                     i_skip = 0;
1105                 }
1106                 x264_macroblock_write_cavlc( h, &h->out.bs );
1107             }
1108         }
1109         TIMER_STOP( i_mtime_write );
1110
1111 #if VISUALIZE
1112         if( h->param.b_visualize )
1113             x264_visualize_mb( h );
1114 #endif
1115
1116         /* save cache */
1117         x264_macroblock_cache_save( h );
1118
1119         /* accumulate mb stats */
1120         h->stat.frame.i_mb_count[h->mb.i_type]++;
1121         if( !IS_SKIP(h->mb.i_type) && !IS_INTRA(h->mb.i_type) && !IS_DIRECT(h->mb.i_type) )
1122         {
1123             if( h->mb.i_partition != D_8x8 )
1124                 h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_partition ] ] += 4;
1125             else
1126                 for( i = 0; i < 4; i++ )
1127                     h->stat.frame.i_mb_count_size[ x264_mb_partition_pixel_table[ h->mb.i_sub_partition[i] ] ] ++;
1128             if( h->param.i_frame_reference > 1 )
1129             {
1130                 for( i = 0; i < 4; i++ )
1131                 {
1132                     int i_ref = h->mb.cache.ref[0][ x264_scan8[4*i] ];
1133                     if( i_ref >= 0 )
1134                         h->stat.frame.i_mb_count_ref[i_ref] ++;
1135                 }
1136             }
1137         }
1138         if( h->mb.i_cbp_luma && !IS_INTRA(h->mb.i_type) )
1139         {
1140             h->stat.frame.i_mb_count_8x8dct[0] ++;
1141             h->stat.frame.i_mb_count_8x8dct[1] += h->mb.b_transform_8x8;
1142         }
1143
1144         if( h->mb.b_variable_qp )
1145             x264_ratecontrol_mb(h, bs_pos(&h->out.bs) + x264_cabac_pos(&h->cabac) - mb_spos);
1146
1147         if( h->sh.b_mbaff )
1148         {
1149             if( (i_mb_y&1) && i_mb_x == h->sps->i_mb_width - 1 )
1150                 mb_xy++;
1151             else if( i_mb_y&1 )
1152                 mb_xy += 1 - h->sps->i_mb_width;
1153             else
1154                 mb_xy += h->sps->i_mb_width;
1155         }
1156         else
1157             mb_xy++;
1158     }
1159
1160     if( h->param.b_cabac )
1161     {
1162         x264_cabac_encode_flush( h, &h->cabac );
1163         h->out.bs.p = h->cabac.p;
1164     }
1165     else
1166     {
1167         if( i_skip > 0 )
1168             bs_write_ue( &h->out.bs, i_skip );  /* last skip run */
1169         /* rbsp_slice_trailing_bits */
1170         bs_rbsp_trailing( &h->out.bs );
1171     }
1172
1173     x264_nal_end( h );
1174
1175     /* Compute misc bits */
1176     h->stat.frame.i_misc_bits = bs_pos( &h->out.bs )
1177                               + NALU_OVERHEAD * 8
1178                               - h->stat.frame.i_itex_bits
1179                               - h->stat.frame.i_ptex_bits
1180                               - h->stat.frame.i_hdr_bits;
1181 }
1182
1183 static void x264_thread_sync_context( x264_t *dst, x264_t *src )
1184 {
1185     x264_frame_t **f;
1186     if( dst == src )
1187         return;
1188
1189     // reference counting
1190     for( f = src->frames.reference; *f; f++ )
1191         (*f)->i_reference_count++;
1192     for( f = dst->frames.reference; *f; f++ )
1193         x264_frame_push_unused( src, *f );
1194     src->fdec->i_reference_count++;
1195     x264_frame_push_unused( src, dst->fdec );
1196
1197     // copy everything except the per-thread pointers and the constants.
1198     memcpy( &dst->i_frame, &src->i_frame, offsetof(x264_t, mb.type) - offsetof(x264_t, i_frame) );
1199     memcpy( &dst->mb.i_type, &src->mb.i_type, offsetof(x264_t, rc) - offsetof(x264_t, mb.i_type) );
1200     dst->stat = src->stat;
1201 }
1202
1203 static void x264_thread_sync_stat( x264_t *dst, x264_t *src )
1204 {
1205     if( dst == src )
1206         return;
1207     memcpy( &dst->stat.i_slice_count, &src->stat.i_slice_count, sizeof(dst->stat) - sizeof(dst->stat.frame) );
1208 }
1209
1210 static int x264_slices_write( x264_t *h )
1211 {
1212     int i_frame_size;
1213
1214 #if VISUALIZE
1215     if( h->param.b_visualize )
1216         x264_visualize_init( h );
1217 #endif
1218
1219     x264_stack_align( x264_slice_write, h );
1220     i_frame_size = h->out.nal[h->out.i_nal-1].i_payload;
1221     x264_fdec_filter_row( h, h->sps->i_mb_height );
1222
1223 #if VISUALIZE
1224     if( h->param.b_visualize )
1225     {
1226         x264_visualize_show( h );
1227         x264_visualize_close( h );
1228     }
1229 #endif
1230
1231     h->out.i_frame_size = i_frame_size;
1232     return 0;
1233 }
1234
1235 /****************************************************************************
1236  * x264_encoder_encode:
1237  *  XXX: i_poc   : is the poc of the current given picture
1238  *       i_frame : is the number of the frame being coded
1239  *  ex:  type frame poc
1240  *       I      0   2*0
1241  *       P      1   2*3
1242  *       B      2   2*1
1243  *       B      3   2*2
1244  *       P      4   2*6
1245  *       B      5   2*4
1246  *       B      6   2*5
1247  ****************************************************************************/
1248 int     x264_encoder_encode( x264_t *h,
1249                              x264_nal_t **pp_nal, int *pi_nal,
1250                              x264_picture_t *pic_in,
1251                              x264_picture_t *pic_out )
1252 {
1253     x264_t *thread_current, *thread_prev, *thread_oldest;
1254     int     i_nal_type;
1255     int     i_nal_ref_idc;
1256
1257     int   i_global_qp;
1258
1259     if( h->param.i_threads > 1)
1260     {
1261         int i = ++h->i_thread_phase;
1262         int t = h->param.i_threads;
1263         thread_current = h->thread[ i%t ];
1264         thread_prev    = h->thread[ (i-1)%t ];
1265         thread_oldest  = h->thread[ (i+1)%t ];
1266         x264_thread_sync_context( thread_current, thread_prev );
1267         x264_thread_sync_ratecontrol( thread_current, thread_prev, thread_oldest );
1268         h = thread_current;
1269 //      fprintf(stderr, "current: %p  prev: %p  oldest: %p \n", thread_current, thread_prev, thread_oldest);
1270     }
1271     else
1272     {
1273         thread_current =
1274         thread_prev    =
1275         thread_oldest  = h;
1276     }
1277
1278     // ok to call this before encoding any frames, since the initial values of fdec have b_kept_as_ref=0
1279     x264_reference_update( h );
1280     h->fdec->i_lines_completed = -1;
1281
1282     /* no data out */
1283     *pi_nal = 0;
1284     *pp_nal = NULL;
1285
1286     /* ------------------- Setup new frame from picture -------------------- */
1287     TIMER_START( i_mtime_encode_frame );
1288     if( pic_in != NULL )
1289     {
1290         /* 1: Copy the picture to a frame and move it to a buffer */
1291         x264_frame_t *fenc = x264_frame_pop_unused( h );
1292
1293         x264_frame_copy_picture( h, fenc, pic_in );
1294
1295         if( h->param.i_width != 16 * h->sps->i_mb_width ||
1296             h->param.i_height != 16 * h->sps->i_mb_height )
1297             x264_frame_expand_border_mod16( h, fenc );
1298
1299         fenc->i_frame = h->frames.i_input++;
1300
1301         x264_frame_push( h->frames.next, fenc );
1302
1303         if( h->frames.b_have_lowres )
1304             x264_frame_init_lowres( h, fenc );
1305
1306         if( h->frames.i_input <= h->frames.i_delay + 1 - h->param.i_threads )
1307         {
1308             /* Nothing yet to encode */
1309             /* waiting for filling bframe buffer */
1310             pic_out->i_type = X264_TYPE_AUTO;
1311             return 0;
1312         }
1313     }
1314
1315     if( h->frames.current[0] == NULL )
1316     {
1317         int bframes = 0;
1318         /* 2: Select frame types */
1319         if( h->frames.next[0] == NULL )
1320         {
1321             x264_encoder_frame_end( thread_oldest, thread_current, pp_nal, pi_nal, pic_out );
1322             return 0;
1323         }
1324
1325         x264_slicetype_decide( h );
1326
1327         /* 3: move some B-frames and 1 non-B to encode queue */
1328         while( IS_X264_TYPE_B( h->frames.next[bframes]->i_type ) )
1329             bframes++;
1330         x264_frame_push( h->frames.current, x264_frame_shift( &h->frames.next[bframes] ) );
1331         /* FIXME: when max B-frames > 3, BREF may no longer be centered after GOP closing */
1332         if( h->param.b_bframe_pyramid && bframes > 1 )
1333         {
1334             x264_frame_t *mid = x264_frame_shift( &h->frames.next[bframes/2] );
1335             mid->i_type = X264_TYPE_BREF;
1336             x264_frame_push( h->frames.current, mid );
1337             bframes--;
1338         }
1339         while( bframes-- )
1340             x264_frame_push( h->frames.current, x264_frame_shift( h->frames.next ) );
1341     }
1342     TIMER_STOP( i_mtime_encode_frame );
1343
1344     /* ------------------- Get frame to be encoded ------------------------- */
1345     /* 4: get picture to encode */
1346     h->fenc = x264_frame_shift( h->frames.current );
1347     if( h->fenc == NULL )
1348     {
1349         /* Nothing yet to encode (ex: waiting for I/P with B frames) */
1350         /* waiting for filling bframe buffer */
1351         pic_out->i_type = X264_TYPE_AUTO;
1352         return 0;
1353     }
1354
1355 do_encode:
1356
1357     if( h->fenc->i_type == X264_TYPE_IDR )
1358     {
1359         h->frames.i_last_idr = h->fenc->i_frame;
1360     }
1361
1362     /* ------------------- Setup frame context ----------------------------- */
1363     /* 5: Init data dependent of frame type */
1364     TIMER_START( i_mtime_encode_frame );
1365     if( h->fenc->i_type == X264_TYPE_IDR )
1366     {
1367         /* reset ref pictures */
1368         x264_reference_reset( h );
1369
1370         i_nal_type    = NAL_SLICE_IDR;
1371         i_nal_ref_idc = NAL_PRIORITY_HIGHEST;
1372         h->sh.i_type = SLICE_TYPE_I;
1373     }
1374     else if( h->fenc->i_type == X264_TYPE_I )
1375     {
1376         i_nal_type    = NAL_SLICE;
1377         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
1378         h->sh.i_type = SLICE_TYPE_I;
1379     }
1380     else if( h->fenc->i_type == X264_TYPE_P )
1381     {
1382         i_nal_type    = NAL_SLICE;
1383         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
1384         h->sh.i_type = SLICE_TYPE_P;
1385     }
1386     else if( h->fenc->i_type == X264_TYPE_BREF )
1387     {
1388         i_nal_type    = NAL_SLICE;
1389         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* maybe add MMCO to forget it? -> low */
1390         h->sh.i_type = SLICE_TYPE_B;
1391     }
1392     else    /* B frame */
1393     {
1394         i_nal_type    = NAL_SLICE;
1395         i_nal_ref_idc = NAL_PRIORITY_DISPOSABLE;
1396         h->sh.i_type = SLICE_TYPE_B;
1397     }
1398
1399     h->fdec->i_poc =
1400     h->fenc->i_poc = 2 * (h->fenc->i_frame - h->frames.i_last_idr);
1401     h->fdec->i_type = h->fenc->i_type;
1402     h->fdec->i_frame = h->fenc->i_frame;
1403     h->fenc->b_kept_as_ref =
1404     h->fdec->b_kept_as_ref = i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE && h->param.i_keyint_max > 1;
1405
1406
1407
1408     /* ------------------- Init                ----------------------------- */
1409     /* build ref list 0/1 */
1410     x264_reference_build_list( h, h->fdec->i_poc );
1411
1412     /* Init the rate control */
1413     x264_ratecontrol_start( h, h->fenc->i_qpplus1 );
1414     i_global_qp = x264_ratecontrol_qp( h );
1415
1416     pic_out->i_qpplus1 =
1417     h->fdec->i_qpplus1 = i_global_qp + 1;
1418
1419     if( h->sh.i_type == SLICE_TYPE_B )
1420         x264_macroblock_bipred_init( h );
1421
1422     /* ------------------------ Create slice header  ----------------------- */
1423     x264_slice_init( h, i_nal_type, i_global_qp );
1424
1425     if( i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE )
1426         h->i_frame_num++;
1427
1428     /* ---------------------- Write the bitstream -------------------------- */
1429     /* Init bitstream context */
1430     h->out.i_nal = 0;
1431     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
1432
1433     if(h->param.b_aud){
1434         int pic_type;
1435
1436         if(h->sh.i_type == SLICE_TYPE_I)
1437             pic_type = 0;
1438         else if(h->sh.i_type == SLICE_TYPE_P)
1439             pic_type = 1;
1440         else if(h->sh.i_type == SLICE_TYPE_B)
1441             pic_type = 2;
1442         else
1443             pic_type = 7;
1444
1445         x264_nal_start(h, NAL_AUD, NAL_PRIORITY_DISPOSABLE);
1446         bs_write(&h->out.bs, 3, pic_type);
1447         bs_rbsp_trailing(&h->out.bs);
1448         x264_nal_end(h);
1449     }
1450
1451     h->i_nal_type = i_nal_type;
1452     h->i_nal_ref_idc = i_nal_ref_idc;
1453
1454     /* Write SPS and PPS */
1455     if( i_nal_type == NAL_SLICE_IDR && h->param.b_repeat_headers )
1456     {
1457         if( h->fenc->i_frame == 0 )
1458         {
1459             /* identify ourself */
1460             x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
1461             x264_sei_version_write( h, &h->out.bs );
1462             x264_nal_end( h );
1463         }
1464
1465         /* generate sequence parameters */
1466         x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
1467         x264_sps_write( &h->out.bs, h->sps );
1468         x264_nal_end( h );
1469
1470         /* generate picture parameters */
1471         x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
1472         x264_pps_write( &h->out.bs, h->pps );
1473         x264_nal_end( h );
1474     }
1475
1476     /* Write frame */
1477     if( h->param.i_threads > 1 )
1478     {
1479         x264_pthread_create( &h->thread_handle, NULL, (void*)x264_slices_write, h );
1480         h->b_thread_active = 1;
1481     }
1482     else
1483         x264_slices_write( h );
1484
1485     /* restore CPU state (before using float again) */
1486     x264_cpu_restore( h->param.cpu );
1487
1488     if( h->sh.i_type == SLICE_TYPE_P && !h->param.rc.b_stat_read
1489         && h->param.i_scenecut_threshold >= 0
1490         && !h->param.b_pre_scenecut )
1491     {
1492         const int *mbs = h->stat.frame.i_mb_count;
1493         int i_mb_i = mbs[I_16x16] + mbs[I_8x8] + mbs[I_4x4];
1494         int i_mb_p = mbs[P_L0] + mbs[P_8x8];
1495         int i_mb_s = mbs[P_SKIP];
1496         int i_mb   = h->sps->i_mb_width * h->sps->i_mb_height;
1497         int64_t i_inter_cost = h->stat.frame.i_inter_cost;
1498         int64_t i_intra_cost = h->stat.frame.i_intra_cost;
1499
1500         float f_bias;
1501         int i_gop_size = h->fenc->i_frame - h->frames.i_last_idr;
1502         float f_thresh_max = h->param.i_scenecut_threshold / 100.0;
1503         /* magic numbers pulled out of thin air */
1504         float f_thresh_min = f_thresh_max * h->param.i_keyint_min
1505                              / ( h->param.i_keyint_max * 4 );
1506         if( h->param.i_keyint_min == h->param.i_keyint_max )
1507              f_thresh_min= f_thresh_max;
1508
1509         /* macroblock_analyse() doesn't further analyse skipped mbs,
1510          * so we have to guess their cost */
1511         if( h->stat.frame.i_mbs_analysed > 0 )
1512             i_intra_cost = i_intra_cost * i_mb / h->stat.frame.i_mbs_analysed;
1513
1514         if( i_gop_size < h->param.i_keyint_min / 4 )
1515             f_bias = f_thresh_min / 4;
1516         else if( i_gop_size <= h->param.i_keyint_min )
1517             f_bias = f_thresh_min * i_gop_size / h->param.i_keyint_min;
1518         else
1519         {
1520             f_bias = f_thresh_min
1521                      + ( f_thresh_max - f_thresh_min )
1522                        * ( i_gop_size - h->param.i_keyint_min )
1523                        / ( h->param.i_keyint_max - h->param.i_keyint_min );
1524         }
1525         f_bias = X264_MIN( f_bias, 1.0 );
1526
1527         /* Bad P will be reencoded as I */
1528         if( h->stat.frame.i_mbs_analysed > 0 &&
1529             i_inter_cost >= (1.0 - f_bias) * i_intra_cost )
1530         {
1531             int b;
1532
1533             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",
1534                       h->fenc->i_frame,
1535                       (double)i_intra_cost, (double)i_inter_cost,
1536                       1. - (double)i_inter_cost / i_intra_cost,
1537                       f_bias, i_gop_size,
1538                       i_mb_i, i_mb_p, i_mb_s );
1539
1540             /* Restore frame num */
1541             h->i_frame_num--;
1542
1543             for( b = 0; h->frames.current[b] && IS_X264_TYPE_B( h->frames.current[b]->i_type ); b++ );
1544             if( b > 0 )
1545             {
1546                 /* If using B-frames, force GOP to be closed.
1547                  * Even if this frame is going to be I and not IDR, forcing a
1548                  * P-frame before the scenecut will probably help compression.
1549                  *
1550                  * We don't yet know exactly which frame is the scene cut, so
1551                  * we can't assign an I-frame. Instead, change the previous
1552                  * B-frame to P, and rearrange coding order. */
1553
1554                 if( h->param.b_bframe_adaptive || b > 1 )
1555                     h->fenc->i_type = X264_TYPE_AUTO;
1556                 x264_frame_sort_pts( h->frames.current );
1557                 x264_frame_unshift( h->frames.next, h->fenc );
1558                 h->fenc = h->frames.current[b-1];
1559                 h->frames.current[b-1] = NULL;
1560                 h->fenc->i_type = X264_TYPE_P;
1561                 x264_frame_sort_dts( h->frames.current );
1562             }
1563             /* Do IDR if needed */
1564             else if( i_gop_size >= h->param.i_keyint_min )
1565             {
1566                 /* Reset */
1567                 h->i_frame_num = 0;
1568
1569                 /* Reinit field of fenc */
1570                 h->fenc->i_type = X264_TYPE_IDR;
1571                 h->fenc->i_poc = 0;
1572
1573                 /* Put enqueued frames back in the pool */
1574                 while( h->frames.current[0] )
1575                     x264_frame_push( h->frames.next, x264_frame_shift( h->frames.current ) );
1576                 x264_frame_sort_pts( h->frames.next );
1577             }
1578             else
1579             {
1580                 h->fenc->i_type = X264_TYPE_I;
1581             }
1582             goto do_encode;
1583         }
1584     }
1585
1586     x264_encoder_frame_end( thread_oldest, thread_current, pp_nal, pi_nal, pic_out );
1587     return 0;
1588 }
1589
1590 static void x264_encoder_frame_end( x264_t *h, x264_t *thread_current,
1591                                     x264_nal_t **pp_nal, int *pi_nal,
1592                                     x264_picture_t *pic_out )
1593 {
1594     int i;
1595     char psz_message[80];
1596
1597     if( h->b_thread_active )
1598     {
1599         x264_pthread_join( h->thread_handle, NULL );
1600         h->b_thread_active = 0;
1601     }
1602     if( !h->out.i_nal )
1603     {
1604         pic_out->i_type = X264_TYPE_AUTO;
1605         return;
1606     }
1607
1608     x264_frame_push_unused( thread_current, h->fenc );
1609
1610     /* End bitstream, set output  */
1611     *pi_nal = h->out.i_nal;
1612     *pp_nal = h->out.nal;
1613     h->out.i_nal = 0;
1614
1615     /* Set output picture properties */
1616     if( h->sh.i_type == SLICE_TYPE_I )
1617         pic_out->i_type = h->i_nal_type == NAL_SLICE_IDR ? X264_TYPE_IDR : X264_TYPE_I;
1618     else if( h->sh.i_type == SLICE_TYPE_P )
1619         pic_out->i_type = X264_TYPE_P;
1620     else
1621         pic_out->i_type = X264_TYPE_B;
1622     pic_out->i_pts = h->fenc->i_pts;
1623
1624     pic_out->img.i_plane = h->fdec->i_plane;
1625     for(i = 0; i < 4; i++){
1626         pic_out->img.i_stride[i] = h->fdec->i_stride[i];
1627         pic_out->img.plane[i] = h->fdec->plane[i];
1628     }
1629
1630     /* ---------------------- Update encoder state ------------------------- */
1631
1632     /* update rc */
1633     x264_cpu_restore( h->param.cpu );
1634     x264_ratecontrol_end( h, h->out.i_frame_size * 8 );
1635
1636     /* restore CPU state (before using float again) */
1637     x264_cpu_restore( h->param.cpu );
1638
1639     x264_noise_reduction_update( h );
1640
1641     TIMER_STOP( i_mtime_encode_frame );
1642
1643     /* ---------------------- Compute/Print statistics --------------------- */
1644     x264_thread_sync_stat( h, h->thread[0] );
1645
1646     /* Slice stat */
1647     h->stat.i_slice_count[h->sh.i_type]++;
1648     h->stat.i_slice_size[h->sh.i_type] += h->out.i_frame_size + NALU_OVERHEAD;
1649     h->stat.i_slice_qp[h->sh.i_type] += h->fdec->i_qpplus1 - 1;
1650
1651     for( i = 0; i < X264_MBTYPE_MAX; i++ )
1652         h->stat.i_mb_count[h->sh.i_type][i] += h->stat.frame.i_mb_count[i];
1653     for( i = 0; i < 2; i++ )
1654         h->stat.i_mb_count_8x8dct[i] += h->stat.frame.i_mb_count_8x8dct[i];
1655     if( h->sh.i_type != SLICE_TYPE_I )
1656     {
1657         for( i = 0; i < 7; i++ )
1658             h->stat.i_mb_count_size[h->sh.i_type][i] += h->stat.frame.i_mb_count_size[i];
1659         for( i = 0; i < 32; i++ )
1660             h->stat.i_mb_count_ref[h->sh.i_type][i] += h->stat.frame.i_mb_count_ref[i];
1661     }
1662     if( h->sh.i_type == SLICE_TYPE_B )
1663     {
1664         h->stat.i_direct_frames[ h->sh.b_direct_spatial_mv_pred ] ++;
1665         if( h->mb.b_direct_auto_write )
1666         {
1667             //FIXME somewhat arbitrary time constants
1668             if( h->stat.i_direct_score[0] + h->stat.i_direct_score[1] > h->mb.i_mb_count )
1669             {
1670                 for( i = 0; i < 2; i++ )
1671                     h->stat.i_direct_score[i] = h->stat.i_direct_score[i] * 9/10;
1672             }
1673             for( i = 0; i < 2; i++ )
1674                 h->stat.i_direct_score[i] += h->stat.frame.i_direct_score[i];
1675         }
1676     }
1677
1678     psz_message[0] = '\0';
1679     if( h->param.analyse.b_psnr )
1680     {
1681         int64_t sqe[3];
1682
1683         for( i=0; i<3; i++ )
1684         {
1685             sqe[i] = x264_pixel_ssd_wxh( &h->pixf,
1686                          h->fdec->plane[i], h->fdec->i_stride[i],
1687                          h->fenc->plane[i], h->fenc->i_stride[i],
1688                          h->param.i_width >> !!i, h->param.i_height >> !!i );
1689         }
1690         x264_cpu_restore( h->param.cpu );
1691
1692         h->stat.i_sqe_global[h->sh.i_type] += sqe[0] + sqe[1] + sqe[2];
1693         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 );
1694         h->stat.f_psnr_mean_y[h->sh.i_type] += x264_psnr( sqe[0], h->param.i_width * h->param.i_height );
1695         h->stat.f_psnr_mean_u[h->sh.i_type] += x264_psnr( sqe[1], h->param.i_width * h->param.i_height / 4 );
1696         h->stat.f_psnr_mean_v[h->sh.i_type] += x264_psnr( sqe[2], h->param.i_width * h->param.i_height / 4 );
1697
1698         snprintf( psz_message, 80, " PSNR Y:%5.2f U:%5.2f V:%5.2f",
1699                   x264_psnr( sqe[0], h->param.i_width * h->param.i_height ),
1700                   x264_psnr( sqe[1], h->param.i_width * h->param.i_height / 4),
1701                   x264_psnr( sqe[2], h->param.i_width * h->param.i_height / 4) );
1702     }
1703
1704     if( h->param.analyse.b_ssim )
1705     {
1706         // offset by 2 pixels to avoid alignment of ssim blocks with dct blocks
1707         float ssim_y = x264_pixel_ssim_wxh( &h->pixf,
1708                          h->fdec->plane[0] + 2+2*h->fdec->i_stride[0], h->fdec->i_stride[0],
1709                          h->fenc->plane[0] + 2+2*h->fenc->i_stride[0], h->fenc->i_stride[0],
1710                          h->param.i_width-2, h->param.i_height-2 );
1711         h->stat.f_ssim_mean_y[h->sh.i_type] += ssim_y;
1712         snprintf( psz_message + strlen(psz_message), 80 - strlen(psz_message),
1713                   " SSIM Y:%.5f", ssim_y );
1714     }
1715     psz_message[79] = '\0';
1716
1717     x264_log( h, X264_LOG_DEBUG,
1718                   "frame=%4d QP=%i NAL=%d Slice:%c Poc:%-3d I:%-4d P:%-4d SKIP:%-4d size=%d bytes%s\n",
1719               h->i_frame,
1720               h->fdec->i_qpplus1 - 1,
1721               h->i_nal_ref_idc,
1722               h->sh.i_type == SLICE_TYPE_I ? 'I' : (h->sh.i_type == SLICE_TYPE_P ? 'P' : 'B' ),
1723               h->fdec->i_poc,
1724               h->stat.frame.i_mb_count_i,
1725               h->stat.frame.i_mb_count_p,
1726               h->stat.frame.i_mb_count_skip,
1727               h->out.i_frame_size,
1728               psz_message );
1729
1730     // keep stats all in one place
1731     x264_thread_sync_stat( h->thread[0], h );
1732     // for the use of the next frame
1733     x264_thread_sync_stat( thread_current, h );
1734
1735 #ifdef DEBUG_MB_TYPE
1736 {
1737     static const char mb_chars[] = { 'i', 'i', 'I', 'C', 'P', '8', 'S',
1738         'D', '<', 'X', 'B', 'X', '>', 'B', 'B', 'B', 'B', '8', 'S' };
1739     int mb_xy;
1740     for( mb_xy = 0; mb_xy < h->sps->i_mb_width * h->sps->i_mb_height; mb_xy++ )
1741     {
1742         if( h->mb.type[mb_xy] < X264_MBTYPE_MAX && h->mb.type[mb_xy] >= 0 )
1743             fprintf( stderr, "%c ", mb_chars[ h->mb.type[mb_xy] ] );
1744         else
1745             fprintf( stderr, "? " );
1746
1747         if( (mb_xy+1) % h->sps->i_mb_width == 0 )
1748             fprintf( stderr, "\n" );
1749     }
1750 }
1751 #endif
1752
1753 #ifdef DEBUG_DUMP_FRAME
1754     /* Dump reconstructed frame */
1755     x264_frame_dump( h, h->fdec, "fdec.yuv" );
1756 #endif
1757 }
1758
1759 /****************************************************************************
1760  * x264_encoder_close:
1761  ****************************************************************************/
1762 void    x264_encoder_close  ( x264_t *h )
1763 {
1764 #ifdef DEBUG_BENCHMARK
1765     int64_t i_mtime_total = i_mtime_analyse + i_mtime_encode + i_mtime_write + i_mtime_filter + 1;
1766 #endif
1767     int64_t i_yuv_size = 3 * h->param.i_width * h->param.i_height / 2;
1768     int i;
1769
1770     for( i=0; i<h->param.i_threads; i++ )
1771     {
1772         // don't strictly have to wait for the other threads, but it's simpler than cancelling them
1773         if( h->thread[i]->b_thread_active )
1774             x264_pthread_join( h->thread[i]->thread_handle, NULL );
1775     }
1776
1777 #ifdef DEBUG_BENCHMARK
1778     x264_log( h, X264_LOG_INFO,
1779               "analyse=%d(%lldms) encode=%d(%lldms) write=%d(%lldms) filter=%d(%lldms)\n",
1780               (int)(100*i_mtime_analyse/i_mtime_total), i_mtime_analyse/1000,
1781               (int)(100*i_mtime_encode/i_mtime_total), i_mtime_encode/1000,
1782               (int)(100*i_mtime_write/i_mtime_total), i_mtime_write/1000,
1783               (int)(100*i_mtime_filter/i_mtime_total), i_mtime_filter/1000 );
1784 #endif
1785
1786     /* Slices used and PSNR */
1787     for( i=0; i<5; i++ )
1788     {
1789         static const int slice_order[] = { SLICE_TYPE_I, SLICE_TYPE_SI, SLICE_TYPE_P, SLICE_TYPE_SP, SLICE_TYPE_B };
1790         static const char *slice_name[] = { "P", "B", "I", "SP", "SI" };
1791         int i_slice = slice_order[i];
1792
1793         if( h->stat.i_slice_count[i_slice] > 0 )
1794         {
1795             const int i_count = h->stat.i_slice_count[i_slice];
1796             if( h->param.analyse.b_psnr )
1797             {
1798                 x264_log( h, X264_LOG_INFO,
1799                           "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",
1800                           slice_name[i_slice],
1801                           i_count,
1802                           (double)h->stat.i_slice_qp[i_slice] / i_count,
1803                           (double)h->stat.i_slice_size[i_slice] / i_count,
1804                           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,
1805                           h->stat.f_psnr_average[i_slice] / i_count,
1806                           x264_psnr( h->stat.i_sqe_global[i_slice], i_count * i_yuv_size ) );
1807             }
1808             else
1809             {
1810                 x264_log( h, X264_LOG_INFO,
1811                           "slice %s:%-5d Avg QP:%5.2f  size:%6.0f\n",
1812                           slice_name[i_slice],
1813                           i_count,
1814                           (double)h->stat.i_slice_qp[i_slice] / i_count,
1815                           (double)h->stat.i_slice_size[i_slice] / i_count );
1816             }
1817         }
1818     }
1819
1820     /* MB types used */
1821     if( h->stat.i_slice_count[SLICE_TYPE_I] > 0 )
1822     {
1823         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_I];
1824         const double i_count = h->stat.i_slice_count[SLICE_TYPE_I] * h->mb.i_mb_count / 100.0;
1825         x264_log( h, X264_LOG_INFO,
1826                   "mb I  I16..4: %4.1f%% %4.1f%% %4.1f%%\n",
1827                   i_mb_count[I_16x16]/ i_count,
1828                   i_mb_count[I_8x8]  / i_count,
1829                   i_mb_count[I_4x4]  / i_count );
1830     }
1831     if( h->stat.i_slice_count[SLICE_TYPE_P] > 0 )
1832     {
1833         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_P];
1834         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_P];
1835         const double i_count = h->stat.i_slice_count[SLICE_TYPE_P] * h->mb.i_mb_count / 100.0;
1836         x264_log( h, X264_LOG_INFO,
1837                   "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",
1838                   i_mb_count[I_16x16]/ i_count,
1839                   i_mb_count[I_8x8]  / i_count,
1840                   i_mb_count[I_4x4]  / i_count,
1841                   i_mb_size[PIXEL_16x16] / (i_count*4),
1842                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
1843                   i_mb_size[PIXEL_8x8] / (i_count*4),
1844                   (i_mb_size[PIXEL_8x4] + i_mb_size[PIXEL_4x8]) / (i_count*4),
1845                   i_mb_size[PIXEL_4x4] / (i_count*4),
1846                   i_mb_count[P_SKIP] / i_count );
1847     }
1848     if( h->stat.i_slice_count[SLICE_TYPE_B] > 0 )
1849     {
1850         const int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_B];
1851         const int64_t *i_mb_size = h->stat.i_mb_count_size[SLICE_TYPE_B];
1852         const double i_count = h->stat.i_slice_count[SLICE_TYPE_B] * h->mb.i_mb_count / 100.0;
1853         x264_log( h, X264_LOG_INFO,
1854                   "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",
1855                   i_mb_count[I_16x16]  / i_count,
1856                   i_mb_count[I_8x8]    / i_count,
1857                   i_mb_count[I_4x4]    / i_count,
1858                   i_mb_size[PIXEL_16x16] / (i_count*4),
1859                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
1860                   i_mb_size[PIXEL_8x8] / (i_count*4),
1861                   i_mb_count[B_DIRECT] / i_count,
1862                   i_mb_count[B_SKIP]   / i_count );
1863     }
1864
1865     x264_ratecontrol_summary( h );
1866
1867     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 )
1868     {
1869         const int i_count = h->stat.i_slice_count[SLICE_TYPE_I] +
1870                             h->stat.i_slice_count[SLICE_TYPE_P] +
1871                             h->stat.i_slice_count[SLICE_TYPE_B];
1872         float fps = (float) h->param.i_fps_num / h->param.i_fps_den;
1873 #define SUM3(p) (p[SLICE_TYPE_I] + p[SLICE_TYPE_P] + p[SLICE_TYPE_B])
1874 #define SUM3b(p,o) (p[SLICE_TYPE_I][o] + p[SLICE_TYPE_P][o] + p[SLICE_TYPE_B][o])
1875         float f_bitrate = fps * SUM3(h->stat.i_slice_size) / i_count / 125;
1876
1877         if( h->pps->b_transform_8x8_mode )
1878         {
1879             int64_t i_i8x8 = SUM3b( h->stat.i_mb_count, I_8x8 );
1880             int64_t i_intra = i_i8x8 + SUM3b( h->stat.i_mb_count, I_4x4 )
1881                                      + SUM3b( h->stat.i_mb_count, I_16x16 );
1882             x264_log( h, X264_LOG_INFO, "8x8 transform  intra:%.1f%%  inter:%.1f%%\n",
1883                       100. * i_i8x8 / i_intra,
1884                       100. * h->stat.i_mb_count_8x8dct[1] / h->stat.i_mb_count_8x8dct[0] );
1885         }
1886
1887         if( h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
1888             && h->stat.i_slice_count[SLICE_TYPE_B] )
1889         {
1890             x264_log( h, X264_LOG_INFO, "direct mvs  spatial:%.1f%%  temporal:%.1f%%\n",
1891                       h->stat.i_direct_frames[1] * 100. / h->stat.i_slice_count[SLICE_TYPE_B],
1892                       h->stat.i_direct_frames[0] * 100. / h->stat.i_slice_count[SLICE_TYPE_B] );
1893         }
1894
1895         if( h->frames.i_max_ref0 > 1 )
1896         {
1897             int i_slice;
1898             for( i_slice = 0; i_slice < 2; i_slice++ )
1899             {
1900                 char buf[200];
1901                 char *p = buf;
1902                 int64_t i_den = 0;
1903                 int i_max = 0;
1904                 for( i = 0; i < h->frames.i_max_ref0 << h->param.b_interlaced; i++ )
1905                     if( h->stat.i_mb_count_ref[i_slice][i] )
1906                     {
1907                         i_den += h->stat.i_mb_count_ref[i_slice][i];
1908                         i_max = i;
1909                     }
1910                 if( i_max == 0 )
1911                     continue;
1912                 for( i = 0; i <= i_max; i++ )
1913                     p += sprintf( p, " %4.1f%%", 100. * h->stat.i_mb_count_ref[i_slice][i] / i_den );
1914                 x264_log( h, X264_LOG_INFO, "ref %c %s\n", i_slice==SLICE_TYPE_P ? 'P' : 'B', buf );
1915             }
1916         }
1917
1918         if( h->param.analyse.b_ssim )
1919         {
1920             x264_log( h, X264_LOG_INFO,
1921                       "SSIM Mean Y:%.7f\n",
1922                       SUM3( h->stat.f_ssim_mean_y ) / i_count );
1923         }
1924         if( h->param.analyse.b_psnr )
1925         {
1926             x264_log( h, X264_LOG_INFO,
1927                       "PSNR Mean Y:%6.3f U:%6.3f V:%6.3f Avg:%6.3f Global:%6.3f kb/s:%.2f\n",
1928                       SUM3( h->stat.f_psnr_mean_y ) / i_count,
1929                       SUM3( h->stat.f_psnr_mean_u ) / i_count,
1930                       SUM3( h->stat.f_psnr_mean_v ) / i_count,
1931                       SUM3( h->stat.f_psnr_average ) / i_count,
1932                       x264_psnr( SUM3( h->stat.i_sqe_global ), i_count * i_yuv_size ),
1933                       f_bitrate );
1934         }
1935         else
1936             x264_log( h, X264_LOG_INFO, "kb/s:%.1f\n", f_bitrate );
1937     }
1938
1939     /* frames */
1940     for( i = 0; h->frames.current[i]; i++ )
1941         x264_frame_delete( h->frames.current[i] );
1942     for( i = 0; h->frames.next[i]; i++ )
1943         x264_frame_delete( h->frames.next[i] );
1944     for( i = 0; h->frames.unused[i]; i++ )
1945         x264_frame_delete( h->frames.unused[i] );
1946     for( i = 0; h->frames.reference[i]; i++ )
1947         x264_frame_delete( h->frames.reference[i] );
1948
1949     /* rc */
1950     x264_ratecontrol_delete( h );
1951
1952     /* param */
1953     if( h->param.rc.psz_stat_out )
1954         free( h->param.rc.psz_stat_out );
1955     if( h->param.rc.psz_stat_in )
1956         free( h->param.rc.psz_stat_in );
1957     if( h->param.rc.psz_rc_eq )
1958         free( h->param.rc.psz_rc_eq );
1959
1960     x264_cqm_delete( h );
1961     for( i = h->param.i_threads - 1; i >= 0; i-- )
1962     {
1963         x264_macroblock_cache_end( h->thread[i] );
1964         x264_free( h->thread[i]->out.p_bitstream );
1965         x264_free( h->thread[i] );
1966     }
1967 }