git.sesse.net Git - x264/blob - encoder/encoder.c

   1 /*****************************************************************************
   2  * x264: h264 encoder
   3  *****************************************************************************
   4  * Copyright (C) 2003-2008 x264 project
   5  *
   6  * Authors: Laurent Aimar <fenrir@via.ecp.fr>
   7  *          Loren Merritt <lorenm@u.washington.edu>
   8  *          Fiona Glaser <fiona@x264.com>
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published by
  12  * the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  * GNU General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * along with this program; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
  23  *****************************************************************************/
  24
  25 #include <math.h>
  26
  27 #include "common/common.h"
  28 #include "common/cpu.h"
  29
  30 #include "set.h"
  31 #include "analyse.h"
  32 #include "ratecontrol.h"
  33 #include "macroblock.h"
  34
  35 #if VISUALIZE
  36 #include "common/visualize.h"
  37 #endif
  38
  39 //#define DEBUG_MB_TYPE
  40
  41 #define NALU_OVERHEAD 5 // startcode + NAL type costs 5 bytes per frame
  42
  43 #define bs_write_ue bs_write_ue_big
  44
  45 static int x264_encoder_frame_end( x264_t *h, x264_t *thread_current,
  46                                    x264_nal_t **pp_nal, int *pi_nal,
  47                                    x264_picture_t *pic_out );
  48
  49 /****************************************************************************
  50  *
  51  ******************************* x264 libs **********************************
  52  *
  53  ****************************************************************************/
  54 static float x264_psnr( int64_t i_sqe, int64_t i_size )
  55 {
  56     double f_mse = (double)i_sqe / ((double)65025.0 * (double)i_size);
  57     if( f_mse <= 0.0000000001 ) /* Max 100dB */
  58         return 100;
  59
  60     return (float)(-10.0 * log( f_mse ) / log( 10.0 ));
  61 }
  62
  63 static void x264_frame_dump( x264_t *h )
  64 {
  65     FILE *f = fopen( h->param.psz_dump_yuv, "r+b" );
  66     int i, y;
  67     if( !f )
  68         return;
  69     /* Write the frame in display order */
  70     fseek( f, h->fdec->i_frame * h->param.i_height * h->param.i_width * 3/2, SEEK_SET );
  71     for( i = 0; i < h->fdec->i_plane; i++ )
  72         for( y = 0; y < h->param.i_height >> !!i; y++ )
  73             fwrite( &h->fdec->plane[i][y*h->fdec->i_stride[i]], 1, h->param.i_width >> !!i, f );
  74     fclose( f );
  75 }
  76
  77
  78 /* Fill "default" values */
  79 static void x264_slice_header_init( x264_t *h, x264_slice_header_t *sh,
  80                                     x264_sps_t *sps, x264_pps_t *pps,
  81                                     int i_idr_pic_id, int i_frame, int i_qp )
  82 {
  83     x264_param_t *param = &h->param;
  84     int i;
  85
  86     /* First we fill all field */
  87     sh->sps = sps;
  88     sh->pps = pps;
  89
  90     sh->i_first_mb  = 0;
  91     sh->i_last_mb   = h->mb.i_mb_count - 1;
  92     sh->i_pps_id    = pps->i_id;
  93
  94     sh->i_frame_num = i_frame;
  95
  96     sh->b_mbaff = h->param.b_interlaced;
  97     sh->b_field_pic = 0;    /* no field support for now */
  98     sh->b_bottom_field = 0; /* not yet used */
  99
 100     sh->i_idr_pic_id = i_idr_pic_id;
 101
 102     /* poc stuff, fixed later */
 103     sh->i_poc_lsb = 0;
 104     sh->i_delta_poc_bottom = 0;
 105     sh->i_delta_poc[0] = 0;
 106     sh->i_delta_poc[1] = 0;
 107
 108     sh->i_redundant_pic_cnt = 0;
 109
 110     if( !h->mb.b_direct_auto_read )
 111     {
 112         if( h->mb.b_direct_auto_write )
 113             sh->b_direct_spatial_mv_pred = ( h->stat.i_direct_score[1] > h->stat.i_direct_score[0] );
 114         else
 115             sh->b_direct_spatial_mv_pred = ( param->analyse.i_direct_mv_pred == X264_DIRECT_PRED_SPATIAL );
 116     }
 117     /* else b_direct_spatial_mv_pred was read from the 2pass statsfile */
 118
 119     sh->b_num_ref_idx_override = 0;
 120     sh->i_num_ref_idx_l0_active = 1;
 121     sh->i_num_ref_idx_l1_active = 1;
 122
 123     sh->b_ref_pic_list_reordering_l0 = h->b_ref_reorder[0];
 124     sh->b_ref_pic_list_reordering_l1 = h->b_ref_reorder[1];
 125
 126     /* If the ref list isn't in the default order, construct reordering header */
 127     /* List1 reordering isn't needed yet */
 128     if( sh->b_ref_pic_list_reordering_l0 )
 129     {
 130         int pred_frame_num = i_frame;
 131         for( i = 0; i < h->i_ref0; i++ )
 132         {
 133             int diff = h->fref0[i]->i_frame_num - pred_frame_num;
 134             if( diff == 0 )
 135                 x264_log( h, X264_LOG_ERROR, "diff frame num == 0\n" );
 136             sh->ref_pic_list_order[0][i].idc = ( diff > 0 );
 137             sh->ref_pic_list_order[0][i].arg = abs( diff ) - 1;
 138             pred_frame_num = h->fref0[i]->i_frame_num;
 139         }
 140     }
 141
 142     sh->i_cabac_init_idc = param->i_cabac_init_idc;
 143
 144     sh->i_qp = i_qp;
 145     sh->i_qp_delta = i_qp - pps->i_pic_init_qp;
 146     sh->b_sp_for_swidth = 0;
 147     sh->i_qs_delta = 0;
 148
 149     /* If effective qp <= 15, deblocking would have no effect anyway */
 150     if( param->b_deblocking_filter
 151         && ( h->mb.b_variable_qp
 152         || 15 < i_qp + 2 * X264_MIN(param->i_deblocking_filter_alphac0, param->i_deblocking_filter_beta) ) )
 153     {
 154         sh->i_disable_deblocking_filter_idc = 0;
 155     }
 156     else
 157     {
 158         sh->i_disable_deblocking_filter_idc = 1;
 159     }
 160     sh->i_alpha_c0_offset = param->i_deblocking_filter_alphac0 << 1;
 161     sh->i_beta_offset = param->i_deblocking_filter_beta << 1;
 162 }
 163
 164 static void x264_slice_header_write( bs_t *s, x264_slice_header_t *sh, int i_nal_ref_idc )
 165 {
 166     int i;
 167
 168     if( sh->b_mbaff )
 169     {
 170         assert( sh->i_first_mb % (2*sh->sps->i_mb_width) == 0 );
 171         bs_write_ue( s, sh->i_first_mb >> 1 );
 172     }
 173     else
 174         bs_write_ue( s, sh->i_first_mb );
 175
 176     bs_write_ue( s, sh->i_type + 5 );   /* same type things */
 177     bs_write_ue( s, sh->i_pps_id );
 178     bs_write( s, sh->sps->i_log2_max_frame_num, sh->i_frame_num );
 179
 180     if( !sh->sps->b_frame_mbs_only )
 181     {
 182         bs_write1( s, sh->b_field_pic );
 183         if( sh->b_field_pic )
 184             bs_write1( s, sh->b_bottom_field );
 185     }
 186
 187     if( sh->i_idr_pic_id >= 0 ) /* NAL IDR */
 188     {
 189         bs_write_ue( s, sh->i_idr_pic_id );
 190     }
 191
 192     if( sh->sps->i_poc_type == 0 )
 193     {
 194         bs_write( s, sh->sps->i_log2_max_poc_lsb, sh->i_poc_lsb );
 195         if( sh->pps->b_pic_order && !sh->b_field_pic )
 196         {
 197             bs_write_se( s, sh->i_delta_poc_bottom );
 198         }
 199     }
 200     else if( sh->sps->i_poc_type == 1 && !sh->sps->b_delta_pic_order_always_zero )
 201     {
 202         bs_write_se( s, sh->i_delta_poc[0] );
 203         if( sh->pps->b_pic_order && !sh->b_field_pic )
 204         {
 205             bs_write_se( s, sh->i_delta_poc[1] );
 206         }
 207     }
 208
 209     if( sh->pps->b_redundant_pic_cnt )
 210     {
 211         bs_write_ue( s, sh->i_redundant_pic_cnt );
 212     }
 213
 214     if( sh->i_type == SLICE_TYPE_B )
 215     {
 216         bs_write1( s, sh->b_direct_spatial_mv_pred );
 217     }
 218     if( sh->i_type == SLICE_TYPE_P || sh->i_type == SLICE_TYPE_SP || sh->i_type == SLICE_TYPE_B )
 219     {
 220         bs_write1( s, sh->b_num_ref_idx_override );
 221         if( sh->b_num_ref_idx_override )
 222         {
 223             bs_write_ue( s, sh->i_num_ref_idx_l0_active - 1 );
 224             if( sh->i_type == SLICE_TYPE_B )
 225             {
 226                 bs_write_ue( s, sh->i_num_ref_idx_l1_active - 1 );
 227             }
 228         }
 229     }
 230
 231     /* ref pic list reordering */
 232     if( sh->i_type != SLICE_TYPE_I )
 233     {
 234         bs_write1( s, sh->b_ref_pic_list_reordering_l0 );
 235         if( sh->b_ref_pic_list_reordering_l0 )
 236         {
 237             for( i = 0; i < sh->i_num_ref_idx_l0_active; i++ )
 238             {
 239                 bs_write_ue( s, sh->ref_pic_list_order[0][i].idc );
 240                 bs_write_ue( s, sh->ref_pic_list_order[0][i].arg );
 241
 242             }
 243             bs_write_ue( s, 3 );
 244         }
 245     }
 246     if( sh->i_type == SLICE_TYPE_B )
 247     {
 248         bs_write1( s, sh->b_ref_pic_list_reordering_l1 );
 249         if( sh->b_ref_pic_list_reordering_l1 )
 250         {
 251             for( i = 0; i < sh->i_num_ref_idx_l1_active; i++ )
 252             {
 253                 bs_write_ue( s, sh->ref_pic_list_order[1][i].idc );
 254                 bs_write_ue( s, sh->ref_pic_list_order[1][i].arg );
 255             }
 256             bs_write_ue( s, 3 );
 257         }
 258     }
 259
 260     if( ( sh->pps->b_weighted_pred && ( sh->i_type == SLICE_TYPE_P || sh->i_type == SLICE_TYPE_SP ) ) ||
 261         ( sh->pps->b_weighted_bipred == 1 && sh->i_type == SLICE_TYPE_B ) )
 262     {
 263         /* FIXME */
 264     }
 265
 266     if( i_nal_ref_idc != 0 )
 267     {
 268         if( sh->i_idr_pic_id >= 0 )
 269         {
 270             bs_write1( s, 0 );  /* no output of prior pics flag */
 271             bs_write1( s, 0 );  /* long term reference flag */
 272         }
 273         else
 274         {
 275             bs_write1( s, 0 );  /* adaptive_ref_pic_marking_mode_flag */
 276         }
 277     }
 278
 279     if( sh->pps->b_cabac && sh->i_type != SLICE_TYPE_I )
 280     {
 281         bs_write_ue( s, sh->i_cabac_init_idc );
 282     }
 283     bs_write_se( s, sh->i_qp_delta );      /* slice qp delta */
 284
 285     if( sh->pps->b_deblocking_filter_control )
 286     {
 287         bs_write_ue( s, sh->i_disable_deblocking_filter_idc );
 288         if( sh->i_disable_deblocking_filter_idc != 1 )
 289         {
 290             bs_write_se( s, sh->i_alpha_c0_offset >> 1 );
 291             bs_write_se( s, sh->i_beta_offset >> 1 );
 292         }
 293     }
 294 }
 295
 296 /* If we are within a reasonable distance of the end of the memory allocated for the bitstream, */
 297 /* reallocate, adding an arbitrary amount of space (100 kilobytes). */
 298 static int x264_bitstream_check_buffer( x264_t *h )
 299 {
 300     uint8_t *bs_bak = h->out.p_bitstream;
 301     if( ( h->param.b_cabac && (h->cabac.p_end - h->cabac.p < 2500) )
 302      || ( h->out.bs.p_end - h->out.bs.p < 2500 ) )
 303     {
 304         intptr_t delta;
 305         int i;
 306
 307         h->out.i_bitstream += 100000;
 308         CHECKED_MALLOC( h->out.p_bitstream, h->out.i_bitstream );
 309         h->mc.memcpy_aligned( h->out.p_bitstream, bs_bak, (h->out.i_bitstream - 100000) & ~15 );
 310         delta = h->out.p_bitstream - bs_bak;
 311
 312         h->out.bs.p_start += delta;
 313         h->out.bs.p += delta;
 314         h->out.bs.p_end = h->out.p_bitstream + h->out.i_bitstream;
 315
 316         h->cabac.p_start += delta;
 317         h->cabac.p += delta;
 318         h->cabac.p_end = h->out.p_bitstream + h->out.i_bitstream;
 319
 320         for( i = 0; i <= h->out.i_nal; i++ )
 321             h->out.nal[i].p_payload += delta;
 322         x264_free( bs_bak );
 323     }
 324     return 0;
 325 fail:
 326     x264_free( bs_bak );
 327     return -1;
 328 }
 329
 330 /****************************************************************************
 331  *
 332  ****************************************************************************
 333  ****************************** External API*********************************
 334  ****************************************************************************
 335  *
 336  ****************************************************************************/
 337
 338 static int x264_validate_parameters( x264_t *h )
 339 {
 340 #ifdef HAVE_MMX
 341     if( !(x264_cpu_detect() & X264_CPU_SSE) )
 342     {
 343         x264_log( h, X264_LOG_ERROR, "your cpu does not support SSE1, but x264 was compiled with asm support\n");
 344         x264_log( h, X264_LOG_ERROR, "to run x264, recompile without asm support (configure --disable-asm)\n");
 345         return -1;
 346     }
 347 #endif
 348     if( h->param.i_width <= 0 || h->param.i_height <= 0 )
 349     {
 350         x264_log( h, X264_LOG_ERROR, "invalid width x height (%dx%d)\n",
 351                   h->param.i_width, h->param.i_height );
 352         return -1;
 353     }
 354
 355     if( h->param.i_width % 2 || h->param.i_height % 2 )
 356     {
 357         x264_log( h, X264_LOG_ERROR, "width or height not divisible by 2 (%dx%d)\n",
 358                   h->param.i_width, h->param.i_height );
 359         return -1;
 360     }
 361     if( h->param.i_csp != X264_CSP_I420 )
 362     {
 363         x264_log( h, X264_LOG_ERROR, "invalid CSP (only I420 supported)\n" );
 364         return -1;
 365     }
 366
 367     if( h->param.i_threads == X264_THREADS_AUTO )
 368         h->param.i_threads = x264_cpu_num_processors() * 3/2;
 369     h->param.i_threads = x264_clip3( h->param.i_threads, 1, X264_THREAD_MAX );
 370     if( h->param.i_threads > 1 )
 371     {
 372 #ifndef HAVE_PTHREAD
 373         x264_log( h, X264_LOG_WARNING, "not compiled with pthread support!\n");
 374         h->param.i_threads = 1;
 375 #endif
 376     }
 377
 378     if( h->param.b_interlaced )
 379     {
 380         if( h->param.analyse.i_me_method >= X264_ME_ESA )
 381         {
 382             x264_log( h, X264_LOG_WARNING, "interlace + me=esa is not implemented\n" );
 383             h->param.analyse.i_me_method = X264_ME_UMH;
 384         }
 385         if( h->param.analyse.i_direct_mv_pred > X264_DIRECT_PRED_SPATIAL )
 386         {
 387             x264_log( h, X264_LOG_WARNING, "interlace + direct=temporal is not implemented\n" );
 388             h->param.analyse.i_direct_mv_pred = X264_DIRECT_PRED_SPATIAL;
 389         }
 390     }
 391
 392     /* Detect default ffmpeg settings and terminate with an error. */
 393     {
 394         int score = 0;
 395         score += h->param.analyse.i_me_range == 0;
 396         score += h->param.rc.i_qp_step == 3;
 397         score += h->param.i_keyint_max == 12;
 398         score += h->param.rc.i_qp_min == 2;
 399         score += h->param.rc.i_qp_max == 31;
 400         score += h->param.rc.f_qcompress == 0.5;
 401         score += fabs(h->param.rc.f_ip_factor - 1.25) < 0.01;
 402         score += fabs(h->param.rc.f_pb_factor - 1.25) < 0.01;
 403         score += h->param.analyse.inter == 0 && h->param.analyse.i_subpel_refine == 8;
 404         if( score >= 5 )
 405         {
 406             x264_log( h, X264_LOG_ERROR, "broken ffmpeg default settings detected\n" );
 407             x264_log( h, X264_LOG_ERROR, "use an encoding preset (vpre)\n" );
 408             return -1;
 409         }
 410     }
 411
 412     if( h->param.rc.i_rc_method < 0 || h->param.rc.i_rc_method > 2 )
 413     {
 414         x264_log( h, X264_LOG_ERROR, "no ratecontrol method specified\n" );
 415         return -1;
 416     }
 417     h->param.rc.f_rf_constant = x264_clip3f( h->param.rc.f_rf_constant, 0, 51 );
 418     h->param.rc.i_qp_constant = x264_clip3( h->param.rc.i_qp_constant, 0, 51 );
 419     if( h->param.rc.i_rc_method == X264_RC_CRF )
 420     {
 421         h->param.rc.i_qp_constant = h->param.rc.f_rf_constant;
 422         h->param.rc.i_bitrate = 0;
 423     }
 424     if( (h->param.rc.i_rc_method == X264_RC_CQP || h->param.rc.i_rc_method == X264_RC_CRF)
 425         && h->param.rc.i_qp_constant == 0 )
 426     {
 427         h->mb.b_lossless = 1;
 428         h->param.i_cqm_preset = X264_CQM_FLAT;
 429         h->param.psz_cqm_file = NULL;
 430         h->param.rc.i_rc_method = X264_RC_CQP;
 431         h->param.rc.f_ip_factor = 1;
 432         h->param.rc.f_pb_factor = 1;
 433         h->param.analyse.b_psnr = 0;
 434         h->param.analyse.b_ssim = 0;
 435         h->param.analyse.i_chroma_qp_offset = 0;
 436         h->param.analyse.i_trellis = 0;
 437         h->param.analyse.b_fast_pskip = 0;
 438         h->param.analyse.i_noise_reduction = 0;
 439         h->param.analyse.f_psy_rd = 0;
 440         h->param.i_bframe = 0;
 441         /* 8x8dct is not useful at all in CAVLC lossless */
 442         if( !h->param.b_cabac )
 443             h->param.analyse.b_transform_8x8 = 0;
 444     }
 445     if( h->param.rc.i_rc_method == X264_RC_CQP )
 446     {
 447         float qp_p = h->param.rc.i_qp_constant;
 448         float qp_i = qp_p - 6*log(h->param.rc.f_ip_factor)/log(2);
 449         float qp_b = qp_p + 6*log(h->param.rc.f_pb_factor)/log(2);
 450         h->param.rc.i_qp_min = x264_clip3( (int)(X264_MIN3( qp_p, qp_i, qp_b )), 0, 51 );
 451         h->param.rc.i_qp_max = x264_clip3( (int)(X264_MAX3( qp_p, qp_i, qp_b ) + .999), 0, 51 );
 452         h->param.rc.i_aq_mode = 0;
 453         h->param.rc.b_mb_tree = 0;
 454     }
 455     h->param.rc.i_qp_max = x264_clip3( h->param.rc.i_qp_max, 0, 51 );
 456     h->param.rc.i_qp_min = x264_clip3( h->param.rc.i_qp_min, 0, h->param.rc.i_qp_max );
 457
 458     if( ( h->param.i_width % 16 || h->param.i_height % 16 )
 459         && h->param.i_height != 1080 && !h->mb.b_lossless )
 460     {
 461         // There's nothing special about 1080 in that the warning still applies to it,
 462         // but chances are the user can't help it if his content is already 1080p,
 463         // so there's no point in warning in that case.
 464         x264_log( h, X264_LOG_WARNING,
 465                   "width or height not divisible by 16 (%dx%d), compression will suffer.\n",
 466                   h->param.i_width, h->param.i_height );
 467     }
 468
 469     int max_slices = (h->param.i_height+((16<<h->param.b_interlaced)-1))/(16<<h->param.b_interlaced);
 470     h->param.i_slice_count = x264_clip3( h->param.i_slice_count, 0, max_slices );
 471     h->param.i_slice_max_size = X264_MAX( h->param.i_slice_max_size, 0 );
 472     h->param.i_slice_max_mbs = X264_MAX( h->param.i_slice_max_mbs, 0 );
 473     if( h->param.b_interlaced && h->param.i_slice_max_size )
 474     {
 475         x264_log( h, X264_LOG_WARNING, "interlaced + slice-max-size is not implemented\n" );
 476         h->param.i_slice_max_size = 0;
 477     }
 478     if( h->param.b_interlaced && h->param.i_slice_max_mbs )
 479     {
 480         x264_log( h, X264_LOG_WARNING, "interlaced + slice-max-mbs is not implemented\n" );
 481         h->param.i_slice_max_mbs = 0;
 482     }
 483     if( h->param.i_slice_max_mbs || h->param.i_slice_max_size )
 484         h->param.i_slice_count = 0;
 485
 486     h->param.i_frame_reference = x264_clip3( h->param.i_frame_reference, 1, 16 );
 487     if( h->param.i_keyint_max <= 0 )
 488         h->param.i_keyint_max = 1;
 489     if( h->param.i_scenecut_threshold < 0 )
 490         h->param.i_scenecut_threshold = 0;
 491     h->param.i_keyint_min = x264_clip3( h->param.i_keyint_min, 1, h->param.i_keyint_max/2+1 );
 492     if( !h->param.analyse.i_subpel_refine && h->param.analyse.i_direct_mv_pred > X264_DIRECT_PRED_SPATIAL )
 493     {
 494         x264_log( h, X264_LOG_WARNING, "subme=0 + direct=temporal is not supported\n" );
 495         h->param.analyse.i_direct_mv_pred = X264_DIRECT_PRED_SPATIAL;
 496     }
 497     h->param.i_bframe = x264_clip3( h->param.i_bframe, 0, X264_BFRAME_MAX );
 498     if( h->param.i_keyint_max == 1 )
 499         h->param.i_bframe = 0;
 500     h->param.i_bframe_bias = x264_clip3( h->param.i_bframe_bias, -90, 100 );
 501     h->param.b_bframe_pyramid = h->param.b_bframe_pyramid && h->param.i_bframe > 1;
 502     if( !h->param.i_bframe )
 503     {
 504         h->param.i_bframe_adaptive = X264_B_ADAPT_NONE;
 505         h->param.analyse.i_direct_mv_pred = 0;
 506         h->param.analyse.b_weighted_bipred = 0;
 507     }
 508     h->param.rc.i_lookahead = x264_clip3( h->param.rc.i_lookahead, 0, X264_LOOKAHEAD_MAX );
 509     {
 510         int maxrate = X264_MAX( h->param.rc.i_vbv_max_bitrate, h->param.rc.i_bitrate );
 511         float bufsize = maxrate ? (float)h->param.rc.i_vbv_buffer_size / maxrate : 0;
 512         float fps = h->param.i_fps_num > 0 && h->param.i_fps_den > 0 ? (float) h->param.i_fps_num / h->param.i_fps_den : 25.0;
 513         h->param.rc.i_lookahead = X264_MIN( h->param.rc.i_lookahead, X264_MAX( h->param.i_keyint_max, bufsize*fps ) );
 514     }
 515
 516     h->param.rc.f_qcompress = x264_clip3f( h->param.rc.f_qcompress, 0.0, 1.0 );
 517     if( !h->param.rc.i_lookahead || h->param.i_keyint_max == 1 || h->param.rc.f_qcompress == 1 )
 518         h->param.rc.b_mb_tree = 0;
 519     if( h->param.rc.b_stat_read )
 520         h->param.rc.i_lookahead = 0;
 521 #ifdef HAVE_PTHREAD
 522     if( h->param.i_sync_lookahead )
 523         h->param.i_sync_lookahead = x264_clip3( h->param.i_sync_lookahead, h->param.i_threads + h->param.i_bframe, X264_LOOKAHEAD_MAX );
 524     if( h->param.rc.b_stat_read || h->param.i_threads == 1 )
 525         h->param.i_sync_lookahead = 0;
 526 #else
 527     h->param.i_sync_lookahead = 0;
 528 #endif
 529
 530     h->mb.b_direct_auto_write = h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
 531                                 && h->param.i_bframe
 532                                 && ( h->param.rc.b_stat_write || !h->param.rc.b_stat_read );
 533
 534     h->param.i_deblocking_filter_alphac0 = x264_clip3( h->param.i_deblocking_filter_alphac0, -6, 6 );
 535     h->param.i_deblocking_filter_beta    = x264_clip3( h->param.i_deblocking_filter_beta, -6, 6 );
 536     h->param.analyse.i_luma_deadzone[0] = x264_clip3( h->param.analyse.i_luma_deadzone[0], 0, 32 );
 537     h->param.analyse.i_luma_deadzone[1] = x264_clip3( h->param.analyse.i_luma_deadzone[1], 0, 32 );
 538
 539     h->param.i_cabac_init_idc = x264_clip3( h->param.i_cabac_init_idc, 0, 2 );
 540
 541     if( h->param.i_cqm_preset < X264_CQM_FLAT || h->param.i_cqm_preset > X264_CQM_CUSTOM )
 542         h->param.i_cqm_preset = X264_CQM_FLAT;
 543
 544     if( h->param.analyse.i_me_method < X264_ME_DIA ||
 545         h->param.analyse.i_me_method > X264_ME_TESA )
 546         h->param.analyse.i_me_method = X264_ME_HEX;
 547     if( h->param.analyse.i_me_range < 4 )
 548         h->param.analyse.i_me_range = 4;
 549     if( h->param.analyse.i_me_range > 16 && h->param.analyse.i_me_method <= X264_ME_HEX )
 550         h->param.analyse.i_me_range = 16;
 551     if( h->param.analyse.i_me_method == X264_ME_TESA &&
 552         (h->mb.b_lossless || h->param.analyse.i_subpel_refine <= 1) )
 553         h->param.analyse.i_me_method = X264_ME_ESA;
 554     h->param.analyse.i_subpel_refine = x264_clip3( h->param.analyse.i_subpel_refine, 0, 10 );
 555     h->param.analyse.b_mixed_references = h->param.analyse.b_mixed_references && h->param.i_frame_reference > 1;
 556     h->param.analyse.inter &= X264_ANALYSE_PSUB16x16|X264_ANALYSE_PSUB8x8|X264_ANALYSE_BSUB16x16|
 557                               X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 558     h->param.analyse.intra &= X264_ANALYSE_I4x4|X264_ANALYSE_I8x8;
 559     if( !(h->param.analyse.inter & X264_ANALYSE_PSUB16x16) )
 560         h->param.analyse.inter &= ~X264_ANALYSE_PSUB8x8;
 561     if( !h->param.analyse.b_transform_8x8 )
 562     {
 563         h->param.analyse.inter &= ~X264_ANALYSE_I8x8;
 564         h->param.analyse.intra &= ~X264_ANALYSE_I8x8;
 565     }
 566     h->param.analyse.i_chroma_qp_offset = x264_clip3(h->param.analyse.i_chroma_qp_offset, -12, 12);
 567     if( !h->param.b_cabac )
 568         h->param.analyse.i_trellis = 0;
 569     h->param.analyse.i_trellis = x264_clip3( h->param.analyse.i_trellis, 0, 2 );
 570     if( !h->param.analyse.b_psy )
 571     {
 572         h->param.analyse.f_psy_rd = 0;
 573         h->param.analyse.f_psy_trellis = 0;
 574     }
 575     if( !h->param.analyse.i_trellis )
 576         h->param.analyse.f_psy_trellis = 0;
 577     h->param.analyse.f_psy_rd = x264_clip3f( h->param.analyse.f_psy_rd, 0, 10 );
 578     h->param.analyse.f_psy_trellis = x264_clip3f( h->param.analyse.f_psy_trellis, 0, 10 );
 579     if( h->param.analyse.i_subpel_refine < 6 )
 580         h->param.analyse.f_psy_rd = 0;
 581     h->mb.i_psy_rd = FIX8( h->param.analyse.f_psy_rd );
 582     /* Psy RDO increases overall quantizers to improve the quality of luma--this indirectly hurts chroma quality */
 583     /* so we lower the chroma QP offset to compensate */
 584     /* This can be triggered repeatedly on multiple calls to parameter_validate, but since encoding
 585      * uses the pps chroma qp offset not the param chroma qp offset, this is not a problem. */
 586     if( h->mb.i_psy_rd )
 587         h->param.analyse.i_chroma_qp_offset -= h->param.analyse.f_psy_rd < 0.25 ? 1 : 2;
 588     h->mb.i_psy_trellis = FIX8( h->param.analyse.f_psy_trellis / 4 );
 589     /* Psy trellis has a similar effect. */
 590     if( h->mb.i_psy_trellis )
 591         h->param.analyse.i_chroma_qp_offset -= h->param.analyse.f_psy_trellis < 0.25 ? 1 : 2;
 592     else
 593         h->mb.i_psy_trellis = 0;
 594     h->param.analyse.i_chroma_qp_offset = x264_clip3(h->param.analyse.i_chroma_qp_offset, -12, 12);
 595     h->param.rc.i_aq_mode = x264_clip3( h->param.rc.i_aq_mode, 0, 2 );
 596     h->param.rc.f_aq_strength = x264_clip3f( h->param.rc.f_aq_strength, 0, 3 );
 597     if( h->param.rc.f_aq_strength == 0 )
 598         h->param.rc.i_aq_mode = 0;
 599     /* MB-tree requires AQ to be on, even if the strength is zero. */
 600     if( !h->param.rc.i_aq_mode && h->param.rc.b_mb_tree )
 601     {
 602         h->param.rc.i_aq_mode = 1;
 603         h->param.rc.f_aq_strength = 0;
 604     }
 605     if( h->param.rc.b_mb_tree && h->param.b_bframe_pyramid )
 606     {
 607         x264_log( h, X264_LOG_WARNING, "b-pyramid + mb-tree is not supported\n" );
 608         h->param.b_bframe_pyramid = 0;
 609     }
 610     h->param.analyse.i_noise_reduction = x264_clip3( h->param.analyse.i_noise_reduction, 0, 1<<16 );
 611     if( h->param.analyse.i_subpel_refine == 10 && (h->param.analyse.i_trellis != 2 || !h->param.rc.i_aq_mode) )
 612         h->param.analyse.i_subpel_refine = 9;
 613
 614     {
 615         const x264_level_t *l = x264_levels;
 616         if( h->param.i_level_idc < 0 )
 617         {
 618             int maxrate_bak = h->param.rc.i_vbv_max_bitrate;
 619             if( h->param.rc.i_rc_method == X264_RC_ABR && h->param.rc.i_vbv_buffer_size <= 0 )
 620                 h->param.rc.i_vbv_max_bitrate = h->param.rc.i_bitrate * 2;
 621             h->sps = h->sps_array;
 622             x264_sps_init( h->sps, h->param.i_sps_id, &h->param );
 623             do h->param.i_level_idc = l->level_idc;
 624                 while( l[1].level_idc && x264_validate_levels( h, 0 ) && l++ );
 625             h->param.rc.i_vbv_max_bitrate = maxrate_bak;
 626         }
 627         else
 628         {
 629             while( l->level_idc && l->level_idc != h->param.i_level_idc )
 630                 l++;
 631             if( l->level_idc == 0 )
 632             {
 633                 x264_log( h, X264_LOG_ERROR, "invalid level_idc: %d\n", h->param.i_level_idc );
 634                 return -1;
 635             }
 636         }
 637         if( h->param.analyse.i_mv_range <= 0 )
 638             h->param.analyse.i_mv_range = l->mv_range >> h->param.b_interlaced;
 639         else
 640             h->param.analyse.i_mv_range = x264_clip3(h->param.analyse.i_mv_range, 32, 512 >> h->param.b_interlaced);
 641     }
 642
 643     if( h->param.i_threads > 1 )
 644     {
 645         int r = h->param.analyse.i_mv_range_thread;
 646         int r2;
 647         if( r <= 0 )
 648         {
 649             // half of the available space is reserved and divided evenly among the threads,
 650             // the rest is allocated to whichever thread is far enough ahead to use it.
 651             // reserving more space increases quality for some videos, but costs more time
 652             // in thread synchronization.
 653             int max_range = (h->param.i_height + X264_THREAD_HEIGHT) / h->param.i_threads - X264_THREAD_HEIGHT;
 654             r = max_range / 2;
 655         }
 656         r = X264_MAX( r, h->param.analyse.i_me_range );
 657         r = X264_MIN( r, h->param.analyse.i_mv_range );
 658         // round up to use the whole mb row
 659         r2 = (r & ~15) + ((-X264_THREAD_HEIGHT) & 15);
 660         if( r2 < r )
 661             r2 += 16;
 662         x264_log( h, X264_LOG_DEBUG, "using mv_range_thread = %d\n", r2 );
 663         h->param.analyse.i_mv_range_thread = r2;
 664     }
 665
 666     if( h->param.rc.f_qblur < 0 )
 667         h->param.rc.f_qblur = 0;
 668     if( h->param.rc.f_complexity_blur < 0 )
 669         h->param.rc.f_complexity_blur = 0;
 670
 671     h->param.i_sps_id &= 31;
 672
 673     if( h->param.i_log_level < X264_LOG_INFO )
 674     {
 675         h->param.analyse.b_psnr = 0;
 676         h->param.analyse.b_ssim = 0;
 677     }
 678
 679     /* ensure the booleans are 0 or 1 so they can be used in math */
 680 #define BOOLIFY(x) h->param.x = !!h->param.x
 681     BOOLIFY( b_cabac );
 682     BOOLIFY( b_deblocking_filter );
 683     BOOLIFY( b_interlaced );
 684     BOOLIFY( analyse.b_transform_8x8 );
 685     BOOLIFY( analyse.b_chroma_me );
 686     BOOLIFY( analyse.b_fast_pskip );
 687     BOOLIFY( rc.b_stat_write );
 688     BOOLIFY( rc.b_stat_read );
 689     BOOLIFY( rc.b_mb_tree );
 690 #undef BOOLIFY
 691
 692     return 0;
 693 }
 694
 695 static void mbcmp_init( x264_t *h )
 696 {
 697     int satd = !h->mb.b_lossless && h->param.analyse.i_subpel_refine > 1;
 698     memcpy( h->pixf.mbcmp, satd ? h->pixf.satd : h->pixf.sad_aligned, sizeof(h->pixf.mbcmp) );
 699     memcpy( h->pixf.mbcmp_unaligned, satd ? h->pixf.satd : h->pixf.sad, sizeof(h->pixf.mbcmp_unaligned) );
 700     h->pixf.intra_mbcmp_x3_16x16 = satd ? h->pixf.intra_satd_x3_16x16 : h->pixf.intra_sad_x3_16x16;
 701     h->pixf.intra_mbcmp_x3_8x8c = satd ? h->pixf.intra_satd_x3_8x8c : h->pixf.intra_sad_x3_8x8c;
 702     h->pixf.intra_mbcmp_x3_4x4 = satd ? h->pixf.intra_satd_x3_4x4 : h->pixf.intra_sad_x3_4x4;
 703     satd &= h->param.analyse.i_me_method == X264_ME_TESA;
 704     memcpy( h->pixf.fpelcmp, satd ? h->pixf.satd : h->pixf.sad, sizeof(h->pixf.fpelcmp) );
 705     memcpy( h->pixf.fpelcmp_x3, satd ? h->pixf.satd_x3 : h->pixf.sad_x3, sizeof(h->pixf.fpelcmp_x3) );
 706     memcpy( h->pixf.fpelcmp_x4, satd ? h->pixf.satd_x4 : h->pixf.sad_x4, sizeof(h->pixf.fpelcmp_x4) );
 707 }
 708
 709 static void x264_set_aspect_ratio( x264_t *h, x264_param_t *param, int initial )
 710 {
 711     /* VUI */
 712     if( param->vui.i_sar_width > 0 && param->vui.i_sar_height > 0 )
 713     {
 714         int i_w = param->vui.i_sar_width;
 715         int i_h = param->vui.i_sar_height;
 716         int old_w = h->param.vui.i_sar_width;
 717         int old_h = h->param.vui.i_sar_height;
 718
 719         x264_reduce_fraction( &i_w, &i_h );
 720
 721         while( i_w > 65535 || i_h > 65535 )
 722         {
 723             i_w /= 2;
 724             i_h /= 2;
 725         }
 726
 727         if( i_w != old_w || i_h != old_h || initial )
 728         {
 729             h->param.vui.i_sar_width = 0;
 730             h->param.vui.i_sar_height = 0;
 731             if( i_w == 0 || i_h == 0 )
 732                 x264_log( h, X264_LOG_WARNING, "cannot create valid sample aspect ratio\n" );
 733             else
 734             {
 735                 x264_log( h, initial?X264_LOG_INFO:X264_LOG_DEBUG, "using SAR=%d/%d\n", i_w, i_h );
 736                 h->param.vui.i_sar_width = i_w;
 737                 h->param.vui.i_sar_height = i_h;
 738             }
 739         }
 740     }
 741 }
 742
 743 /****************************************************************************
 744  * x264_encoder_open:
 745  ****************************************************************************/
 746 x264_t *x264_encoder_open( x264_param_t *param )
 747 {
 748     x264_t *h;
 749     char buf[1000], *p;
 750     int i, qp, i_slicetype_length;
 751
 752     CHECKED_MALLOCZERO( h, sizeof(x264_t) );
 753
 754     /* Create a copy of param */
 755     memcpy( &h->param, param, sizeof(x264_param_t) );
 756
 757     if( param->param_free )
 758         param->param_free( param );
 759
 760     if( x264_validate_parameters( h ) < 0 )
 761         goto fail;
 762
 763     if( h->param.psz_cqm_file )
 764         if( x264_cqm_parse_file( h, h->param.psz_cqm_file ) < 0 )
 765             goto fail;
 766
 767     if( h->param.rc.psz_stat_out )
 768         h->param.rc.psz_stat_out = strdup( h->param.rc.psz_stat_out );
 769     if( h->param.rc.psz_stat_in )
 770         h->param.rc.psz_stat_in = strdup( h->param.rc.psz_stat_in );
 771
 772     x264_set_aspect_ratio( h, param, 1 );
 773
 774     x264_reduce_fraction( &h->param.i_fps_num, &h->param.i_fps_den );
 775
 776     /* Init x264_t */
 777     h->i_frame = 0;
 778     h->i_frame_num = 0;
 779     h->i_idr_pic_id = 0;
 780
 781     h->sps = &h->sps_array[0];
 782     x264_sps_init( h->sps, h->param.i_sps_id, &h->param );
 783
 784     h->pps = &h->pps_array[0];
 785     x264_pps_init( h->pps, h->param.i_sps_id, &h->param, h->sps );
 786
 787     x264_validate_levels( h, 1 );
 788
 789     h->chroma_qp_table = i_chroma_qp_table + 12 + h->pps->i_chroma_qp_index_offset;
 790
 791     if( x264_cqm_init( h ) < 0 )
 792         goto fail;
 793
 794     h->mb.i_mb_count = h->sps->i_mb_width * h->sps->i_mb_height;
 795
 796     /* Init frames. */
 797     if( h->param.i_bframe_adaptive == X264_B_ADAPT_TRELLIS )
 798         h->frames.i_delay = X264_MAX(h->param.i_bframe,3)*4;
 799     else
 800         h->frames.i_delay = h->param.i_bframe;
 801     if( h->param.rc.b_mb_tree || h->param.rc.i_vbv_buffer_size )
 802         h->frames.i_delay = X264_MAX( h->frames.i_delay, h->param.rc.i_lookahead );
 803     i_slicetype_length = h->frames.i_delay;
 804     h->frames.i_delay += h->param.i_threads - 1;
 805     h->frames.i_delay = X264_MIN( h->frames.i_delay, X264_LOOKAHEAD_MAX );
 806     h->frames.i_delay += h->param.i_sync_lookahead;
 807
 808     h->frames.i_max_ref0 = h->param.i_frame_reference;
 809     h->frames.i_max_ref1 = h->sps->vui.i_num_reorder_frames;
 810     h->frames.i_max_dpb  = h->sps->vui.i_max_dec_frame_buffering;
 811     h->frames.b_have_lowres = !h->param.rc.b_stat_read
 812         && ( h->param.rc.i_rc_method == X264_RC_ABR
 813           || h->param.rc.i_rc_method == X264_RC_CRF
 814           || h->param.i_bframe_adaptive
 815           || h->param.i_scenecut_threshold
 816           || h->param.rc.b_mb_tree );
 817     h->frames.b_have_lowres |= h->param.rc.b_stat_read && h->param.rc.i_vbv_buffer_size > 0;
 818     h->frames.b_have_sub8x8_esa = !!(h->param.analyse.inter & X264_ANALYSE_PSUB8x8);
 819
 820     h->frames.i_last_idr = - h->param.i_keyint_max;
 821     h->frames.i_input    = 0;
 822
 823     CHECKED_MALLOCZERO( h->frames.unused[0], (h->frames.i_delay + 3) * sizeof(x264_frame_t *) );
 824     /* Allocate room for max refs plus a few extra just in case. */
 825     CHECKED_MALLOCZERO( h->frames.unused[1], (h->param.i_threads + 20) * sizeof(x264_frame_t *) );
 826     CHECKED_MALLOCZERO( h->frames.current, (h->param.i_sync_lookahead + h->param.i_bframe
 827                         + h->param.i_threads + 3) * sizeof(x264_frame_t *) );
 828
 829     h->i_ref0 = 0;
 830     h->i_ref1 = 0;
 831
 832     x264_rdo_init();
 833
 834     /* init CPU functions */
 835     x264_predict_16x16_init( h->param.cpu, h->predict_16x16 );
 836     x264_predict_8x8c_init( h->param.cpu, h->predict_8x8c );
 837     x264_predict_8x8_init( h->param.cpu, h->predict_8x8, &h->predict_8x8_filter );
 838     x264_predict_4x4_init( h->param.cpu, h->predict_4x4 );
 839     if( !h->param.b_cabac )
 840         x264_init_vlc_tables();
 841     x264_pixel_init( h->param.cpu, &h->pixf );
 842     x264_dct_init( h->param.cpu, &h->dctf );
 843     x264_zigzag_init( h->param.cpu, &h->zigzagf, h->param.b_interlaced );
 844     x264_mc_init( h->param.cpu, &h->mc );
 845     x264_quant_init( h, h->param.cpu, &h->quantf );
 846     x264_deblock_init( h->param.cpu, &h->loopf );
 847     x264_dct_init_weights();
 848
 849     mbcmp_init( h );
 850
 851     p = buf + sprintf( buf, "using cpu capabilities:" );
 852     for( i=0; x264_cpu_names[i].flags; i++ )
 853     {
 854         if( !strcmp(x264_cpu_names[i].name, "SSE2")
 855             && param->cpu & (X264_CPU_SSE2_IS_FAST|X264_CPU_SSE2_IS_SLOW) )
 856             continue;
 857         if( !strcmp(x264_cpu_names[i].name, "SSE3")
 858             && (param->cpu & X264_CPU_SSSE3 || !(param->cpu & X264_CPU_CACHELINE_64)) )
 859             continue;
 860         if( !strcmp(x264_cpu_names[i].name, "SSE4.1")
 861             && (param->cpu & X264_CPU_SSE42) )
 862             continue;
 863         if( (param->cpu & x264_cpu_names[i].flags) == x264_cpu_names[i].flags
 864             && (!i || x264_cpu_names[i].flags != x264_cpu_names[i-1].flags) )
 865             p += sprintf( p, " %s", x264_cpu_names[i].name );
 866     }
 867     if( !param->cpu )
 868         p += sprintf( p, " none!" );
 869     x264_log( h, X264_LOG_INFO, "%s\n", buf );
 870
 871     for( qp = h->param.rc.i_qp_min; qp <= h->param.rc.i_qp_max; qp++ )
 872         if( x264_analyse_init_costs( h, qp ) )
 873             goto fail;
 874     if( x264_analyse_init_costs( h, X264_LOOKAHEAD_QP ) )
 875         goto fail;
 876     if( h->cost_mv[1][2013] != 24 )
 877     {
 878         x264_log( h, X264_LOG_ERROR, "MV cost test failed: x264 has been miscompiled!\n" );
 879         goto fail;
 880     }
 881
 882     h->out.i_nal = 0;
 883     h->out.i_bitstream = X264_MAX( 1000000, h->param.i_width * h->param.i_height * 4
 884         * ( h->param.rc.i_rc_method == X264_RC_ABR ? pow( 0.95, h->param.rc.i_qp_min )
 885           : pow( 0.95, h->param.rc.i_qp_constant ) * X264_MAX( 1, h->param.rc.f_ip_factor )));
 886
 887     CHECKED_MALLOC( h->nal_buffer, h->out.i_bitstream * 3/2 + 4 );
 888     h->nal_buffer_size = h->out.i_bitstream * 3/2 + 4;
 889
 890     h->thread[0] = h;
 891     h->i_thread_num = 0;
 892     for( i = 1; i < h->param.i_threads + !!h->param.i_sync_lookahead; i++ )
 893         CHECKED_MALLOC( h->thread[i], sizeof(x264_t) );
 894
 895     for( i = 0; i < h->param.i_threads; i++ )
 896     {
 897         if( i > 0 )
 898             *h->thread[i] = *h;
 899         h->thread[i]->fdec = x264_frame_pop_unused( h, 1 );
 900         if( !h->thread[i]->fdec )
 901             goto fail;
 902         CHECKED_MALLOC( h->thread[i]->out.p_bitstream, h->out.i_bitstream );
 903         /* Start each thread with room for 8 NAL units; it'll realloc later if needed. */
 904         CHECKED_MALLOC( h->thread[i]->out.nal, 8*sizeof(x264_nal_t) );
 905         h->thread[i]->out.i_nals_allocated = 8;
 906         if( x264_macroblock_cache_init( h->thread[i] ) < 0 )
 907             goto fail;
 908     }
 909
 910     if( x264_lookahead_init( h, i_slicetype_length ) )
 911         goto fail;
 912
 913     if( x264_ratecontrol_new( h ) < 0 )
 914         goto fail;
 915
 916     if( h->param.psz_dump_yuv )
 917     {
 918         /* create or truncate the reconstructed video file */
 919         FILE *f = fopen( h->param.psz_dump_yuv, "w" );
 920         if( f )
 921             fclose( f );
 922         else
 923         {
 924             x264_log( h, X264_LOG_ERROR, "can't write to fdec.yuv\n" );
 925             goto fail;
 926         }
 927     }
 928
 929     x264_log( h, X264_LOG_INFO, "profile %s, level %d.%d\n",
 930         h->sps->i_profile_idc == PROFILE_BASELINE ? "Baseline" :
 931         h->sps->i_profile_idc == PROFILE_MAIN ? "Main" :
 932         h->sps->i_profile_idc == PROFILE_HIGH ? "High" :
 933         "High 4:4:4 Predictive", h->sps->i_level_idc/10, h->sps->i_level_idc%10 );
 934
 935     return h;
 936 fail:
 937     x264_free( h );
 938     return NULL;
 939 }
 940
 941 /****************************************************************************
 942  * x264_encoder_reconfig:
 943  ****************************************************************************/
 944 int x264_encoder_reconfig( x264_t *h, x264_param_t *param )
 945 {
 946     h = h->thread[h->i_thread_phase%h->param.i_threads];
 947     x264_set_aspect_ratio( h, param, 0 );
 948 #define COPY(var) h->param.var = param->var
 949     COPY( i_frame_reference ); // but never uses more refs than initially specified
 950     COPY( i_bframe_bias );
 951     if( h->param.i_scenecut_threshold )
 952         COPY( i_scenecut_threshold ); // can't turn it on or off, only vary the threshold
 953     COPY( b_deblocking_filter );
 954     COPY( i_deblocking_filter_alphac0 );
 955     COPY( i_deblocking_filter_beta );
 956     COPY( analyse.intra );
 957     COPY( analyse.inter );
 958     COPY( analyse.i_direct_mv_pred );
 959     /* Scratch buffer prevents me_range from being increased for esa/tesa */
 960     if( h->param.analyse.i_me_method < X264_ME_ESA || param->analyse.i_me_range < h->param.analyse.i_me_range )
 961         COPY( analyse.i_me_range );
 962     COPY( analyse.i_noise_reduction );
 963     /* We can't switch out of subme=0 during encoding. */
 964     if( h->param.analyse.i_subpel_refine )
 965         COPY( analyse.i_subpel_refine );
 966     COPY( analyse.i_trellis );
 967     COPY( analyse.b_chroma_me );
 968     COPY( analyse.b_dct_decimate );
 969     COPY( analyse.b_fast_pskip );
 970     COPY( analyse.b_mixed_references );
 971     COPY( analyse.f_psy_rd );
 972     COPY( analyse.f_psy_trellis );
 973     // can only twiddle these if they were enabled to begin with:
 974     if( h->param.analyse.i_me_method >= X264_ME_ESA || param->analyse.i_me_method < X264_ME_ESA )
 975         COPY( analyse.i_me_method );
 976     if( h->param.analyse.i_me_method >= X264_ME_ESA && !h->frames.b_have_sub8x8_esa )
 977         h->param.analyse.inter &= ~X264_ANALYSE_PSUB8x8;
 978     if( h->pps->b_transform_8x8_mode )
 979         COPY( analyse.b_transform_8x8 );
 980     if( h->frames.i_max_ref1 > 1 )
 981         COPY( b_bframe_pyramid );
 982     COPY( i_slice_max_size );
 983     COPY( i_slice_max_mbs );
 984     COPY( i_slice_count );
 985 #undef COPY
 986
 987     mbcmp_init( h );
 988
 989     return x264_validate_parameters( h );
 990 }
 991
 992 /* internal usage */
 993 static void x264_nal_start( x264_t *h, int i_type, int i_ref_idc )
 994 {
 995     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
 996
 997     nal->i_ref_idc = i_ref_idc;
 998     nal->i_type    = i_type;
 999
1000     nal->i_payload= 0;
1001     nal->p_payload= &h->out.p_bitstream[bs_pos( &h->out.bs ) / 8];
1002 }
1003 static int x264_nal_end( x264_t *h )
1004 {
1005     x264_nal_t *nal = &h->out.nal[h->out.i_nal];
1006     nal->i_payload = &h->out.p_bitstream[bs_pos( &h->out.bs ) / 8] - nal->p_payload;
1007     h->out.i_nal++;
1008
1009     /* if number of allocated nals is not enough, re-allocate a larger one. */
1010     if( h->out.i_nal >= h->out.i_nals_allocated )
1011     {
1012         x264_nal_t *new_out = x264_malloc( sizeof(x264_nal_t) * (h->out.i_nals_allocated*2) );
1013         if( !new_out )
1014             return -1;
1015         memcpy( new_out, h->out.nal, sizeof(x264_nal_t) * (h->out.i_nals_allocated) );
1016         x264_free( h->out.nal );
1017         h->out.nal = new_out;
1018         h->out.i_nals_allocated *= 2;
1019     }
1020     return 0;
1021 }
1022
1023 static int x264_encoder_encapsulate_nals( x264_t *h )
1024 {
1025     int nal_size = 0, i;
1026     for( i = 0; i < h->out.i_nal; i++ )
1027         nal_size += h->out.nal[i].i_payload;
1028
1029     /* Worst-case NAL unit escaping: reallocate the buffer if it's too small. */
1030     if( h->nal_buffer_size < nal_size * 3/2 + h->out.i_nal * 4 )
1031     {
1032         uint8_t *buf = x264_malloc( nal_size * 2 + h->out.i_nal * 4 );
1033         if( !buf )
1034             return -1;
1035         x264_free( h->nal_buffer );
1036         h->nal_buffer = buf;
1037     }
1038
1039     uint8_t *nal_buffer = h->nal_buffer;
1040
1041     for( i = 0; i < h->out.i_nal; i++ )
1042     {
1043         int size = x264_nal_encode( nal_buffer, h->param.b_annexb, &h->out.nal[i] );
1044         h->out.nal[i].i_payload = size;
1045         h->out.nal[i].p_payload = nal_buffer;
1046         nal_buffer += size;
1047     }
1048
1049     return nal_buffer - h->nal_buffer;
1050 }
1051
1052 /****************************************************************************
1053  * x264_encoder_headers:
1054  ****************************************************************************/
1055 int x264_encoder_headers( x264_t *h, x264_nal_t **pp_nal, int *pi_nal )
1056 {
1057     int frame_size = 0;
1058     /* init bitstream context */
1059     h->out.i_nal = 0;
1060     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
1061
1062     /* Write SEI, SPS and PPS. */
1063     x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
1064     if( x264_sei_version_write( h, &h->out.bs ) )
1065         return -1;
1066     if( x264_nal_end( h ) )
1067         return -1;
1068
1069     /* generate sequence parameters */
1070     x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
1071     x264_sps_write( &h->out.bs, h->sps );
1072     if( x264_nal_end( h ) )
1073         return -1;
1074
1075     /* generate picture parameters */
1076     x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
1077     x264_pps_write( &h->out.bs, h->pps );
1078     if( x264_nal_end( h ) )
1079         return -1;
1080     bs_flush( &h->out.bs );
1081
1082     frame_size = x264_encoder_encapsulate_nals( h );
1083
1084     /* now set output*/
1085     *pi_nal = h->out.i_nal;
1086     *pp_nal = &h->out.nal[0];
1087     h->out.i_nal = 0;
1088
1089     return frame_size;
1090 }
1091
1092 static inline void x264_reference_build_list( x264_t *h, int i_poc )
1093 {
1094     int i;
1095     int b_ok;
1096
1097     /* build ref list 0/1 */
1098     h->i_ref0 = 0;
1099     h->i_ref1 = 0;
1100     for( i = 0; h->frames.reference[i]; i++ )
1101     {
1102         if( h->frames.reference[i]->i_poc < i_poc )
1103         {
1104             h->fref0[h->i_ref0++] = h->frames.reference[i];
1105         }
1106         else if( h->frames.reference[i]->i_poc > i_poc )
1107         {
1108             h->fref1[h->i_ref1++] = h->frames.reference[i];
1109         }
1110     }
1111
1112     /* Order ref0 from higher to lower poc */
1113     do
1114     {
1115         b_ok = 1;
1116         for( i = 0; i < h->i_ref0 - 1; i++ )
1117         {
1118             if( h->fref0[i]->i_poc < h->fref0[i+1]->i_poc )
1119             {
1120                 XCHG( x264_frame_t*, h->fref0[i], h->fref0[i+1] );
1121                 b_ok = 0;
1122                 break;
1123             }
1124         }
1125     } while( !b_ok );
1126     /* Order ref1 from lower to higher poc (bubble sort) for B-frame */
1127     do
1128     {
1129         b_ok = 1;
1130         for( i = 0; i < h->i_ref1 - 1; i++ )
1131         {
1132             if( h->fref1[i]->i_poc > h->fref1[i+1]->i_poc )
1133             {
1134                 XCHG( x264_frame_t*, h->fref1[i], h->fref1[i+1] );
1135                 b_ok = 0;
1136                 break;
1137             }
1138         }
1139     } while( !b_ok );
1140
1141     /* In the standard, a P-frame's ref list is sorted by frame_num.
1142      * We use POC, but check whether explicit reordering is needed */
1143     h->b_ref_reorder[0] =
1144     h->b_ref_reorder[1] = 0;
1145     if( h->sh.i_type == SLICE_TYPE_P )
1146     {
1147         for( i = 0; i < h->i_ref0 - 1; i++ )
1148             if( h->fref0[i]->i_frame_num < h->fref0[i+1]->i_frame_num )
1149             {
1150                 h->b_ref_reorder[0] = 1;
1151                 break;
1152             }
1153     }
1154
1155     h->i_ref1 = X264_MIN( h->i_ref1, h->frames.i_max_ref1 );
1156     h->i_ref0 = X264_MIN( h->i_ref0, h->frames.i_max_ref0 );
1157     h->i_ref0 = X264_MIN( h->i_ref0, h->param.i_frame_reference ); // if reconfig() has lowered the limit
1158     assert( h->i_ref0 + h->i_ref1 <= 16 );
1159     h->mb.pic.i_fref[0] = h->i_ref0;
1160     h->mb.pic.i_fref[1] = h->i_ref1;
1161 }
1162
1163 static void x264_fdec_filter_row( x264_t *h, int mb_y )
1164 {
1165     /* mb_y is the mb to be encoded next, not the mb to be filtered here */
1166     int b_hpel = h->fdec->b_kept_as_ref;
1167     int b_deblock = !h->sh.i_disable_deblocking_filter_idc;
1168     int b_end = mb_y == h->sps->i_mb_height;
1169     int min_y = mb_y - (1 << h->sh.b_mbaff);
1170     int max_y = b_end ? h->sps->i_mb_height : mb_y;
1171     b_deblock &= b_hpel || h->param.psz_dump_yuv;
1172     if( mb_y & h->sh.b_mbaff )
1173         return;
1174     if( min_y < 0 )
1175         return;
1176
1177     if( !b_end )
1178     {
1179         int i, j;
1180         for( j=0; j<=h->sh.b_mbaff; j++ )
1181             for( i=0; i<3; i++ )
1182             {
1183                 memcpy( h->mb.intra_border_backup[j][i],
1184                         h->fdec->plane[i] + ((mb_y*16 >> !!i) + j - 1 - h->sh.b_mbaff) * h->fdec->i_stride[i],
1185                         h->sps->i_mb_width*16 >> !!i );
1186             }
1187     }
1188
1189     if( b_deblock )
1190     {
1191         int y;
1192         for( y = min_y; y < max_y; y += (1 << h->sh.b_mbaff) )
1193             x264_frame_deblock_row( h, y );
1194     }
1195
1196     if( b_hpel )
1197     {
1198         x264_frame_expand_border( h, h->fdec, min_y, b_end );
1199         if( h->param.analyse.i_subpel_refine )
1200         {
1201             x264_frame_filter( h, h->fdec, min_y, b_end );
1202             x264_frame_expand_border_filtered( h, h->fdec, min_y, b_end );
1203         }
1204     }
1205
1206     if( h->param.i_threads > 1 && h->fdec->b_kept_as_ref )
1207     {
1208         x264_frame_cond_broadcast( h->fdec, mb_y*16 + (b_end ? 10000 : -(X264_THREAD_HEIGHT << h->sh.b_mbaff)) );
1209     }
1210
1211     min_y = X264_MAX( min_y*16-8, 0 );
1212     max_y = b_end ? h->param.i_height : mb_y*16-8;
1213
1214     if( h->param.analyse.b_psnr )
1215     {
1216         int i;
1217         for( i=0; i<3; i++ )
1218             h->stat.frame.i_ssd[i] +=
1219                 x264_pixel_ssd_wxh( &h->pixf,
1220                     h->fdec->plane[i] + (min_y>>!!i) * h->fdec->i_stride[i], h->fdec->i_stride[i],
1221                     h->fenc->plane[i] + (min_y>>!!i) * h->fenc->i_stride[i], h->fenc->i_stride[i],
1222                     h->param.i_width >> !!i, (max_y-min_y) >> !!i );
1223     }
1224
1225     if( h->param.analyse.b_ssim )
1226     {
1227         x264_emms();
1228         /* offset by 2 pixels to avoid alignment of ssim blocks with dct blocks,
1229          * and overlap by 4 */
1230         min_y += min_y == 0 ? 2 : -6;
1231         h->stat.frame.f_ssim +=
1232             x264_pixel_ssim_wxh( &h->pixf,
1233                 h->fdec->plane[0] + 2+min_y*h->fdec->i_stride[0], h->fdec->i_stride[0],
1234                 h->fenc->plane[0] + 2+min_y*h->fenc->i_stride[0], h->fenc->i_stride[0],
1235                 h->param.i_width-2, max_y-min_y, h->scratch_buffer );
1236     }
1237 }
1238
1239 static inline int x264_reference_update( x264_t *h )
1240 {
1241     if( h->fdec->i_frame >= 0 )
1242         h->i_frame++;
1243
1244     if( !h->fdec->b_kept_as_ref )
1245     {
1246         if( h->param.i_threads > 1 )
1247         {
1248             x264_frame_push_unused( h, h->fdec );
1249             h->fdec = x264_frame_pop_unused( h, 1 );
1250             if( !h->fdec )
1251                 return -1;
1252         }
1253         return 0;
1254     }
1255
1256     /* move frame in the buffer */
1257     x264_frame_push( h->frames.reference, h->fdec );
1258     if( h->frames.reference[h->frames.i_max_dpb] )
1259         x264_frame_push_unused( h, x264_frame_shift( h->frames.reference ) );
1260     h->fdec = x264_frame_pop_unused( h, 1 );
1261     if( !h->fdec )
1262         return -1;
1263     return 0;
1264 }
1265
1266 static inline void x264_reference_reset( x264_t *h )
1267 {
1268     while( h->frames.reference[0] )
1269         x264_frame_push_unused( h, x264_frame_pop( h->frames.reference ) );
1270     h->fdec->i_poc =
1271     h->fenc->i_poc = 0;
1272 }
1273
1274 static inline void x264_slice_init( x264_t *h, int i_nal_type, int i_global_qp )
1275 {
1276     /* ------------------------ Create slice header  ----------------------- */
1277     if( i_nal_type == NAL_SLICE_IDR )
1278     {
1279         x264_slice_header_init( h, &h->sh, h->sps, h->pps, h->i_idr_pic_id, h->i_frame_num, i_global_qp );
1280
1281         /* increment id */
1282         h->i_idr_pic_id = ( h->i_idr_pic_id + 1 ) % 65536;
1283     }
1284     else
1285     {
1286         x264_slice_header_init( h, &h->sh, h->sps, h->pps, -1, h->i_frame_num, i_global_qp );
1287
1288         /* always set the real higher num of ref frame used */
1289         h->sh.b_num_ref_idx_override = 1;
1290         h->sh.i_num_ref_idx_l0_active = h->i_ref0 <= 0 ? 1 : h->i_ref0;
1291         h->sh.i_num_ref_idx_l1_active = h->i_ref1 <= 0 ? 1 : h->i_ref1;
1292     }
1293
1294     h->fdec->i_frame_num = h->sh.i_frame_num;
1295
1296     if( h->sps->i_poc_type == 0 )
1297     {
1298         h->sh.i_poc_lsb = h->fdec->i_poc & ( (1 << h->sps->i_log2_max_poc_lsb) - 1 );
1299         h->sh.i_delta_poc_bottom = 0;   /* XXX won't work for field */
1300     }
1301     else if( h->sps->i_poc_type == 1 )
1302     {
1303         /* FIXME TODO FIXME */
1304     }
1305     else
1306     {
1307         /* Nothing to do ? */
1308     }
1309
1310     x264_macroblock_slice_init( h );
1311 }
1312
1313 static int x264_slice_write( x264_t *h )
1314 {
1315     int i_skip;
1316     int mb_xy, i_mb_x, i_mb_y;
1317     int i, i_list, i_ref, i_skip_bak = 0; /* Shut up GCC. */
1318     bs_t bs_bak;
1319     x264_cabac_t cabac_bak;
1320     uint8_t cabac_prevbyte_bak = 0; /* Shut up GCC. */
1321     /* Assume no more than 3 bytes of NALU escaping. */
1322     int slice_max_size = h->param.i_slice_max_size > 0 ? (h->param.i_slice_max_size-3-NALU_OVERHEAD)*8 : INT_MAX;
1323     int starting_bits = bs_pos(&h->out.bs);
1324
1325     /* Slice */
1326     x264_nal_start( h, h->i_nal_type, h->i_nal_ref_idc );
1327
1328     /* Slice header */
1329     x264_slice_header_write( &h->out.bs, &h->sh, h->i_nal_ref_idc );
1330     if( h->param.b_cabac )
1331     {
1332         /* alignment needed */
1333         bs_align_1( &h->out.bs );
1334
1335         /* init cabac */
1336         x264_cabac_context_init( &h->cabac, h->sh.i_type, h->sh.i_qp, h->sh.i_cabac_init_idc );
1337         x264_cabac_encode_init ( &h->cabac, h->out.bs.p, h->out.bs.p_end );
1338     }
1339     h->mb.i_last_qp = h->sh.i_qp;
1340     h->mb.i_last_dqp = 0;
1341
1342     i_mb_y = h->sh.i_first_mb / h->sps->i_mb_width;
1343     i_mb_x = h->sh.i_first_mb % h->sps->i_mb_width;
1344     i_skip = 0;
1345
1346     while( (mb_xy = i_mb_x + i_mb_y * h->sps->i_mb_width) <= h->sh.i_last_mb )
1347     {
1348         int mb_spos = bs_pos(&h->out.bs) + x264_cabac_pos(&h->cabac);
1349         if( h->param.i_slice_max_size > 0 )
1350         {
1351             /* We don't need the contexts because flushing the CABAC encoder has no context
1352              * dependency and macroblocks are only re-encoded in the case where a slice is
1353              * ended (and thus the content of all contexts are thrown away). */
1354             if( h->param.b_cabac )
1355             {
1356                 memcpy( &cabac_bak, &h->cabac, offsetof(x264_cabac_t, f8_bits_encoded) );
1357                 /* x264's CABAC writer modifies the previous byte during carry, so it has to be
1358                  * backed up. */
1359                 cabac_prevbyte_bak = h->cabac.p[-1];
1360             }
1361             else
1362             {
1363                 bs_bak = h->out.bs;
1364                 i_skip_bak = i_skip;
1365             }
1366         }
1367
1368         if( i_mb_x == 0 && !h->mb.b_reencode_mb )
1369             x264_fdec_filter_row( h, i_mb_y );
1370
1371         /* load cache */
1372         x264_macroblock_cache_load( h, i_mb_x, i_mb_y );
1373
1374         x264_macroblock_analyse( h );
1375
1376         /* encode this macroblock -> be careful it can change the mb type to P_SKIP if needed */
1377         x264_macroblock_encode( h );
1378
1379         if( x264_bitstream_check_buffer( h ) )
1380             return -1;
1381
1382         if( h->param.b_cabac )
1383         {
1384             if( mb_xy > h->sh.i_first_mb && !(h->sh.b_mbaff && (i_mb_y&1)) )
1385                 x264_cabac_encode_terminal( &h->cabac );
1386
1387             if( IS_SKIP( h->mb.i_type ) )
1388                 x264_cabac_mb_skip( h, 1 );
1389             else
1390             {
1391                 if( h->sh.i_type != SLICE_TYPE_I )
1392                     x264_cabac_mb_skip( h, 0 );
1393                 x264_macroblock_write_cabac( h, &h->cabac );
1394             }
1395         }
1396         else
1397         {
1398             if( IS_SKIP( h->mb.i_type ) )
1399                 i_skip++;
1400             else
1401             {
1402                 if( h->sh.i_type != SLICE_TYPE_I )
1403                 {
1404                     bs_write_ue( &h->out.bs, i_skip );  /* skip run */
1405                     i_skip = 0;
1406                 }
1407                 x264_macroblock_write_cavlc( h, &h->out.bs );
1408             }
1409         }
1410
1411         int total_bits = bs_pos(&h->out.bs) + x264_cabac_pos(&h->cabac);
1412         int mb_size = total_bits - mb_spos;
1413
1414         /* We'll just re-encode this last macroblock if we go over the max slice size. */
1415         if( total_bits - starting_bits > slice_max_size && !h->mb.b_reencode_mb )
1416         {
1417             if( mb_xy != h->sh.i_first_mb )
1418             {
1419                 if( h->param.b_cabac )
1420                 {
1421                     memcpy( &h->cabac, &cabac_bak, offsetof(x264_cabac_t, f8_bits_encoded) );
1422                     h->cabac.p[-1] = cabac_prevbyte_bak;
1423                 }
1424                 else
1425                 {
1426                     h->out.bs = bs_bak;
1427                     i_skip = i_skip_bak;
1428                 }
1429                 h->mb.b_reencode_mb = 1;
1430                 h->sh.i_last_mb = mb_xy-1;
1431                 break;
1432             }
1433             else
1434             {
1435                 h->sh.i_last_mb = mb_xy;
1436                 h->mb.b_reencode_mb = 0;
1437             }
1438         }
1439         else
1440             h->mb.b_reencode_mb = 0;
1441
1442 #if VISUALIZE
1443         if( h->param.b_visualize )
1444             x264_visualize_mb( h );
1445 #endif
1446
1447         /* save cache */
1448         x264_macroblock_cache_save( h );
1449
1450         /* accumulate mb stats */
1451         h->stat.frame.i_mb_count[h->mb.i_type]++;
1452         if( h->param.i_log_level >= X264_LOG_INFO )
1453         {
1454             if( !IS_SKIP(h->mb.i_type) && !IS_INTRA(h->mb.i_type) && !IS_DIRECT(h->mb.i_type) )
1455             {
1456                 if( h->mb.i_partition != D_8x8 )
1457                     h->stat.frame.i_mb_partition[h->mb.i_partition] += 4;
1458                 else
1459                     for( i = 0; i < 4; i++ )
1460                         h->stat.frame.i_mb_partition[h->mb.i_sub_partition[i]] ++;
1461                 if( h->param.i_frame_reference > 1 )
1462                     for( i_list = 0; i_list <= (h->sh.i_type == SLICE_TYPE_B); i_list++ )
1463                         for( i = 0; i < 4; i++ )
1464                         {
1465                             i_ref = h->mb.cache.ref[i_list][ x264_scan8[4*i] ];
1466                             if( i_ref >= 0 )
1467                                 h->stat.frame.i_mb_count_ref[i_list][i_ref] ++;
1468                         }
1469             }
1470             if( h->mb.i_cbp_luma || h->mb.i_cbp_chroma )
1471             {
1472                 int cbpsum = (h->mb.i_cbp_luma&1) + ((h->mb.i_cbp_luma>>1)&1)
1473                            + ((h->mb.i_cbp_luma>>2)&1) + (h->mb.i_cbp_luma>>3);
1474                 int b_intra = IS_INTRA(h->mb.i_type);
1475                 h->stat.frame.i_mb_cbp[!b_intra + 0] += cbpsum;
1476                 h->stat.frame.i_mb_cbp[!b_intra + 2] += h->mb.i_cbp_chroma >= 1;
1477                 h->stat.frame.i_mb_cbp[!b_intra + 4] += h->mb.i_cbp_chroma == 2;
1478             }
1479             if( h->mb.i_cbp_luma && !IS_INTRA(h->mb.i_type) )
1480             {
1481                 h->stat.frame.i_mb_count_8x8dct[0] ++;
1482                 h->stat.frame.i_mb_count_8x8dct[1] += h->mb.b_transform_8x8;
1483             }
1484             if( IS_INTRA(h->mb.i_type) && h->mb.i_type != I_PCM )
1485             {
1486                 if( h->mb.i_type == I_16x16 )
1487                     h->stat.frame.i_mb_pred_mode[0][h->mb.i_intra16x16_pred_mode]++;
1488                 else if( h->mb.i_type == I_8x8 )
1489                     for( i = 0; i < 16; i += 4 )
1490                         h->stat.frame.i_mb_pred_mode[1][h->mb.cache.intra4x4_pred_mode[x264_scan8[i]]]++;
1491                 else //if( h->mb.i_type == I_4x4 )
1492                     for( i = 0; i < 16; i++ )
1493                         h->stat.frame.i_mb_pred_mode[2][h->mb.cache.intra4x4_pred_mode[x264_scan8[i]]]++;
1494             }
1495         }
1496
1497         x264_ratecontrol_mb( h, mb_size );
1498
1499         if( h->sh.b_mbaff )
1500         {
1501             i_mb_x += i_mb_y & 1;
1502             i_mb_y ^= i_mb_x < h->sps->i_mb_width;
1503         }
1504         else
1505             i_mb_x++;
1506         if( i_mb_x == h->sps->i_mb_width )
1507         {
1508             i_mb_y++;
1509             i_mb_x = 0;
1510         }
1511     }
1512
1513     if( h->param.b_cabac )
1514     {
1515         x264_cabac_encode_flush( h, &h->cabac );
1516         h->out.bs.p = h->cabac.p;
1517     }
1518     else
1519     {
1520         if( i_skip > 0 )
1521             bs_write_ue( &h->out.bs, i_skip );  /* last skip run */
1522         /* rbsp_slice_trailing_bits */
1523         bs_rbsp_trailing( &h->out.bs );
1524         bs_flush( &h->out.bs );
1525     }
1526     if( x264_nal_end( h ) )
1527         return -1;
1528
1529     if( h->sh.i_last_mb == h->mb.i_mb_count-1 )
1530     {
1531         h->stat.frame.i_misc_bits = bs_pos( &h->out.bs )
1532                                   + (h->out.i_nal*NALU_OVERHEAD * 8)
1533                                   - h->stat.frame.i_tex_bits
1534                                   - h->stat.frame.i_mv_bits;
1535         x264_fdec_filter_row( h, h->sps->i_mb_height );
1536     }
1537
1538     return 0;
1539 }
1540
1541 static void x264_thread_sync_context( x264_t *dst, x264_t *src )
1542 {
1543     x264_frame_t **f;
1544     if( dst == src )
1545         return;
1546
1547     // reference counting
1548     for( f = src->frames.reference; *f; f++ )
1549         (*f)->i_reference_count++;
1550     for( f = dst->frames.reference; *f; f++ )
1551         x264_frame_push_unused( src, *f );
1552     src->fdec->i_reference_count++;
1553     x264_frame_push_unused( src, dst->fdec );
1554
1555     // copy everything except the per-thread pointers and the constants.
1556     memcpy( &dst->i_frame, &src->i_frame, offsetof(x264_t, mb.type) - offsetof(x264_t, i_frame) );
1557     dst->param = src->param;
1558     dst->stat = src->stat;
1559 }
1560
1561 static void x264_thread_sync_stat( x264_t *dst, x264_t *src )
1562 {
1563     if( dst == src )
1564         return;
1565     memcpy( &dst->stat.i_frame_count, &src->stat.i_frame_count, sizeof(dst->stat) - sizeof(dst->stat.frame) );
1566 }
1567
1568 static void *x264_slices_write( x264_t *h )
1569 {
1570     int i_slice_num = 0;
1571     if( h->param.i_sync_lookahead )
1572         x264_lower_thread_priority( 10 );
1573
1574 #ifdef HAVE_MMX
1575     /* Misalign mask has to be set separately for each thread. */
1576     if( h->param.cpu&X264_CPU_SSE_MISALIGN )
1577         x264_cpu_mask_misalign_sse();
1578 #endif
1579
1580 #if VISUALIZE
1581     if( h->param.b_visualize )
1582         if( x264_visualize_init( h ) )
1583             return (void *)-1;
1584 #endif
1585
1586     /* init stats */
1587     memset( &h->stat.frame, 0, sizeof(h->stat.frame) );
1588     h->mb.b_reencode_mb = 0;
1589     while( h->sh.i_first_mb < h->mb.i_mb_count )
1590     {
1591         h->sh.i_last_mb = h->mb.i_mb_count - 1;
1592         if( h->param.i_slice_max_mbs )
1593             h->sh.i_last_mb = h->sh.i_first_mb + h->param.i_slice_max_mbs - 1;
1594         else if( h->param.i_slice_count )
1595         {
1596             x264_emms();
1597             i_slice_num++;
1598             double height = h->sps->i_mb_height >> h->param.b_interlaced;
1599             int width = h->sps->i_mb_width << h->param.b_interlaced;
1600             h->sh.i_last_mb = (int)(height * i_slice_num / h->param.i_slice_count + 0.5) * width - 1;
1601         }
1602         h->sh.i_last_mb = X264_MIN( h->sh.i_last_mb, h->mb.i_mb_count - 1 );
1603         if( x264_stack_align( x264_slice_write, h ) )
1604             return (void *)-1;
1605         h->sh.i_first_mb = h->sh.i_last_mb + 1;
1606     }
1607
1608 #if VISUALIZE
1609     if( h->param.b_visualize )
1610     {
1611         x264_visualize_show( h );
1612         x264_visualize_close( h );
1613     }
1614 #endif
1615
1616     return (void *)0;
1617 }
1618
1619 /****************************************************************************
1620  * x264_encoder_encode:
1621  *  XXX: i_poc   : is the poc of the current given picture
1622  *       i_frame : is the number of the frame being coded
1623  *  ex:  type frame poc
1624  *       I      0   2*0
1625  *       P      1   2*3
1626  *       B      2   2*1
1627  *       B      3   2*2
1628  *       P      4   2*6
1629  *       B      5   2*4
1630  *       B      6   2*5
1631  ****************************************************************************/
1632 int     x264_encoder_encode( x264_t *h,
1633                              x264_nal_t **pp_nal, int *pi_nal,
1634                              x264_picture_t *pic_in,
1635                              x264_picture_t *pic_out )
1636 {
1637     x264_t *thread_current, *thread_prev, *thread_oldest;
1638     int     i_nal_type;
1639     int     i_nal_ref_idc;
1640
1641     int   i_global_qp;
1642
1643     if( h->param.i_threads > 1)
1644     {
1645         int i = ++h->i_thread_phase;
1646         int t = h->param.i_threads;
1647         thread_current = h->thread[ i%t ];
1648         thread_prev    = h->thread[ (i-1)%t ];
1649         thread_oldest  = h->thread[ (i+1)%t ];
1650         x264_thread_sync_context( thread_current, thread_prev );
1651         x264_thread_sync_ratecontrol( thread_current, thread_prev, thread_oldest );
1652         h = thread_current;
1653 //      fprintf(stderr, "current: %p  prev: %p  oldest: %p \n", thread_current, thread_prev, thread_oldest);
1654     }
1655     else
1656     {
1657         thread_current =
1658         thread_oldest  = h;
1659     }
1660
1661     // ok to call this before encoding any frames, since the initial values of fdec have b_kept_as_ref=0
1662     if( x264_reference_update( h ) )
1663         return -1;
1664     h->fdec->i_lines_completed = -1;
1665
1666     /* no data out */
1667     *pi_nal = 0;
1668     *pp_nal = NULL;
1669
1670     /* ------------------- Setup new frame from picture -------------------- */
1671     if( pic_in != NULL )
1672     {
1673         /* 1: Copy the picture to a frame and move it to a buffer */
1674         x264_frame_t *fenc = x264_frame_pop_unused( h, 0 );
1675         if( !fenc )
1676             return -1;
1677
1678         if( x264_frame_copy_picture( h, fenc, pic_in ) < 0 )
1679             return -1;
1680
1681         if( h->param.i_width != 16 * h->sps->i_mb_width ||
1682             h->param.i_height != 16 * h->sps->i_mb_height )
1683             x264_frame_expand_border_mod16( h, fenc );
1684
1685         fenc->i_frame = h->frames.i_input++;
1686
1687         if( h->frames.b_have_lowres )
1688             x264_frame_init_lowres( h, fenc );
1689
1690         if( h->param.rc.b_mb_tree && h->param.rc.b_stat_read )
1691         {
1692             if( x264_macroblock_tree_read( h, fenc ) )
1693                 return -1;
1694         }
1695         else if( h->param.rc.i_aq_mode )
1696             x264_adaptive_quant_frame( h, fenc );
1697
1698         /* 2: Place the frame into the queue for its slice type decision */
1699         x264_lookahead_put_frame( h, fenc );
1700
1701         if( h->frames.i_input <= h->frames.i_delay + 1 - h->param.i_threads )
1702         {
1703             /* Nothing yet to encode, waiting for filling of buffers */
1704             pic_out->i_type = X264_TYPE_AUTO;
1705             return 0;
1706         }
1707     }
1708     else
1709     {
1710         /* signal kills for lookahead thread */
1711         h->lookahead->b_exit_thread = 1;
1712         x264_pthread_cond_broadcast( &h->lookahead->ifbuf.cv_fill );
1713     }
1714
1715     /* 3: The picture is analyzed in the lookahead */
1716     if( !h->frames.current[0] )
1717         x264_lookahead_get_frames( h );
1718
1719     if( !h->frames.current[0] && x264_lookahead_is_empty( h ) )
1720         return x264_encoder_frame_end( thread_oldest, thread_current, pp_nal, pi_nal, pic_out );
1721
1722     /* ------------------- Get frame to be encoded ------------------------- */
1723     /* 4: get picture to encode */
1724     h->fenc = x264_frame_shift( h->frames.current );
1725     if( h->fenc->param )
1726     {
1727         x264_encoder_reconfig( h, h->fenc->param );
1728         if( h->fenc->param->param_free )
1729             h->fenc->param->param_free( h->fenc->param );
1730     }
1731
1732     if( h->fenc->i_type == X264_TYPE_IDR )
1733     {
1734         h->frames.i_last_idr = h->fenc->i_frame;
1735         h->i_frame_num = 0;
1736     }
1737
1738     /* ------------------- Setup frame context ----------------------------- */
1739     /* 5: Init data dependent of frame type */
1740     if( h->fenc->i_type == X264_TYPE_IDR )
1741     {
1742         /* reset ref pictures */
1743         x264_reference_reset( h );
1744
1745         i_nal_type    = NAL_SLICE_IDR;
1746         i_nal_ref_idc = NAL_PRIORITY_HIGHEST;
1747         h->sh.i_type = SLICE_TYPE_I;
1748     }
1749     else if( h->fenc->i_type == X264_TYPE_I )
1750     {
1751         i_nal_type    = NAL_SLICE;
1752         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
1753         h->sh.i_type = SLICE_TYPE_I;
1754     }
1755     else if( h->fenc->i_type == X264_TYPE_P )
1756     {
1757         i_nal_type    = NAL_SLICE;
1758         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* Not completely true but for now it is (as all I/P are kept as ref)*/
1759         h->sh.i_type = SLICE_TYPE_P;
1760     }
1761     else if( h->fenc->i_type == X264_TYPE_BREF )
1762     {
1763         i_nal_type    = NAL_SLICE;
1764         i_nal_ref_idc = NAL_PRIORITY_HIGH; /* maybe add MMCO to forget it? -> low */
1765         h->sh.i_type = SLICE_TYPE_B;
1766     }
1767     else    /* B frame */
1768     {
1769         i_nal_type    = NAL_SLICE;
1770         i_nal_ref_idc = NAL_PRIORITY_DISPOSABLE;
1771         h->sh.i_type = SLICE_TYPE_B;
1772     }
1773
1774     h->fdec->i_poc =
1775     h->fenc->i_poc = 2 * (h->fenc->i_frame - h->frames.i_last_idr);
1776     h->fdec->i_type = h->fenc->i_type;
1777     h->fdec->i_frame = h->fenc->i_frame;
1778     h->fenc->b_kept_as_ref =
1779     h->fdec->b_kept_as_ref = i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE && h->param.i_keyint_max > 1;
1780
1781
1782
1783     /* ------------------- Init                ----------------------------- */
1784     /* build ref list 0/1 */
1785     x264_reference_build_list( h, h->fdec->i_poc );
1786
1787     if( h->sh.i_type == SLICE_TYPE_B )
1788         x264_macroblock_bipred_init( h );
1789
1790     /* ---------------------- Write the bitstream -------------------------- */
1791     /* Init bitstream context */
1792     h->out.i_nal = 0;
1793     bs_init( &h->out.bs, h->out.p_bitstream, h->out.i_bitstream );
1794
1795     if(h->param.b_aud){
1796         int pic_type;
1797
1798         if(h->sh.i_type == SLICE_TYPE_I)
1799             pic_type = 0;
1800         else if(h->sh.i_type == SLICE_TYPE_P)
1801             pic_type = 1;
1802         else if(h->sh.i_type == SLICE_TYPE_B)
1803             pic_type = 2;
1804         else
1805             pic_type = 7;
1806
1807         x264_nal_start(h, NAL_AUD, NAL_PRIORITY_DISPOSABLE);
1808         bs_write(&h->out.bs, 3, pic_type);
1809         bs_rbsp_trailing(&h->out.bs);
1810         if( x264_nal_end( h ) )
1811             return -1;
1812     }
1813
1814     h->i_nal_type = i_nal_type;
1815     h->i_nal_ref_idc = i_nal_ref_idc;
1816
1817     int overhead = NALU_OVERHEAD;
1818
1819     /* Write SPS and PPS */
1820     if( i_nal_type == NAL_SLICE_IDR && h->param.b_repeat_headers )
1821     {
1822         if( h->fenc->i_frame == 0 )
1823         {
1824             /* identify ourself */
1825             x264_nal_start( h, NAL_SEI, NAL_PRIORITY_DISPOSABLE );
1826             if( x264_sei_version_write( h, &h->out.bs ) )
1827                 return -1;
1828             if( x264_nal_end( h ) )
1829                 return -1;
1830             overhead += h->out.nal[h->out.i_nal-1].i_payload + NALU_OVERHEAD;
1831         }
1832
1833         /* generate sequence parameters */
1834         x264_nal_start( h, NAL_SPS, NAL_PRIORITY_HIGHEST );
1835         x264_sps_write( &h->out.bs, h->sps );
1836         if( x264_nal_end( h ) )
1837             return -1;
1838         overhead += h->out.nal[h->out.i_nal-1].i_payload + NALU_OVERHEAD;
1839
1840         /* generate picture parameters */
1841         x264_nal_start( h, NAL_PPS, NAL_PRIORITY_HIGHEST );
1842         x264_pps_write( &h->out.bs, h->pps );
1843         if( x264_nal_end( h ) )
1844             return -1;
1845         overhead += h->out.nal[h->out.i_nal-1].i_payload + NALU_OVERHEAD;
1846     }
1847
1848     /* Init the rate control */
1849     /* FIXME: Include slice header bit cost. */
1850     x264_ratecontrol_start( h, h->fenc->i_qpplus1, overhead*8 );
1851     i_global_qp = x264_ratecontrol_qp( h );
1852
1853     pic_out->i_qpplus1 =
1854     h->fdec->i_qpplus1 = i_global_qp + 1;
1855
1856     /* ------------------------ Create slice header  ----------------------- */
1857     x264_slice_init( h, i_nal_type, i_global_qp );
1858
1859     if( i_nal_ref_idc != NAL_PRIORITY_DISPOSABLE )
1860         h->i_frame_num++;
1861
1862     /* Write frame */
1863     if( h->param.i_threads > 1 )
1864     {
1865         if( x264_pthread_create( &h->thread_handle, NULL, (void*)x264_slices_write, h ) )
1866             return -1;
1867         h->b_thread_active = 1;
1868     }
1869     else
1870         if( (intptr_t)x264_slices_write( h ) )
1871             return -1;
1872
1873     return x264_encoder_frame_end( thread_oldest, thread_current, pp_nal, pi_nal, pic_out );
1874 }
1875
1876 static int x264_encoder_frame_end( x264_t *h, x264_t *thread_current,
1877                                    x264_nal_t **pp_nal, int *pi_nal,
1878                                    x264_picture_t *pic_out )
1879 {
1880     int i, j, i_list, frame_size;
1881     char psz_message[80];
1882
1883     if( h->b_thread_active )
1884     {
1885         void *ret = NULL;
1886         x264_pthread_join( h->thread_handle, &ret );
1887         if( (intptr_t)ret )
1888             return (intptr_t)ret;
1889         h->b_thread_active = 0;
1890     }
1891     if( !h->out.i_nal )
1892     {
1893         pic_out->i_type = X264_TYPE_AUTO;
1894         return 0;
1895     }
1896
1897     x264_frame_push_unused( thread_current, h->fenc );
1898
1899     /* End bitstream, set output  */
1900     *pi_nal = h->out.i_nal;
1901     *pp_nal = h->out.nal;
1902
1903     frame_size = x264_encoder_encapsulate_nals( h );
1904
1905     h->out.i_nal = 0;
1906
1907     /* Set output picture properties */
1908     if( h->sh.i_type == SLICE_TYPE_I )
1909         pic_out->i_type = h->i_nal_type == NAL_SLICE_IDR ? X264_TYPE_IDR : X264_TYPE_I;
1910     else if( h->sh.i_type == SLICE_TYPE_P )
1911         pic_out->i_type = X264_TYPE_P;
1912     else
1913         pic_out->i_type = X264_TYPE_B;
1914     pic_out->i_pts = h->fenc->i_pts;
1915
1916     pic_out->img.i_plane = h->fdec->i_plane;
1917     for(i = 0; i < 3; i++)
1918     {
1919         pic_out->img.i_stride[i] = h->fdec->i_stride[i];
1920         pic_out->img.plane[i] = h->fdec->plane[i];
1921     }
1922
1923     /* ---------------------- Update encoder state ------------------------- */
1924
1925     /* update rc */
1926     x264_emms();
1927     if( x264_ratecontrol_end( h, frame_size * 8 ) < 0 )
1928         return -1;
1929
1930     x264_noise_reduction_update( thread_current );
1931
1932     /* ---------------------- Compute/Print statistics --------------------- */
1933     x264_thread_sync_stat( h, h->thread[0] );
1934
1935     /* Slice stat */
1936     h->stat.i_frame_count[h->sh.i_type]++;
1937     h->stat.i_frame_size[h->sh.i_type] += frame_size;
1938     h->stat.f_frame_qp[h->sh.i_type] += h->fdec->f_qp_avg_aq;
1939
1940     for( i = 0; i < X264_MBTYPE_MAX; i++ )
1941         h->stat.i_mb_count[h->sh.i_type][i] += h->stat.frame.i_mb_count[i];
1942     for( i = 0; i < X264_PARTTYPE_MAX; i++ )
1943         h->stat.i_mb_partition[h->sh.i_type][i] += h->stat.frame.i_mb_partition[i];
1944     for( i = 0; i < 2; i++ )
1945         h->stat.i_mb_count_8x8dct[i] += h->stat.frame.i_mb_count_8x8dct[i];
1946     for( i = 0; i < 6; i++ )
1947         h->stat.i_mb_cbp[i] += h->stat.frame.i_mb_cbp[i];
1948     for( i = 0; i < 3; i++ )
1949         for( j = 0; j < 13; j++ )
1950             h->stat.i_mb_pred_mode[i][j] += h->stat.frame.i_mb_pred_mode[i][j];
1951     if( h->sh.i_type != SLICE_TYPE_I )
1952         for( i_list = 0; i_list < 2; i_list++ )
1953             for( i = 0; i < 32; i++ )
1954                 h->stat.i_mb_count_ref[h->sh.i_type][i_list][i] += h->stat.frame.i_mb_count_ref[i_list][i];
1955     if( h->sh.i_type == SLICE_TYPE_P )
1956         h->stat.i_consecutive_bframes[h->fdec->i_frame - h->fref0[0]->i_frame - 1]++;
1957     if( h->sh.i_type == SLICE_TYPE_B )
1958     {
1959         h->stat.i_direct_frames[ h->sh.b_direct_spatial_mv_pred ] ++;
1960         if( h->mb.b_direct_auto_write )
1961         {
1962             //FIXME somewhat arbitrary time constants
1963             if( h->stat.i_direct_score[0] + h->stat.i_direct_score[1] > h->mb.i_mb_count )
1964             {
1965                 for( i = 0; i < 2; i++ )
1966                     h->stat.i_direct_score[i] = h->stat.i_direct_score[i] * 9/10;
1967             }
1968             for( i = 0; i < 2; i++ )
1969                 h->stat.i_direct_score[i] += h->stat.frame.i_direct_score[i];
1970         }
1971     }
1972
1973     psz_message[0] = '\0';
1974     if( h->param.analyse.b_psnr )
1975     {
1976         int64_t ssd[3] = {
1977             h->stat.frame.i_ssd[0],
1978             h->stat.frame.i_ssd[1],
1979             h->stat.frame.i_ssd[2],
1980         };
1981
1982         h->stat.i_ssd_global[h->sh.i_type] += ssd[0] + ssd[1] + ssd[2];
1983         h->stat.f_psnr_average[h->sh.i_type] += x264_psnr( ssd[0] + ssd[1] + ssd[2], 3 * h->param.i_width * h->param.i_height / 2 );
1984         h->stat.f_psnr_mean_y[h->sh.i_type] += x264_psnr( ssd[0], h->param.i_width * h->param.i_height );
1985         h->stat.f_psnr_mean_u[h->sh.i_type] += x264_psnr( ssd[1], h->param.i_width * h->param.i_height / 4 );
1986         h->stat.f_psnr_mean_v[h->sh.i_type] += x264_psnr( ssd[2], h->param.i_width * h->param.i_height / 4 );
1987
1988         snprintf( psz_message, 80, " PSNR Y:%5.2f U:%5.2f V:%5.2f",
1989                   x264_psnr( ssd[0], h->param.i_width * h->param.i_height ),
1990                   x264_psnr( ssd[1], h->param.i_width * h->param.i_height / 4),
1991                   x264_psnr( ssd[2], h->param.i_width * h->param.i_height / 4) );
1992     }
1993
1994     if( h->param.analyse.b_ssim )
1995     {
1996         double ssim_y = h->stat.frame.f_ssim
1997                       / (((h->param.i_width-6)>>2) * ((h->param.i_height-6)>>2));
1998         h->stat.f_ssim_mean_y[h->sh.i_type] += ssim_y;
1999         snprintf( psz_message + strlen(psz_message), 80 - strlen(psz_message),
2000                   " SSIM Y:%.5f", ssim_y );
2001     }
2002     psz_message[79] = '\0';
2003
2004     x264_log( h, X264_LOG_DEBUG,
2005                   "frame=%4d QP=%.2f NAL=%d Slice:%c Poc:%-3d I:%-4d P:%-4d SKIP:%-4d size=%d bytes%s\n",
2006               h->i_frame,
2007               h->fdec->f_qp_avg_aq,
2008               h->i_nal_ref_idc,
2009               h->sh.i_type == SLICE_TYPE_I ? 'I' : (h->sh.i_type == SLICE_TYPE_P ? 'P' : 'B' ),
2010               h->fdec->i_poc,
2011               h->stat.frame.i_mb_count_i,
2012               h->stat.frame.i_mb_count_p,
2013               h->stat.frame.i_mb_count_skip,
2014               frame_size,
2015               psz_message );
2016
2017     // keep stats all in one place
2018     x264_thread_sync_stat( h->thread[0], h );
2019     // for the use of the next frame
2020     x264_thread_sync_stat( thread_current, h );
2021
2022 #ifdef DEBUG_MB_TYPE
2023 {
2024     static const char mb_chars[] = { 'i', 'i', 'I', 'C', 'P', '8', 'S',
2025         'D', '<', 'X', 'B', 'X', '>', 'B', 'B', 'B', 'B', '8', 'S' };
2026     int mb_xy;
2027     for( mb_xy = 0; mb_xy < h->sps->i_mb_width * h->sps->i_mb_height; mb_xy++ )
2028     {
2029         if( h->mb.type[mb_xy] < X264_MBTYPE_MAX && h->mb.type[mb_xy] >= 0 )
2030             fprintf( stderr, "%c ", mb_chars[ h->mb.type[mb_xy] ] );
2031         else
2032             fprintf( stderr, "? " );
2033
2034         if( (mb_xy+1) % h->sps->i_mb_width == 0 )
2035             fprintf( stderr, "\n" );
2036     }
2037 }
2038 #endif
2039
2040     if( h->param.psz_dump_yuv )
2041         x264_frame_dump( h );
2042
2043     return frame_size;
2044 }
2045
2046 static void x264_print_intra( int64_t *i_mb_count, double i_count, int b_print_pcm, char *intra )
2047 {
2048     intra += sprintf( intra, "I16..4%s: %4.1f%% %4.1f%% %4.1f%%",
2049         b_print_pcm ? "..PCM" : "",
2050         i_mb_count[I_16x16]/ i_count,
2051         i_mb_count[I_8x8]  / i_count,
2052         i_mb_count[I_4x4]  / i_count );
2053     if( b_print_pcm )
2054         sprintf( intra, " %4.1f%%", i_mb_count[I_PCM]  / i_count );
2055 }
2056
2057 /****************************************************************************
2058  * x264_encoder_close:
2059  ****************************************************************************/
2060 void    x264_encoder_close  ( x264_t *h )
2061 {
2062     int64_t i_yuv_size = 3 * h->param.i_width * h->param.i_height / 2;
2063     int64_t i_mb_count_size[2][7] = {{0}};
2064     char buf[200];
2065     int i, j, i_list, i_type;
2066     int b_print_pcm = h->stat.i_mb_count[SLICE_TYPE_I][I_PCM]
2067                    || h->stat.i_mb_count[SLICE_TYPE_P][I_PCM]
2068                    || h->stat.i_mb_count[SLICE_TYPE_B][I_PCM];
2069
2070     x264_lookahead_delete( h );
2071
2072     for( i=0; i<h->param.i_threads; i++ )
2073     {
2074         // don't strictly have to wait for the other threads, but it's simpler than canceling them
2075         if( h->thread[i]->b_thread_active )
2076         {
2077             x264_pthread_join( h->thread[i]->thread_handle, NULL );
2078             assert( h->thread[i]->fenc->i_reference_count == 1 );
2079             x264_frame_delete( h->thread[i]->fenc );
2080         }
2081     }
2082
2083     /* Slices used and PSNR */
2084     for( i=0; i<5; i++ )
2085     {
2086         static const int slice_order[] = { SLICE_TYPE_I, SLICE_TYPE_SI, SLICE_TYPE_P, SLICE_TYPE_SP, SLICE_TYPE_B };
2087         static const char *slice_name[] = { "P", "B", "I", "SP", "SI" };
2088         int i_slice = slice_order[i];
2089
2090         if( h->stat.i_frame_count[i_slice] > 0 )
2091         {
2092             const int i_count = h->stat.i_frame_count[i_slice];
2093             if( h->param.analyse.b_psnr )
2094             {
2095                 x264_log( h, X264_LOG_INFO,
2096                           "frame %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",
2097                           slice_name[i_slice],
2098                           i_count,
2099                           h->stat.f_frame_qp[i_slice] / i_count,
2100                           (double)h->stat.i_frame_size[i_slice] / i_count,
2101                           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,
2102                           h->stat.f_psnr_average[i_slice] / i_count,
2103                           x264_psnr( h->stat.i_ssd_global[i_slice], i_count * i_yuv_size ) );
2104             }
2105             else
2106             {
2107                 x264_log( h, X264_LOG_INFO,
2108                           "frame %s:%-5d Avg QP:%5.2f  size:%6.0f\n",
2109                           slice_name[i_slice],
2110                           i_count,
2111                           h->stat.f_frame_qp[i_slice] / i_count,
2112                           (double)h->stat.i_frame_size[i_slice] / i_count );
2113             }
2114         }
2115     }
2116     if( h->param.i_bframe && h->stat.i_frame_count[SLICE_TYPE_P] )
2117     {
2118         char *p = buf;
2119         int den = 0;
2120         // weight by number of frames (including the P-frame) that are in a sequence of N B-frames
2121         for( i=0; i<=h->param.i_bframe; i++ )
2122             den += (i+1) * h->stat.i_consecutive_bframes[i];
2123         for( i=0; i<=h->param.i_bframe; i++ )
2124             p += sprintf( p, " %4.1f%%", 100. * (i+1) * h->stat.i_consecutive_bframes[i] / den );
2125         x264_log( h, X264_LOG_INFO, "consecutive B-frames:%s\n", buf );
2126     }
2127
2128     for( i_type = 0; i_type < 2; i_type++ )
2129         for( i = 0; i < X264_PARTTYPE_MAX; i++ )
2130         {
2131             if( i == D_DIRECT_8x8 ) continue; /* direct is counted as its own type */
2132             i_mb_count_size[i_type][x264_mb_partition_pixel_table[i]] += h->stat.i_mb_partition[i_type][i];
2133         }
2134
2135     /* MB types used */
2136     if( h->stat.i_frame_count[SLICE_TYPE_I] > 0 )
2137     {
2138         int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_I];
2139         double i_count = h->stat.i_frame_count[SLICE_TYPE_I] * h->mb.i_mb_count / 100.0;
2140         x264_print_intra( i_mb_count, i_count, b_print_pcm, buf );
2141         x264_log( h, X264_LOG_INFO, "mb I  %s\n", buf );
2142     }
2143     if( h->stat.i_frame_count[SLICE_TYPE_P] > 0 )
2144     {
2145         int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_P];
2146         double i_count = h->stat.i_frame_count[SLICE_TYPE_P] * h->mb.i_mb_count / 100.0;
2147         int64_t *i_mb_size = i_mb_count_size[SLICE_TYPE_P];
2148         x264_print_intra( i_mb_count, i_count, b_print_pcm, buf );
2149         x264_log( h, X264_LOG_INFO,
2150                   "mb P  %s  P16..4: %4.1f%% %4.1f%% %4.1f%% %4.1f%% %4.1f%%    skip:%4.1f%%\n",
2151                   buf,
2152                   i_mb_size[PIXEL_16x16] / (i_count*4),
2153                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
2154                   i_mb_size[PIXEL_8x8] / (i_count*4),
2155                   (i_mb_size[PIXEL_8x4] + i_mb_size[PIXEL_4x8]) / (i_count*4),
2156                   i_mb_size[PIXEL_4x4] / (i_count*4),
2157                   i_mb_count[P_SKIP] / i_count );
2158     }
2159     if( h->stat.i_frame_count[SLICE_TYPE_B] > 0 )
2160     {
2161         int64_t *i_mb_count = h->stat.i_mb_count[SLICE_TYPE_B];
2162         double i_count = h->stat.i_frame_count[SLICE_TYPE_B] * h->mb.i_mb_count / 100.0;
2163         double i_mb_list_count;
2164         int64_t *i_mb_size = i_mb_count_size[SLICE_TYPE_B];
2165         int64_t list_count[3] = {0}; /* 0 == L0, 1 == L1, 2 == BI */
2166         x264_print_intra( i_mb_count, i_count, b_print_pcm, buf );
2167         for( i = 0; i < X264_PARTTYPE_MAX; i++ )
2168             for( j = 0; j < 2; j++ )
2169             {
2170                 int l0 = x264_mb_type_list_table[i][0][j];
2171                 int l1 = x264_mb_type_list_table[i][1][j];
2172                 if( l0 || l1 )
2173                     list_count[l1+l0*l1] += h->stat.i_mb_count[SLICE_TYPE_B][i] * 2;
2174             }
2175         list_count[0] += h->stat.i_mb_partition[SLICE_TYPE_B][D_L0_8x8];
2176         list_count[1] += h->stat.i_mb_partition[SLICE_TYPE_B][D_L1_8x8];
2177         list_count[2] += h->stat.i_mb_partition[SLICE_TYPE_B][D_BI_8x8];
2178         i_mb_count[B_DIRECT] += (h->stat.i_mb_partition[SLICE_TYPE_B][D_DIRECT_8x8]+2)/4;
2179         i_mb_list_count = (list_count[0] + list_count[1] + list_count[2]) / 100.0;
2180         x264_log( h, X264_LOG_INFO,
2181                   "mb B  %s  B16..8: %4.1f%% %4.1f%% %4.1f%%  direct:%4.1f%%  skip:%4.1f%%  L0:%4.1f%% L1:%4.1f%% BI:%4.1f%%\n",
2182                   buf,
2183                   i_mb_size[PIXEL_16x16] / (i_count*4),
2184                   (i_mb_size[PIXEL_16x8] + i_mb_size[PIXEL_8x16]) / (i_count*4),
2185                   i_mb_size[PIXEL_8x8] / (i_count*4),
2186                   i_mb_count[B_DIRECT] / i_count,
2187                   i_mb_count[B_SKIP]   / i_count,
2188                   list_count[0] / i_mb_list_count,
2189                   list_count[1] / i_mb_list_count,
2190                   list_count[2] / i_mb_list_count );
2191     }
2192
2193     x264_ratecontrol_summary( h );
2194
2195     if( h->stat.i_frame_count[SLICE_TYPE_I] + h->stat.i_frame_count[SLICE_TYPE_P] + h->stat.i_frame_count[SLICE_TYPE_B] > 0 )
2196     {
2197 #define SUM3(p) (p[SLICE_TYPE_I] + p[SLICE_TYPE_P] + p[SLICE_TYPE_B])
2198 #define SUM3b(p,o) (p[SLICE_TYPE_I][o] + p[SLICE_TYPE_P][o] + p[SLICE_TYPE_B][o])
2199         int64_t i_i8x8 = SUM3b( h->stat.i_mb_count, I_8x8 );
2200         int64_t i_intra = i_i8x8 + SUM3b( h->stat.i_mb_count, I_4x4 )
2201                                  + SUM3b( h->stat.i_mb_count, I_16x16 );
2202         int64_t i_all_intra = i_intra + SUM3b( h->stat.i_mb_count, I_PCM);
2203         const int i_count = h->stat.i_frame_count[SLICE_TYPE_I] +
2204                             h->stat.i_frame_count[SLICE_TYPE_P] +
2205                             h->stat.i_frame_count[SLICE_TYPE_B];
2206         int64_t i_mb_count = i_count * h->mb.i_mb_count;
2207         float fps = (float) h->param.i_fps_num / h->param.i_fps_den;
2208         float f_bitrate = fps * SUM3(h->stat.i_frame_size) / i_count / 125;
2209
2210         if( h->pps->b_transform_8x8_mode )
2211         {
2212             buf[0] = 0;
2213             if( h->stat.i_mb_count_8x8dct[0] )
2214                 sprintf( buf, " inter:%.1f%%", 100. * h->stat.i_mb_count_8x8dct[1] / h->stat.i_mb_count_8x8dct[0] );
2215             x264_log( h, X264_LOG_INFO, "8x8 transform intra:%.1f%%%s\n", 100. * i_i8x8 / i_intra, buf );
2216         }
2217
2218         if( h->param.analyse.i_direct_mv_pred == X264_DIRECT_PRED_AUTO
2219             && h->stat.i_frame_count[SLICE_TYPE_B] )
2220         {
2221             x264_log( h, X264_LOG_INFO, "direct mvs  spatial:%.1f%% temporal:%.1f%%\n",
2222                       h->stat.i_direct_frames[1] * 100. / h->stat.i_frame_count[SLICE_TYPE_B],
2223                       h->stat.i_direct_frames[0] * 100. / h->stat.i_frame_count[SLICE_TYPE_B] );
2224         }
2225
2226         buf[0] = 0;
2227         if( i_mb_count != i_all_intra )
2228             sprintf( buf, " inter: %.1f%% %.1f%% %.1f%%",
2229                      h->stat.i_mb_cbp[1] * 100.0 / ((i_mb_count - i_all_intra)*4),
2230                      h->stat.i_mb_cbp[3] * 100.0 / ((i_mb_count - i_all_intra)  ),
2231                      h->stat.i_mb_cbp[5] * 100.0 / ((i_mb_count - i_all_intra)) );
2232         x264_log( h, X264_LOG_INFO, "coded y,uvDC,uvAC intra: %.1f%% %.1f%% %.1f%%%s\n",
2233                   h->stat.i_mb_cbp[0] * 100.0 / (i_all_intra*4),
2234                   h->stat.i_mb_cbp[2] * 100.0 / (i_all_intra  ),
2235                   h->stat.i_mb_cbp[4] * 100.0 / (i_all_intra  ), buf );
2236
2237         int64_t fixed_pred_modes[3][9] = {{0}};
2238         int64_t sum_pred_modes[3] = {0};
2239         for( i = 0; i <= I_PRED_16x16_DC_128; i++ )
2240         {
2241             fixed_pred_modes[0][x264_mb_pred_mode16x16_fix[i]] += h->stat.i_mb_pred_mode[0][i];
2242             sum_pred_modes[0] += h->stat.i_mb_pred_mode[0][i];
2243         }
2244         if( sum_pred_modes[0] )
2245             x264_log( h, X264_LOG_INFO, "i16 v,h,dc,p: %2.0f%% %2.0f%% %2.0f%% %2.0f%%\n",
2246                       fixed_pred_modes[0][0] * 100.0 / sum_pred_modes[0],
2247                       fixed_pred_modes[0][1] * 100.0 / sum_pred_modes[0],
2248                       fixed_pred_modes[0][2] * 100.0 / sum_pred_modes[0],
2249                       fixed_pred_modes[0][3] * 100.0 / sum_pred_modes[0] );
2250         for( i = 1; i <= 2; i++ )
2251         {
2252             for( j = 0; j <= I_PRED_8x8_DC_128; j++ )
2253             {
2254                 fixed_pred_modes[i][x264_mb_pred_mode4x4_fix(j)] += h->stat.i_mb_pred_mode[i][j];
2255                 sum_pred_modes[i] += h->stat.i_mb_pred_mode[i][j];
2256             }
2257             if( sum_pred_modes[i] )
2258                 x264_log( h, X264_LOG_INFO, "i%d v,h,dc,ddl,ddr,vr,hd,vl,hu: %2.0f%% %2.0f%% %2.0f%% %2.0f%% %2.0f%% %2.0f%% %2.0f%% %2.0f%% %2.0f%%\n", (3-i)*4,
2259                           fixed_pred_modes[i][0] * 100.0 / sum_pred_modes[i],
2260                           fixed_pred_modes[i][1] * 100.0 / sum_pred_modes[i],
2261                           fixed_pred_modes[i][2] * 100.0 / sum_pred_modes[i],
2262                           fixed_pred_modes[i][3] * 100.0 / sum_pred_modes[i],
2263                           fixed_pred_modes[i][4] * 100.0 / sum_pred_modes[i],
2264                           fixed_pred_modes[i][5] * 100.0 / sum_pred_modes[i],
2265                           fixed_pred_modes[i][6] * 100.0 / sum_pred_modes[i],
2266                           fixed_pred_modes[i][7] * 100.0 / sum_pred_modes[i],
2267                           fixed_pred_modes[i][8] * 100.0 / sum_pred_modes[i] );
2268         }
2269
2270         for( i_list = 0; i_list < 2; i_list++ )
2271         {
2272             int i_slice;
2273             for( i_slice = 0; i_slice < 2; i_slice++ )
2274             {
2275                 char *p = buf;
2276                 int64_t i_den = 0;
2277                 int i_max = 0;
2278                 for( i = 0; i < 32; i++ )
2279                     if( h->stat.i_mb_count_ref[i_slice][i_list][i] )
2280                     {
2281                         i_den += h->stat.i_mb_count_ref[i_slice][i_list][i];
2282                         i_max = i;
2283                     }
2284                 if( i_max == 0 )
2285                     continue;
2286                 for( i = 0; i <= i_max; i++ )
2287                     p += sprintf( p, " %4.1f%%", 100. * h->stat.i_mb_count_ref[i_slice][i_list][i] / i_den );
2288                 x264_log( h, X264_LOG_INFO, "ref %c L%d:%s\n", "PB"[i_slice], i_list, buf );
2289             }
2290         }
2291
2292         if( h->param.analyse.b_ssim )
2293         {
2294             x264_log( h, X264_LOG_INFO,
2295                       "SSIM Mean Y:%.7f\n",
2296                       SUM3( h->stat.f_ssim_mean_y ) / i_count );
2297         }
2298         if( h->param.analyse.b_psnr )
2299         {
2300             x264_log( h, X264_LOG_INFO,
2301                       "PSNR Mean Y:%6.3f U:%6.3f V:%6.3f Avg:%6.3f Global:%6.3f kb/s:%.2f\n",
2302                       SUM3( h->stat.f_psnr_mean_y ) / i_count,
2303                       SUM3( h->stat.f_psnr_mean_u ) / i_count,
2304                       SUM3( h->stat.f_psnr_mean_v ) / i_count,
2305                       SUM3( h->stat.f_psnr_average ) / i_count,
2306                       x264_psnr( SUM3( h->stat.i_ssd_global ), i_count * i_yuv_size ),
2307                       f_bitrate );
2308         }
2309         else
2310             x264_log( h, X264_LOG_INFO, "kb/s:%.2f\n", f_bitrate );
2311     }
2312
2313     /* rc */
2314     x264_ratecontrol_delete( h );
2315
2316     /* param */
2317     if( h->param.rc.psz_stat_out )
2318         free( h->param.rc.psz_stat_out );
2319     if( h->param.rc.psz_stat_in )
2320         free( h->param.rc.psz_stat_in );
2321
2322     x264_cqm_delete( h );
2323
2324     x264_analyse_free_costs( h );
2325
2326     if( h->param.i_threads > 1)
2327         h = h->thread[ h->i_thread_phase % h->param.i_threads ];
2328
2329     /* frames */
2330     x264_frame_delete_list( h->frames.unused[0] );
2331     x264_frame_delete_list( h->frames.unused[1] );
2332     x264_frame_delete_list( h->frames.current );
2333
2334     h = h->thread[0];
2335
2336     for( i = h->param.i_threads - 1; i >= 0; i-- )
2337     {
2338         x264_frame_t **frame;
2339
2340         for( frame = h->thread[i]->frames.reference; *frame; frame++ )
2341         {
2342             assert( (*frame)->i_reference_count > 0 );
2343             (*frame)->i_reference_count--;
2344             if( (*frame)->i_reference_count == 0 )
2345                 x264_frame_delete( *frame );
2346         }
2347         frame = &h->thread[i]->fdec;
2348         assert( (*frame)->i_reference_count > 0 );
2349         (*frame)->i_reference_count--;
2350         if( (*frame)->i_reference_count == 0 )
2351             x264_frame_delete( *frame );
2352
2353         x264_macroblock_cache_end( h->thread[i] );
2354         x264_free( h->thread[i]->out.p_bitstream );
2355         x264_free( h->thread[i]->out.nal);
2356         x264_free( h->thread[i] );
2357     }
2358 }
2359
2360 /****************************************************************************
2361  * x264_encoder_delayed_frames:
2362  ****************************************************************************/
2363 int x264_encoder_delayed_frames( x264_t *h )
2364 {
2365     int delayed_frames = 0;
2366     int i;
2367     for( i=0; i<h->param.i_threads; i++ )
2368         delayed_frames += h->thread[i]->b_thread_active;
2369     h = h->thread[ h->i_thread_phase % h->param.i_threads ];
2370     for( i=0; h->frames.current[i]; i++ )
2371         delayed_frames++;
2372     x264_pthread_mutex_lock( &h->lookahead->ofbuf.mutex );
2373     x264_pthread_mutex_lock( &h->lookahead->ifbuf.mutex );
2374     x264_pthread_mutex_lock( &h->lookahead->next.mutex );
2375     delayed_frames += h->lookahead->ifbuf.i_size + h->lookahead->next.i_size + h->lookahead->ofbuf.i_size;
2376     x264_pthread_mutex_unlock( &h->lookahead->next.mutex );
2377     x264_pthread_mutex_unlock( &h->lookahead->ifbuf.mutex );
2378     x264_pthread_mutex_unlock( &h->lookahead->ofbuf.mutex );
2379     return delayed_frames;
2380 }