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

   1 /*****************************************************************************
   2  * macroblock.c: h264 encoder library
   3  *****************************************************************************
   4  * Copyright (C) 2003 Laurent Aimar
   5  * $Id: macroblock.c,v 1.1 2004/06/03 19:27:08 fenrir Exp $
   6  *
   7  * Authors: Laurent Aimar <fenrir@via.ecp.fr>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, write to the Free Software
  21  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
  22  *****************************************************************************/
  23
  24 #include "common/common.h"
  25 #include "macroblock.h"
  26
  27
  28 #define ZIG(i,y,x) level[i] = dct[x][y];
  29 static inline void zigzag_scan_2x2_dc( int16_t level[4], int16_t dct[2][2] )
  30 {
  31     ZIG(0,0,0)
  32     ZIG(1,0,1)
  33     ZIG(2,1,0)
  34     ZIG(3,1,1)
  35 }
  36 #undef ZIG
  37
  38 /* (ref: JVT-B118)
  39  * x264_mb_decimate_score: given dct coeffs it returns a score to see if we could empty this dct coeffs
  40  * to 0 (low score means set it to null)
  41  * Used in inter macroblock (luma and chroma)
  42  *  luma: for a 8x8 block: if score < 4 -> null
  43  *        for the complete mb: if score < 6 -> null
  44  *  chroma: for the complete mb: if score < 7 -> null
  45  */
  46 static int x264_mb_decimate_score( int16_t *dct, int i_max )
  47 {
  48     static const int i_ds_table4[16] = {
  49         3,2,2,1,1,1,0,0,0,0,0,0,0,0,0,0 };
  50     static const int i_ds_table8[64] = {
  51         3,3,3,3,2,2,2,2,2,2,2,2,1,1,1,1,
  52         1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,
  53         0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
  54         0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 };
  55
  56     const int *ds_table = (i_max == 64) ? i_ds_table8 : i_ds_table4;
  57     int i_score = 0;
  58     int idx = i_max - 1;
  59
  60     while( idx >= 0 && dct[idx] == 0 )
  61         idx--;
  62
  63     while( idx >= 0 )
  64     {
  65         int i_run;
  66
  67         if( abs( dct[idx--] ) > 1 )
  68             return 9;
  69
  70         i_run = 0;
  71         while( idx >= 0 && dct[idx] == 0 )
  72         {
  73             idx--;
  74             i_run++;
  75         }
  76         i_score += ds_table[i_run];
  77     }
  78
  79     return i_score;
  80 }
  81
  82 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qscale )
  83 {
  84     int x = 4 * block_idx_x[idx];
  85     int y = 4 * block_idx_y[idx];
  86     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
  87     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
  88     DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
  89
  90     if( h->mb.b_lossless )
  91     {
  92         h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
  93         return;
  94     }
  95
  96     h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
  97
  98     if( h->mb.b_trellis )
  99         x264_quant_4x4_trellis( h, dct4x4, CQM_4IY, i_qscale, DCT_LUMA_4x4, 1 );
 100     else
 101         h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4IY][i_qscale], h->quant4_bias[CQM_4IY][i_qscale] );
 102
 103     h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4 );
 104     h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qscale );
 105
 106     /* output samples to fdec */
 107     h->dctf.add4x4_idct( p_dst, dct4x4 );
 108 }
 109
 110 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qscale )
 111 {
 112     int x = 8 * (idx&1);
 113     int y = 8 * (idx>>1);
 114     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
 115     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
 116     DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 117
 118     h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
 119
 120     if( h->mb.b_trellis )
 121         x264_quant_8x8_trellis( h, dct8x8, CQM_8IY, i_qscale, 1 );
 122     else
 123         h->quantf.quant_8x8( dct8x8, h->quant8_mf[CQM_8IY][i_qscale], h->quant8_bias[CQM_8IY][i_qscale] );
 124
 125     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
 126     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qscale );
 127     h->dctf.add8x8_idct8( p_dst, dct8x8 );
 128 }
 129
 130 static void x264_mb_encode_i16x16( x264_t *h, int i_qscale )
 131 {
 132     uint8_t  *p_src = h->mb.pic.p_fenc[0];
 133     uint8_t  *p_dst = h->mb.pic.p_fdec[0];
 134
 135     DECLARE_ALIGNED_16( int16_t dct4x4[16+1][4][4] );
 136
 137     int i;
 138
 139     if( h->mb.b_lossless )
 140     {
 141         for( i = 0; i < 16; i++ )
 142         {
 143             int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 144             int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 145             h->zigzagf.sub_4x4( h->dct.luma4x4[i], p_src+oe, p_dst+od );
 146             dct4x4[0][block_idx_x[i]][block_idx_y[i]] = h->dct.luma4x4[i][0];
 147         }
 148         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct4x4[0] );
 149         return;
 150     }
 151
 152     h->dctf.sub16x16_dct( &dct4x4[1], p_src, p_dst );
 153     for( i = 0; i < 16; i++ )
 154     {
 155         /* copy dc coeff */
 156         dct4x4[0][block_idx_y[i]][block_idx_x[i]] = dct4x4[1+i][0][0];
 157
 158         /* quant/scan/dequant */
 159         if( h->mb.b_trellis )
 160             x264_quant_4x4_trellis( h, dct4x4[1+i], CQM_4IY, i_qscale, DCT_LUMA_AC, 1 );
 161         else
 162             h->quantf.quant_4x4( dct4x4[1+i], h->quant4_mf[CQM_4IY][i_qscale], h->quant4_bias[CQM_4IY][i_qscale] );
 163
 164         h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[1+i] );
 165         h->quantf.dequant_4x4( dct4x4[1+i], h->dequant4_mf[CQM_4IY], i_qscale );
 166     }
 167
 168     h->dctf.dct4x4dc( dct4x4[0] );
 169     h->quantf.quant_4x4_dc( dct4x4[0], h->quant4_mf[CQM_4IY][i_qscale][0]>>1, h->quant4_bias[CQM_4IY][i_qscale][0]<<1 );
 170     h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct4x4[0] );
 171
 172     /* output samples to fdec */
 173     h->dctf.idct4x4dc( dct4x4[0] );
 174     x264_mb_dequant_4x4_dc( dct4x4[0], h->dequant4_mf[CQM_4IY], i_qscale );  /* XXX not inversed */
 175
 176     /* calculate dct coeffs */
 177     for( i = 0; i < 16; i++ )
 178     {
 179         /* copy dc coeff */
 180         dct4x4[1+i][0][0] = dct4x4[0][block_idx_y[i]][block_idx_x[i]];
 181     }
 182     /* put pixels to fdec */
 183     h->dctf.add16x16_idct( p_dst, &dct4x4[1] );
 184 }
 185
 186 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qscale )
 187 {
 188     int i, ch;
 189     int b_decimate = b_inter && (h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate);
 190
 191     for( ch = 0; ch < 2; ch++ )
 192     {
 193         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 194         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 195         int i_decimate_score = 0;
 196
 197         DECLARE_ALIGNED_16( int16_t dct2x2[2][2]  );
 198         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 199
 200         if( h->mb.b_lossless )
 201         {
 202             for( i = 0; i < 4; i++ )
 203             {
 204                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 205                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 206                 h->zigzagf.sub_4x4( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od );
 207                 h->dct.chroma_dc[ch][i] = h->dct.luma4x4[16+i+ch*4][0];
 208             }
 209             continue;
 210         }
 211
 212         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 213         /* calculate dct coeffs */
 214         for( i = 0; i < 4; i++ )
 215         {
 216             /* copy dc coeff */
 217             dct2x2[block_idx_y[i]][block_idx_x[i]] = dct4x4[i][0][0];
 218
 219             /* no trellis; it doesn't seem to help chroma noticeably */
 220             h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[CQM_4IC+b_inter][i_qscale], h->quant4_bias[CQM_4IC+b_inter][i_qscale] );
 221             h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 222
 223             if( b_decimate )
 224             {
 225                 i_decimate_score += x264_mb_decimate_score( h->dct.luma4x4[16+i+ch*4]+1, 15 );
 226             }
 227         }
 228
 229         h->dctf.dct2x2dc( dct2x2 );
 230         h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4IC+b_inter][i_qscale][0]>>1, h->quant4_bias[CQM_4IC+b_inter][i_qscale][0]<<1 );
 231         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 232
 233         /* output samples to fdec */
 234         h->dctf.idct2x2dc( dct2x2 );
 235         x264_mb_dequant_2x2_dc( dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qscale );  /* XXX not inversed */
 236
 237         if( b_decimate && i_decimate_score < 7 )
 238         {
 239             /* Near null chroma 8x8 block so make it null (bits saving) */
 240             memset( &h->dct.luma4x4[16+ch*4], 0, 4 * sizeof( *h->dct.luma4x4 ) );
 241             if( !array_non_zero( dct2x2 ) )
 242                 continue;
 243             memset( dct4x4, 0, sizeof( dct4x4 ) );
 244         }
 245         else
 246         {
 247             for( i = 0; i < 4; i++ )
 248                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qscale );
 249         }
 250
 251         for( i = 0; i < 4; i++ )
 252             dct4x4[i][0][0] = dct2x2[0][i];
 253         h->dctf.add8x8_idct( p_dst, dct4x4 );
 254     }
 255
 256     /* coded block pattern */
 257     h->mb.i_cbp_chroma = 0;
 258     for( i = 0; i < 8; i++ )
 259     {
 260         int nz = array_non_zero_count( h->dct.luma4x4[16+i]+1, 15 );
 261         h->mb.cache.non_zero_count[x264_scan8[16+i]] = nz;
 262         h->mb.i_cbp_chroma |= nz;
 263     }
 264     if( h->mb.i_cbp_chroma )
 265         h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
 266     else if( array_non_zero( h->dct.chroma_dc ) )
 267         h->mb.i_cbp_chroma = 1;    /* dc only */
 268 }
 269
 270 static void x264_macroblock_encode_skip( x264_t *h )
 271 {
 272     int i;
 273     h->mb.i_cbp_luma = 0x00;
 274     h->mb.i_cbp_chroma = 0x00;
 275
 276     for( i = 0; i < 16+8; i++ )
 277     {
 278         h->mb.cache.non_zero_count[x264_scan8[i]] = 0;
 279     }
 280
 281     /* store cbp */
 282     h->mb.cbp[h->mb.i_mb_xy] = 0;
 283 }
 284
 285 /*****************************************************************************
 286  * x264_macroblock_encode_pskip:
 287  *  Encode an already marked skip block
 288  *****************************************************************************/
 289 void x264_macroblock_encode_pskip( x264_t *h )
 290 {
 291     const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 292                                 h->mb.mv_min[0], h->mb.mv_max[0] );
 293     const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 294                                 h->mb.mv_min[1], h->mb.mv_max[1] );
 295
 296     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 297     if( !h->mb.b_skip_pbskip_mc )
 298     {
 299         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 300                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 301                        mvx, mvy, 16, 16 );
 302
 303         h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 304                          h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 305                          mvx, mvy, 8, 8 );
 306
 307         h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 308                          h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 309                          mvx, mvy, 8, 8 );
 310     }
 311
 312     x264_macroblock_encode_skip( h );
 313 }
 314
 315 /*****************************************************************************
 316  * x264_macroblock_encode:
 317  *****************************************************************************/
 318 void x264_macroblock_encode( x264_t *h )
 319 {
 320     int i_cbp_dc = 0;
 321     int i_qp = h->mb.i_qp;
 322     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 323     int b_force_no_skip = 0;
 324     int i;
 325
 326     if( h->sh.b_mbaff
 327         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 328         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 329     {
 330         /* The first skip is predicted to be a frame mb pair.
 331          * We don't yet support the aff part of mbaff, so force it to non-skip
 332          * so that we can pick the aff flag. */
 333         b_force_no_skip = 1;
 334         if( IS_SKIP(h->mb.i_type) )
 335         {
 336             if( h->mb.i_type == P_SKIP )
 337                 h->mb.i_type = P_L0;
 338             else if( h->mb.i_type == B_SKIP )
 339                 h->mb.i_type = B_DIRECT;
 340         }
 341     }
 342
 343     if( h->mb.i_type == P_SKIP )
 344     {
 345         /* A bit special */
 346         x264_macroblock_encode_pskip( h );
 347         return;
 348     }
 349     if( h->mb.i_type == B_SKIP )
 350     {
 351         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 352         if( !h->mb.b_skip_pbskip_mc )
 353             x264_mb_mc( h );
 354         x264_macroblock_encode_skip( h );
 355         return;
 356     }
 357
 358     if( h->mb.i_type == I_16x16 )
 359     {
 360         const int i_mode = h->mb.i_intra16x16_pred_mode;
 361         h->mb.b_transform_8x8 = 0;
 362         /* do the right prediction */
 363         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 364
 365         /* encode the 16x16 macroblock */
 366         x264_mb_encode_i16x16( h, i_qp );
 367     }
 368     else if( h->mb.i_type == I_8x8 )
 369     {
 370         DECLARE_ALIGNED_16( uint8_t edge[33] );
 371         h->mb.b_transform_8x8 = 1;
 372         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 373         if( h->mb.i_skip_intra )
 374         {
 375             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 376             /* In RD mode, restore the now-overwritten DCT data. */
 377             if( h->mb.i_skip_intra == 2 )
 378                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 379         }
 380         for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 381         {
 382             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 383             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 384
 385             x264_predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 386             h->predict_8x8[i_mode]( p_dst, edge );
 387             x264_mb_encode_i8x8( h, i, i_qp );
 388         }
 389     }
 390     else if( h->mb.i_type == I_4x4 )
 391     {
 392         h->mb.b_transform_8x8 = 0;
 393         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 394         if( h->mb.i_skip_intra )
 395         {
 396             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 397             /* In RD mode, restore the now-overwritten DCT data. */
 398             if( h->mb.i_skip_intra == 2 )
 399                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 400         }
 401         for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 402         {
 403             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][4 * block_idx_x[i] + 4 * block_idx_y[i] * FDEC_STRIDE];
 404             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 405
 406             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 407                 /* emulate missing topright samples */
 408                 *(uint32_t*) &p_dst[4-FDEC_STRIDE] = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 409
 410             h->predict_4x4[i_mode]( p_dst );
 411             x264_mb_encode_i4x4( h, i, i_qp );
 412         }
 413     }
 414     else    /* Inter MB */
 415     {
 416         int i8x8, i4x4, idx;
 417         int i_decimate_mb = 0;
 418
 419         /* Motion compensation */
 420         x264_mb_mc( h );
 421
 422         if( h->mb.b_lossless )
 423         {
 424             for( i4x4 = 0; i4x4 < 16; i4x4++ )
 425             {
 426                 int x = 4*block_idx_x[i4x4];
 427                 int y = 4*block_idx_y[i4x4];
 428                 h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 429                                     h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
 430                                     h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
 431             }
 432         }
 433         else if( h->mb.b_transform_8x8 )
 434         {
 435             DECLARE_ALIGNED_16( int16_t dct8x8[4][8][8] );
 436             int nnz8x8[4] = {1,1,1,1};
 437             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 438             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 439
 440             for( idx = 0; idx < 4; idx++ )
 441             {
 442                 if( h->mb.b_noise_reduction )
 443                     x264_denoise_dct( h, (int16_t*)dct8x8[idx] );
 444                 if( h->mb.b_trellis )
 445                     x264_quant_8x8_trellis( h, dct8x8[idx], CQM_8PY, i_qp, 0 );
 446                 else
 447                     h->quantf.quant_8x8( dct8x8[idx], h->quant8_mf[CQM_8PY][i_qp], h->quant8_bias[CQM_8PY][i_qp] );
 448
 449                 h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 450
 451                 if( b_decimate )
 452                 {
 453                     int i_decimate_8x8 = x264_mb_decimate_score( h->dct.luma8x8[idx], 64 );
 454                     i_decimate_mb += i_decimate_8x8;
 455                     if( i_decimate_8x8 < 4 )
 456                     {
 457                         memset( h->dct.luma8x8[idx], 0, sizeof( h->dct.luma8x8[idx] ) );
 458                         memset( dct8x8[idx], 0, sizeof( dct8x8[idx] ) );
 459                         nnz8x8[idx] = 0;
 460                     }
 461                 }
 462                 else
 463                     nnz8x8[idx] = array_non_zero( dct8x8[idx] );
 464             }
 465
 466             if( i_decimate_mb < 6 && b_decimate )
 467                 memset( h->dct.luma8x8, 0, sizeof( h->dct.luma8x8 ) );
 468             else
 469             {
 470                 for( idx = 0; idx < 4; idx++ )
 471                     if( nnz8x8[idx] )
 472                     {
 473                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 474                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 475                     }
 476             }
 477         }
 478         else
 479         {
 480             DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 481             int nnz8x8[4] = {1,1,1,1};
 482             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 483
 484             for( i8x8 = 0; i8x8 < 4; i8x8++ )
 485             {
 486                 int i_decimate_8x8;
 487
 488                 /* encode one 4x4 block */
 489                 i_decimate_8x8 = 0;
 490                 for( i4x4 = 0; i4x4 < 4; i4x4++ )
 491                 {
 492                     idx = i8x8 * 4 + i4x4;
 493
 494                     if( h->mb.b_noise_reduction )
 495                         x264_denoise_dct( h, (int16_t*)dct4x4[idx] );
 496                     if( h->mb.b_trellis )
 497                         x264_quant_4x4_trellis( h, dct4x4[idx], CQM_4PY, i_qp, DCT_LUMA_4x4, 0 );
 498                     else
 499                         h->quantf.quant_4x4( dct4x4[idx], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 500
 501                     h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 502
 503                     if( b_decimate )
 504                         i_decimate_8x8 += x264_mb_decimate_score( h->dct.luma4x4[idx], 16 );
 505                 }
 506
 507                 /* decimate this 8x8 block */
 508                 i_decimate_mb += i_decimate_8x8;
 509                 if( i_decimate_8x8 < 4 && b_decimate )
 510                 {
 511                     memset( &dct4x4[i8x8*4], 0, 4 * sizeof( *dct4x4 ) );
 512                     memset( &h->dct.luma4x4[i8x8*4], 0, 4 * sizeof( *h->dct.luma4x4 ) );
 513                     nnz8x8[i8x8] = 0;
 514                 }
 515             }
 516
 517             if( i_decimate_mb < 6 && b_decimate )
 518                 memset( h->dct.luma4x4, 0, 16 * sizeof( *h->dct.luma4x4 ) );
 519             else
 520             {
 521                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 522                     if( nnz8x8[i8x8] )
 523                     {
 524                         for( i = 0; i < 4; i++ )
 525                             h->quantf.dequant_4x4( dct4x4[i8x8*4+i], h->dequant4_mf[CQM_4PY], i_qp );
 526                         h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 527                     }
 528             }
 529         }
 530     }
 531
 532     /* encode chroma */
 533     if( IS_INTRA( h->mb.i_type ) )
 534     {
 535         const int i_mode = h->mb.i_chroma_pred_mode;
 536         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 537         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 538     }
 539
 540     /* encode the 8x8 blocks */
 541     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 542
 543     /* coded block pattern and non_zero_count */
 544     h->mb.i_cbp_luma = 0x00;
 545     if( h->mb.i_type == I_16x16 )
 546     {
 547         for( i = 0; i < 16; i++ )
 548         {
 549             const int nz = array_non_zero_count( h->dct.luma4x4[i]+1, 15 );
 550             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 551             if( nz > 0 )
 552                 h->mb.i_cbp_luma = 0x0f;
 553         }
 554     }
 555     else if( h->mb.b_transform_8x8 )
 556     {
 557         /* coded_block_flag is enough for CABAC.
 558          * the full non_zero_count is done only in CAVLC. */
 559         for( i = 0; i < 4; i++ )
 560         {
 561             const int nz = array_non_zero( h->dct.luma8x8[i] );
 562             int j;
 563             for( j = 0; j < 4; j++ )
 564                 h->mb.cache.non_zero_count[x264_scan8[4*i+j]] = nz;
 565             if( nz > 0 )
 566                 h->mb.i_cbp_luma |= 1 << i;
 567         }
 568     }
 569     else
 570     {
 571         for( i = 0; i < 16; i++ )
 572         {
 573             const int nz = array_non_zero_count( h->dct.luma4x4[i], 16 );
 574             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 575             if( nz > 0 )
 576                 h->mb.i_cbp_luma |= 1 << (i/4);
 577         }
 578     }
 579
 580     if( h->param.b_cabac )
 581     {
 582         i_cbp_dc = ( h->mb.i_type == I_16x16 && array_non_zero( h->dct.luma16x16_dc ) )
 583                  | array_non_zero( h->dct.chroma_dc[0] ) << 1
 584                  | array_non_zero( h->dct.chroma_dc[1] ) << 2;
 585     }
 586
 587     /* store cbp */
 588     h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
 589
 590     /* Check for P_SKIP
 591      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 592      *      (if multiple mv give same result)*/
 593     if( !b_force_no_skip )
 594     {
 595         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 596             h->mb.i_cbp_luma == 0x00 && h->mb.i_cbp_chroma == 0x00 &&
 597             h->mb.cache.mv[0][x264_scan8[0]][0] == h->mb.cache.pskip_mv[0] &&
 598             h->mb.cache.mv[0][x264_scan8[0]][1] == h->mb.cache.pskip_mv[1] &&
 599             h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 600         {
 601             h->mb.i_type = P_SKIP;
 602         }
 603
 604         /* Check for B_SKIP */
 605         if( h->mb.i_type == B_DIRECT &&
 606             h->mb.i_cbp_luma == 0x00 && h->mb.i_cbp_chroma== 0x00 )
 607         {
 608             h->mb.i_type = B_SKIP;
 609         }
 610     }
 611 }
 612
 613 /*****************************************************************************
 614  * x264_macroblock_probe_skip:
 615  *  Check if the current MB could be encoded as a [PB]_SKIP (it supposes you use
 616  *  the previous QP
 617  *****************************************************************************/
 618 int x264_macroblock_probe_skip( x264_t *h, const int b_bidir )
 619 {
 620     DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 621     DECLARE_ALIGNED_16( int16_t dct2x2[2][2] );
 622     DECLARE_ALIGNED_16( int16_t dctscan[16] );
 623
 624     int i_qp = h->mb.i_qp;
 625     int mvp[2];
 626     int ch;
 627
 628     int i8x8, i4x4;
 629     int i_decimate_mb;
 630
 631     if( !b_bidir )
 632     {
 633         /* Get the MV */
 634         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 635         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 636
 637         /* Motion compensation */
 638         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 639                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 640                        mvp[0], mvp[1], 16, 16 );
 641     }
 642
 643     /* get luma diff */
 644     h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0],
 645                                   h->mb.pic.p_fdec[0] );
 646
 647     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 648     {
 649         /* encode one 4x4 block */
 650         for( i4x4 = 0; i4x4 < 4; i4x4++ )
 651         {
 652             const int idx = i8x8 * 4 + i4x4;
 653
 654             h->quantf.quant_4x4( dct4x4[idx], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 655             h->zigzagf.scan_4x4( dctscan, dct4x4[idx] );
 656
 657             i_decimate_mb += x264_mb_decimate_score( dctscan, 16 );
 658
 659             if( i_decimate_mb >= 6 )
 660             {
 661                 /* not as P_SKIP */
 662                 return 0;
 663             }
 664         }
 665     }
 666
 667     /* encode chroma */
 668     i_qp = h->mb.i_chroma_qp;
 669
 670     for( ch = 0; ch < 2; ch++ )
 671     {
 672         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 673         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 674
 675         if( !b_bidir )
 676         {
 677             h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 678                              h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 679                              mvp[0], mvp[1], 8, 8 );
 680         }
 681
 682         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 683
 684         /* calculate dct DC */
 685         dct2x2[0][0] = dct4x4[0][0][0];
 686         dct2x2[0][1] = dct4x4[1][0][0];
 687         dct2x2[1][0] = dct4x4[2][0][0];
 688         dct2x2[1][1] = dct4x4[3][0][0];
 689         h->dctf.dct2x2dc( dct2x2 );
 690         h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4PC][i_qp][0]>>1, h->quant4_bias[CQM_4PC][i_qp][0]<<1 );
 691         if( dct2x2[0][0] || dct2x2[0][1] || dct2x2[1][0] || dct2x2[1][1]  )
 692         {
 693             /* can't be */
 694             return 0;
 695         }
 696
 697         /* calculate dct coeffs */
 698         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 699         {
 700             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 701             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 702
 703             i_decimate_mb += x264_mb_decimate_score( dctscan+1, 15 );
 704             if( i_decimate_mb >= 7 )
 705             {
 706                 return 0;
 707             }
 708         }
 709     }
 710
 711     h->mb.b_skip_pbskip_mc = 1;
 712     return 1;
 713 }
 714
 715 /****************************************************************************
 716  * DCT-domain noise reduction / adaptive deadzone
 717  * from libavcodec
 718  ****************************************************************************/
 719
 720 void x264_noise_reduction_update( x264_t *h )
 721 {
 722     int cat, i;
 723     for( cat = 0; cat < 2; cat++ )
 724     {
 725         int size = cat ? 64 : 16;
 726         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
 727
 728         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
 729         {
 730             for( i = 0; i < size; i++ )
 731                 h->nr_residual_sum[cat][i] >>= 1;
 732             h->nr_count[cat] >>= 1;
 733         }
 734
 735         for( i = 0; i < size; i++ )
 736             h->nr_offset[cat][i] =
 737                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
 738                  + h->nr_residual_sum[cat][i]/2)
 739               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
 740     }
 741 }
 742
 743 void x264_denoise_dct( x264_t *h, int16_t *dct )
 744 {
 745     const int cat = h->mb.b_transform_8x8;
 746     int i;
 747
 748     h->nr_count[cat]++;
 749
 750     for( i = (cat ? 63 : 15); i >= 1; i-- )
 751     {
 752         int level = dct[i];
 753         if( level )
 754         {
 755             if( level > 0 )
 756             {
 757                 h->nr_residual_sum[cat][i] += level;
 758                 level -= h->nr_offset[cat][i];
 759                 if( level < 0 )
 760                     level = 0;
 761             }
 762             else
 763             {
 764                 h->nr_residual_sum[cat][i] -= level;
 765                 level += h->nr_offset[cat][i];
 766                 if( level > 0 )
 767                     level = 0;
 768             }
 769             dct[i] = level;
 770         }
 771     }
 772 }
 773
 774 /*****************************************************************************
 775  * RD only; 4 calls to this do not make up for one macroblock_encode.
 776  * doesn't transform chroma dc.
 777  *****************************************************************************/
 778 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
 779 {
 780     int i_qp = h->mb.i_qp;
 781     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
 782     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
 783     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 784     int nnz8x8;
 785     int ch;
 786
 787     x264_mb_mc_8x8( h, i8 );
 788
 789     if( h->mb.b_transform_8x8 )
 790     {
 791         DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 792         h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
 793         h->quantf.quant_8x8( dct8x8, h->quant8_mf[CQM_8PY][i_qp], h->quant8_bias[CQM_8PY][i_qp] );
 794         h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
 795
 796         if( b_decimate )
 797             nnz8x8 = 4 <= x264_mb_decimate_score( h->dct.luma8x8[i8], 64 );
 798         else
 799             nnz8x8 = array_non_zero( dct8x8 );
 800
 801         if( nnz8x8 )
 802         {
 803             h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
 804             h->dctf.add8x8_idct8( p_fdec, dct8x8 );
 805         }
 806     }
 807     else
 808     {
 809         int i4;
 810         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 811         h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
 812         h->quantf.quant_4x4( dct4x4[0], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 813         h->quantf.quant_4x4( dct4x4[1], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 814         h->quantf.quant_4x4( dct4x4[2], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 815         h->quantf.quant_4x4( dct4x4[3], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 816         for( i4 = 0; i4 < 4; i4++ )
 817             h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
 818
 819         if( b_decimate )
 820         {
 821             int i_decimate_8x8 = 0;
 822             for( i4 = 0; i4 < 4 && i_decimate_8x8 < 4; i4++ )
 823                 i_decimate_8x8 += x264_mb_decimate_score( h->dct.luma4x4[i8*4+i4], 16 );
 824             nnz8x8 = 4 <= i_decimate_8x8;
 825         }
 826         else
 827             nnz8x8 = array_non_zero( dct4x4 );
 828
 829         if( nnz8x8 )
 830         {
 831             for( i4 = 0; i4 < 4; i4++ )
 832                 h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
 833             h->dctf.add8x8_idct( p_fdec, dct4x4 );
 834         }
 835     }
 836
 837     i_qp = h->mb.i_chroma_qp;
 838
 839     for( ch = 0; ch < 2; ch++ )
 840     {
 841         DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
 842         p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
 843         p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
 844
 845         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
 846         h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 847         h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
 848         if( array_non_zero( dct4x4 ) )
 849         {
 850             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
 851             h->dctf.add4x4_idct( p_fdec, dct4x4 );
 852         }
 853     }
 854
 855     if( nnz8x8 )
 856         h->mb.i_cbp_luma |= (1 << i8);
 857     else
 858         h->mb.i_cbp_luma &= ~(1 << i8);
 859     h->mb.i_cbp_chroma = 0x02;
 860 }