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

   1 /*****************************************************************************
   2  * macroblock.c: h264 encoder library
   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 "common/common.h"
  26 #include "macroblock.h"
  27
  28 /* These chroma DC functions don't have assembly versions and are only used here. */
  29
  30 #define ZIG(i,y,x) level[i] = dct[x][y];
  31 static inline void zigzag_scan_2x2_dc( int16_t level[4], int16_t dct[2][2] )
  32 {
  33     ZIG(0,0,0)
  34     ZIG(1,0,1)
  35     ZIG(2,1,0)
  36     ZIG(3,1,1)
  37 }
  38 #undef ZIG
  39
  40 static inline void idct_dequant_2x2_dc( int16_t dct[2][2], int16_t dct4x4[4][4][4], int dequant_mf[6][4][4], int i_qp )
  41 {
  42     int d0 = dct[0][0] + dct[0][1];
  43     int d1 = dct[1][0] + dct[1][1];
  44     int d2 = dct[0][0] - dct[0][1];
  45     int d3 = dct[1][0] - dct[1][1];
  46     int dmf = dequant_mf[i_qp%6][0][0];
  47     int qbits = i_qp/6 - 5;
  48     if( qbits > 0 )
  49     {
  50         dmf <<= qbits;
  51         qbits = 0;
  52     }
  53     dct4x4[0][0][0] = (d0 + d1) * dmf >> -qbits;
  54     dct4x4[1][0][0] = (d0 - d1) * dmf >> -qbits;
  55     dct4x4[2][0][0] = (d2 + d3) * dmf >> -qbits;
  56     dct4x4[3][0][0] = (d2 - d3) * dmf >> -qbits;
  57 }
  58
  59 static inline void dct2x2dc( int16_t d[2][2], int16_t dct4x4[4][4][4] )
  60 {
  61     int d0 = dct4x4[0][0][0] + dct4x4[1][0][0];
  62     int d1 = dct4x4[2][0][0] + dct4x4[3][0][0];
  63     int d2 = dct4x4[0][0][0] - dct4x4[1][0][0];
  64     int d3 = dct4x4[2][0][0] - dct4x4[3][0][0];
  65     d[0][0] = d0 + d1;
  66     d[1][0] = d2 + d3;
  67     d[0][1] = d0 - d1;
  68     d[1][1] = d2 - d3;
  69     dct4x4[0][0][0] = 0;
  70     dct4x4[1][0][0] = 0;
  71     dct4x4[2][0][0] = 0;
  72     dct4x4[3][0][0] = 0;
  73 }
  74
  75 static ALWAYS_INLINE void x264_quant_4x4( x264_t *h, int16_t dct[4][4], int i_qp, int i_ctxBlockCat, int b_intra, int idx )
  76 {
  77     int i_quant_cat = b_intra ? CQM_4IY : CQM_4PY;
  78     if( h->mb.b_trellis )
  79         x264_quant_4x4_trellis( h, dct, i_quant_cat, i_qp, i_ctxBlockCat, b_intra, idx );
  80     else
  81         h->quantf.quant_4x4( dct, h->quant4_mf[i_quant_cat][i_qp], h->quant4_bias[i_quant_cat][i_qp] );
  82 }
  83
  84 static ALWAYS_INLINE void x264_quant_8x8( x264_t *h, int16_t dct[8][8], int i_qp, int b_intra, int idx )
  85 {
  86     int i_quant_cat = b_intra ? CQM_8IY : CQM_8PY;
  87     if( h->mb.b_trellis )
  88         x264_quant_8x8_trellis( h, dct, i_quant_cat, i_qp, b_intra, idx );
  89     else
  90         h->quantf.quant_8x8( dct, h->quant8_mf[i_quant_cat][i_qp], h->quant8_bias[i_quant_cat][i_qp] );
  91 }
  92
  93 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qp )
  94 {
  95     uint8_t *p_src = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[idx]];
  96     uint8_t *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[idx]];
  97     DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
  98
  99     if( h->mb.b_lossless )
 100     {
 101         h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
 102         return;
 103     }
 104
 105     h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
 106
 107     x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 1, idx );
 108
 109     if( array_non_zero( dct4x4 ) )
 110     {
 111         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4 );
 112         h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qp );
 113
 114         /* output samples to fdec */
 115         h->dctf.add4x4_idct( p_dst, dct4x4 );
 116     }
 117     else
 118         memset( h->dct.luma4x4[idx], 0, sizeof(h->dct.luma4x4[idx]));
 119 }
 120
 121 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qp )
 122 {
 123     int x = 8 * (idx&1);
 124     int y = 8 * (idx>>1);
 125     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
 126     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
 127     DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 128
 129     if( h->mb.b_lossless )
 130     {
 131         h->zigzagf.sub_8x8( h->dct.luma8x8[idx], p_src, p_dst );
 132         return;
 133     }
 134
 135     h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
 136
 137     x264_quant_8x8( h, dct8x8, i_qp, 1, idx );
 138
 139     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
 140     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qp );
 141     h->dctf.add8x8_idct8( p_dst, dct8x8 );
 142 }
 143
 144 static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
 145 {
 146     uint8_t  *p_src = h->mb.pic.p_fenc[0];
 147     uint8_t  *p_dst = h->mb.pic.p_fdec[0];
 148
 149     DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 150     DECLARE_ALIGNED_16( int16_t dct_dc4x4[4][4] );
 151
 152     int i;
 153
 154     if( h->mb.b_lossless )
 155     {
 156         for( i = 0; i < 16; i++ )
 157         {
 158             int oe = block_idx_xy_fenc[i];
 159             int od = block_idx_xy_fdec[i];
 160             h->zigzagf.sub_4x4( h->dct.luma4x4[i], p_src+oe, p_dst+od );
 161             dct_dc4x4[0][block_idx_yx_1d[i]] = h->dct.luma4x4[i][0];
 162             h->dct.luma4x4[i][0] = 0;
 163         }
 164         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 165         return;
 166     }
 167
 168     h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
 169     for( i = 0; i < 16; i++ )
 170     {
 171         /* copy dc coeff */
 172         dct_dc4x4[0][block_idx_xy_1d[i]] = dct4x4[i][0][0];
 173         dct4x4[i][0][0] = 0;
 174
 175         /* quant/scan/dequant */
 176         x264_quant_4x4( h, dct4x4[i], i_qp, DCT_LUMA_AC, 1, i );
 177
 178         h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[i] );
 179         h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IY], i_qp );
 180     }
 181
 182     h->dctf.dct4x4dc( dct_dc4x4 );
 183     if( h->mb.b_trellis )
 184         x264_quant_dc_trellis( h, (int16_t*)dct_dc4x4, CQM_4IY, i_qp, DCT_LUMA_DC, 1);
 185     else
 186         h->quantf.quant_4x4_dc( dct_dc4x4, h->quant4_mf[CQM_4IY][i_qp][0]>>1, h->quant4_bias[CQM_4IY][i_qp][0]<<1 );
 187     h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 188
 189     /* output samples to fdec */
 190     h->dctf.idct4x4dc( dct_dc4x4 );
 191     h->quantf.dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[CQM_4IY], i_qp );  /* XXX not inversed */
 192
 193     /* calculate dct coeffs */
 194     for( i = 0; i < 16; i++ )
 195     {
 196         /* copy dc coeff */
 197         dct4x4[i][0][0] = dct_dc4x4[0][block_idx_xy_1d[i]];
 198     }
 199     /* put pixels to fdec */
 200     h->dctf.add16x16_idct( p_dst, dct4x4 );
 201 }
 202
 203 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
 204 {
 205     int i, ch;
 206     int b_decimate = b_inter && (h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate);
 207
 208     for( ch = 0; ch < 2; ch++ )
 209     {
 210         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 211         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 212         int i_decimate_score = 0;
 213
 214         DECLARE_ALIGNED_16( int16_t dct2x2[2][2]  );
 215         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 216
 217         if( h->mb.b_lossless )
 218         {
 219             for( i = 0; i < 4; i++ )
 220             {
 221                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 222                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 223                 h->zigzagf.sub_4x4( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od );
 224                 h->dct.chroma_dc[ch][i] = h->dct.luma4x4[16+i+ch*4][0];
 225                 h->dct.luma4x4[16+i+ch*4][0] = 0;
 226             }
 227             continue;
 228         }
 229
 230         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 231         dct2x2dc( dct2x2, dct4x4 );
 232         /* calculate dct coeffs */
 233         for( i = 0; i < 4; i++ )
 234         {
 235             if( h->mb.b_trellis )
 236                 x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 0 );
 237             else
 238                 h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[CQM_4IC+b_inter][i_qp], h->quant4_bias[CQM_4IC+b_inter][i_qp] );
 239             h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 240
 241             if( b_decimate )
 242                 i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[16+i+ch*4] );
 243         }
 244
 245         if( h->mb.b_trellis )
 246             x264_quant_dc_trellis( h, (int16_t*)dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter );
 247         else
 248             h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4IC+b_inter][i_qp][0]>>1, h->quant4_bias[CQM_4IC+b_inter][i_qp][0]<<1 );
 249
 250         if( b_decimate && i_decimate_score < 7 )
 251         {
 252             /* Near null chroma 8x8 block so make it null (bits saving) */
 253             memset( &h->dct.luma4x4[16+ch*4], 0, 4 * sizeof( *h->dct.luma4x4 ) );
 254             if( !array_non_zero( dct2x2 ) )
 255             {
 256                 memset( h->dct.chroma_dc[ch], 0, sizeof( h->dct.chroma_dc[ch] ) );
 257                 continue;
 258             }
 259             memset( dct4x4, 0, sizeof( dct4x4 ) );
 260         }
 261         else
 262         {
 263             for( i = 0; i < 4; i++ )
 264                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 265         }
 266
 267         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 268         idct_dequant_2x2_dc( dct2x2, dct4x4, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 269         h->dctf.add8x8_idct( p_dst, dct4x4 );
 270     }
 271
 272     /* coded block pattern */
 273     h->mb.i_cbp_chroma = 0;
 274     for( i = 0; i < 8; i++ )
 275     {
 276         int nz = array_non_zero( h->dct.luma4x4[16+i] );
 277         h->mb.cache.non_zero_count[x264_scan8[16+i]] = nz;
 278         h->mb.i_cbp_chroma |= nz;
 279     }
 280     h->mb.cache.non_zero_count[x264_scan8[25]] = array_non_zero( h->dct.chroma_dc[0] );
 281     h->mb.cache.non_zero_count[x264_scan8[26]] = array_non_zero( h->dct.chroma_dc[1] );
 282     if( h->mb.i_cbp_chroma )
 283         h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
 284     else if( h->mb.cache.non_zero_count[x264_scan8[25]] |
 285              h->mb.cache.non_zero_count[x264_scan8[26]] )
 286         h->mb.i_cbp_chroma = 1;    /* dc only */
 287 }
 288
 289 static void x264_macroblock_encode_skip( x264_t *h )
 290 {
 291     h->mb.i_cbp_luma = 0x00;
 292     h->mb.i_cbp_chroma = 0x00;
 293     memset( h->mb.cache.non_zero_count, 0, X264_SCAN8_SIZE );
 294     /* store cbp */
 295     h->mb.cbp[h->mb.i_mb_xy] = 0;
 296 }
 297
 298 /*****************************************************************************
 299  * x264_macroblock_encode_pskip:
 300  *  Encode an already marked skip block
 301  *****************************************************************************/
 302 static void x264_macroblock_encode_pskip( x264_t *h )
 303 {
 304     const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 305                                 h->mb.mv_min[0], h->mb.mv_max[0] );
 306     const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 307                                 h->mb.mv_min[1], h->mb.mv_max[1] );
 308
 309     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 310     if( !h->mb.b_skip_mc )
 311     {
 312         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 313                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 314                        mvx, mvy, 16, 16 );
 315
 316         h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 317                          h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 318                          mvx, mvy, 8, 8 );
 319
 320         h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 321                          h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 322                          mvx, mvy, 8, 8 );
 323     }
 324
 325     x264_macroblock_encode_skip( h );
 326 }
 327
 328 /*****************************************************************************
 329  * Intra prediction for predictive lossless mode.
 330  *****************************************************************************/
 331
 332 /* Note that these functions take a shortcut (mc.copy instead of actual pixel prediction) which assumes
 333  * that the edge pixels of the reconstructed frame are the same as that of the source frame.  This means
 334  * they will only work correctly if the neighboring blocks are losslessly coded.  In practice, this means
 335  * lossless mode cannot be mixed with lossy mode within a frame. */
 336 /* This can be resolved by explicitly copying the edge pixels after doing the mc.copy, but this doesn't
 337  * need to be done unless we decide to allow mixing lossless and lossy compression. */
 338
 339 void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
 340 {
 341     int stride = h->fenc->i_stride[1] << h->mb.b_interlaced;
 342     if( i_mode == I_PRED_CHROMA_V )
 343     {
 344         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-stride, stride, 8 );
 345         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-stride, stride, 8 );
 346     }
 347     else if( i_mode == I_PRED_CHROMA_H )
 348     {
 349         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-1, stride, 8 );
 350         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-1, stride, 8 );
 351     }
 352     else
 353     {
 354         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 355         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 356     }
 357 }
 358
 359 void x264_predict_lossless_4x4( x264_t *h, uint8_t *p_dst, int idx, int i_mode )
 360 {
 361     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 362     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
 363
 364     if( i_mode == I_PRED_4x4_V )
 365         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
 366     else if( i_mode == I_PRED_4x4_H )
 367         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-1, stride, 4 );
 368     else
 369         h->predict_4x4[i_mode]( p_dst );
 370 }
 371
 372 void x264_predict_lossless_8x8( x264_t *h, uint8_t *p_dst, int idx, int i_mode, uint8_t edge[33] )
 373 {
 374     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 375     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
 376
 377     if( i_mode == I_PRED_8x8_V )
 378         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
 379     else if( i_mode == I_PRED_8x8_H )
 380         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-1, stride, 8 );
 381     else
 382         h->predict_8x8[i_mode]( p_dst, edge );
 383 }
 384
 385 void x264_predict_lossless_16x16( x264_t *h, int i_mode )
 386 {
 387     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 388     if( i_mode == I_PRED_16x16_V )
 389         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-stride, stride, 16 );
 390     else if( i_mode == I_PRED_16x16_H )
 391         h->mc.copy_16x16_unaligned( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-1, stride, 16 );
 392     else
 393         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 394 }
 395
 396 /*****************************************************************************
 397  * x264_macroblock_encode:
 398  *****************************************************************************/
 399 void x264_macroblock_encode( x264_t *h )
 400 {
 401     int i_cbp_dc = 0;
 402     int i_qp = h->mb.i_qp;
 403     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 404     int b_force_no_skip = 0;
 405     int i,j,idx;
 406     uint8_t nnz8x8[4] = {1,1,1,1};
 407
 408     if( h->sh.b_mbaff
 409         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 410         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 411     {
 412         /* The first skip is predicted to be a frame mb pair.
 413          * We don't yet support the aff part of mbaff, so force it to non-skip
 414          * so that we can pick the aff flag. */
 415         b_force_no_skip = 1;
 416         if( IS_SKIP(h->mb.i_type) )
 417         {
 418             if( h->mb.i_type == P_SKIP )
 419                 h->mb.i_type = P_L0;
 420             else if( h->mb.i_type == B_SKIP )
 421                 h->mb.i_type = B_DIRECT;
 422         }
 423     }
 424
 425     if( h->mb.i_type == P_SKIP )
 426     {
 427         /* A bit special */
 428         x264_macroblock_encode_pskip( h );
 429         return;
 430     }
 431     if( h->mb.i_type == B_SKIP )
 432     {
 433         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 434         if( !h->mb.b_skip_mc )
 435             x264_mb_mc( h );
 436         x264_macroblock_encode_skip( h );
 437         return;
 438     }
 439
 440     if( h->mb.i_type == I_16x16 )
 441     {
 442         const int i_mode = h->mb.i_intra16x16_pred_mode;
 443         h->mb.b_transform_8x8 = 0;
 444
 445         if( h->mb.b_lossless )
 446             x264_predict_lossless_16x16( h, i_mode );
 447         else
 448             h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 449
 450         /* encode the 16x16 macroblock */
 451         x264_mb_encode_i16x16( h, i_qp );
 452     }
 453     else if( h->mb.i_type == I_8x8 )
 454     {
 455         DECLARE_ALIGNED_16( uint8_t edge[33] );
 456         h->mb.b_transform_8x8 = 1;
 457         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 458         if( h->mb.i_skip_intra )
 459         {
 460             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 461             /* In RD mode, restore the now-overwritten DCT data. */
 462             if( h->mb.i_skip_intra == 2 )
 463                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 464         }
 465         for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 466         {
 467             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 468             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 469             x264_predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 470
 471             if( h->mb.b_lossless )
 472                 x264_predict_lossless_8x8( h, p_dst, i, i_mode, edge );
 473             else
 474                 h->predict_8x8[i_mode]( p_dst, edge );
 475
 476             x264_mb_encode_i8x8( h, i, i_qp );
 477         }
 478         for( i = 0; i < 4; i++ )
 479             nnz8x8[i] = array_non_zero( h->dct.luma8x8[i] );
 480     }
 481     else if( h->mb.i_type == I_4x4 )
 482     {
 483         h->mb.b_transform_8x8 = 0;
 484         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 485         if( h->mb.i_skip_intra )
 486         {
 487             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 488             /* In RD mode, restore the now-overwritten DCT data. */
 489             if( h->mb.i_skip_intra == 2 )
 490                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 491         }
 492         for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 493         {
 494             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
 495             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 496
 497             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 498                 /* emulate missing topright samples */
 499                 *(uint32_t*) &p_dst[4-FDEC_STRIDE] = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 500
 501             if( h->mb.b_lossless )
 502                 x264_predict_lossless_4x4( h, p_dst, i, i_mode );
 503             else
 504                 h->predict_4x4[i_mode]( p_dst );
 505             x264_mb_encode_i4x4( h, i, i_qp );
 506         }
 507     }
 508     else    /* Inter MB */
 509     {
 510         int i8x8, i4x4;
 511         int i_decimate_mb = 0;
 512
 513         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
 514         if( !h->mb.b_skip_mc )
 515             x264_mb_mc( h );
 516
 517         if( h->mb.b_lossless )
 518         {
 519             if( h->mb.b_transform_8x8 )
 520                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 521                 {
 522                     int x = 8*(i8x8&1);
 523                     int y = 8*(i8x8>>1);
 524                     h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8],
 525                                         h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
 526                                         h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
 527                     nnz8x8[i8x8] = array_non_zero( h->dct.luma8x8[i8x8] );
 528                 }
 529             else
 530                 for( i4x4 = 0; i4x4 < 16; i4x4++ )
 531                 {
 532                     h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 533                                         h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
 534                                         h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4x4] );
 535                 }
 536         }
 537         else if( h->mb.b_transform_8x8 )
 538         {
 539             DECLARE_ALIGNED_16( int16_t dct8x8[4][8][8] );
 540             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 541             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 542             h->nr_count[1] += h->mb.b_noise_reduction * 4;
 543
 544             for( idx = 0; idx < 4; idx++ )
 545             {
 546                 if( h->mb.b_noise_reduction )
 547                     h->quantf.denoise_dct( *dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
 548                 x264_quant_8x8( h, dct8x8[idx], i_qp, 0, idx );
 549
 550                 h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 551
 552                 if( b_decimate )
 553                 {
 554                     int i_decimate_8x8 = h->quantf.decimate_score64( h->dct.luma8x8[idx] );
 555                     i_decimate_mb += i_decimate_8x8;
 556                     if( i_decimate_8x8 < 4 )
 557                         nnz8x8[idx] = 0;
 558                 }
 559                 else
 560                     nnz8x8[idx] = array_non_zero( dct8x8[idx] );
 561             }
 562
 563             if( i_decimate_mb < 6 && b_decimate )
 564                 *(uint32_t*)nnz8x8 = 0;
 565             else
 566             {
 567                 for( idx = 0; idx < 4; idx++ )
 568                     if( nnz8x8[idx] )
 569                     {
 570                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 571                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 572                     }
 573             }
 574         }
 575         else
 576         {
 577             DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 578             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 579             h->nr_count[0] += h->mb.b_noise_reduction * 16;
 580
 581             for( i8x8 = 0; i8x8 < 4; i8x8++ )
 582             {
 583                 int i_decimate_8x8;
 584
 585                 /* encode one 4x4 block */
 586                 i_decimate_8x8 = 0;
 587                 for( i4x4 = 0; i4x4 < 4; i4x4++ )
 588                 {
 589                     idx = i8x8 * 4 + i4x4;
 590
 591                     if( h->mb.b_noise_reduction )
 592                         h->quantf.denoise_dct( *dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
 593                     x264_quant_4x4( h, dct4x4[idx], i_qp, DCT_LUMA_4x4, 0, idx );
 594
 595                     h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 596
 597                     if( b_decimate && i_decimate_8x8 < 6 )
 598                         i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[idx] );
 599                 }
 600
 601                 /* decimate this 8x8 block */
 602                 i_decimate_mb += i_decimate_8x8;
 603                 if( i_decimate_8x8 < 4 && b_decimate )
 604                     nnz8x8[i8x8] = 0;
 605             }
 606
 607             if( i_decimate_mb < 6 && b_decimate )
 608                 *(uint32_t*)nnz8x8 = 0;
 609             else
 610             {
 611                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 612                     if( nnz8x8[i8x8] )
 613                     {
 614                         for( i = 0; i < 4; i++ )
 615                             h->quantf.dequant_4x4( dct4x4[i8x8*4+i], h->dequant4_mf[CQM_4PY], i_qp );
 616                         h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 617                     }
 618             }
 619         }
 620     }
 621
 622     /* encode chroma */
 623     if( IS_INTRA( h->mb.i_type ) )
 624     {
 625         const int i_mode = h->mb.i_chroma_pred_mode;
 626         if( h->mb.b_lossless )
 627             x264_predict_lossless_8x8_chroma( h, i_mode );
 628         else
 629         {
 630             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 631             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 632         }
 633     }
 634
 635     /* encode the 8x8 blocks */
 636     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 637
 638     /* coded block pattern and non_zero_count */
 639     h->mb.i_cbp_luma = 0x00;
 640     if( h->mb.i_type == I_16x16 )
 641     {
 642         for( i = 0; i < 16; i++ )
 643         {
 644             int nz = array_non_zero( h->dct.luma4x4[i] );
 645             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 646             h->mb.i_cbp_luma |= nz;
 647         }
 648         h->mb.i_cbp_luma *= 0xf;
 649         h->mb.cache.non_zero_count[x264_scan8[24]] = array_non_zero( h->dct.luma16x16_dc );
 650     }
 651     else
 652     {
 653         for( i = 0; i < 4; i++)
 654         {
 655             if(!nnz8x8[i])
 656             {
 657                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[0+i*4]] = 0;
 658                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[2+i*4]] = 0;
 659             }
 660             else if( h->mb.b_transform_8x8 )
 661             {
 662                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[0+4*i]] = nnz8x8[i] * 0x0101;
 663                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[2+4*i]] = nnz8x8[i] * 0x0101;
 664                 h->mb.i_cbp_luma |= nnz8x8[i] << i;
 665             }
 666             else
 667             {
 668                 int nz, cbp = 0;
 669                 for( j = 0; j < 4; j++ )
 670                 {
 671                     nz = array_non_zero( h->dct.luma4x4[j+4*i] );
 672                     h->mb.cache.non_zero_count[x264_scan8[j+4*i]] = nz;
 673                     cbp |= nz;
 674                 }
 675                 h->mb.i_cbp_luma |= cbp << i;
 676             }
 677         }
 678         h->mb.cache.non_zero_count[x264_scan8[24]] = 0;
 679     }
 680
 681     if( h->param.b_cabac )
 682     {
 683         i_cbp_dc = h->mb.cache.non_zero_count[x264_scan8[24]]
 684                  | h->mb.cache.non_zero_count[x264_scan8[25]] << 1
 685                  | h->mb.cache.non_zero_count[x264_scan8[26]] << 2;
 686     }
 687
 688     /* store cbp */
 689     h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
 690
 691     /* Check for P_SKIP
 692      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 693      *      (if multiple mv give same result)*/
 694     if( !b_force_no_skip )
 695     {
 696         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 697             !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
 698             *(uint32_t*)h->mb.cache.mv[0][x264_scan8[0]] == *(uint32_t*)h->mb.cache.pskip_mv
 699             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 700         {
 701             h->mb.i_type = P_SKIP;
 702         }
 703
 704         /* Check for B_SKIP */
 705         if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
 706         {
 707             h->mb.i_type = B_SKIP;
 708         }
 709     }
 710 }
 711
 712 /*****************************************************************************
 713  * x264_macroblock_probe_skip:
 714  *  Check if the current MB could be encoded as a [PB]_SKIP (it supposes you use
 715  *  the previous QP
 716  *****************************************************************************/
 717 int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
 718 {
 719     DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 720     DECLARE_ALIGNED_16( int16_t dct2x2[2][2] );
 721     DECLARE_ALIGNED_16( int16_t dctscan[16] );
 722
 723     int i_qp = h->mb.i_qp;
 724     int mvp[2];
 725     int ch, thresh;
 726
 727     int i8x8, i4x4;
 728     int i_decimate_mb;
 729
 730     if( !b_bidir )
 731     {
 732         /* Get the MV */
 733         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 734         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 735
 736         /* Motion compensation */
 737         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 738                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 739                        mvp[0], mvp[1], 16, 16 );
 740     }
 741
 742     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 743     {
 744         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
 745         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
 746         /* get luma diff */
 747         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
 748                                     h->mb.pic.p_fdec[0] + fdec_offset );
 749         /* encode one 4x4 block */
 750         for( i4x4 = 0; i4x4 < 4; i4x4++ )
 751         {
 752             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 753             if( !array_non_zero(dct4x4[i4x4]) )
 754                 continue;
 755             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 756             i_decimate_mb += h->quantf.decimate_score16( dctscan );
 757             if( i_decimate_mb >= 6 )
 758                 return 0;
 759         }
 760     }
 761
 762     /* encode chroma */
 763     i_qp = h->mb.i_chroma_qp;
 764     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 765
 766     for( ch = 0; ch < 2; ch++ )
 767     {
 768         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 769         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 770
 771         if( !b_bidir )
 772         {
 773             h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 774                              h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 775                              mvp[0], mvp[1], 8, 8 );
 776         }
 777
 778         /* there is almost never a termination during chroma, but we can't avoid the check entirely */
 779         /* so instead we check SSD and skip the actual check if the score is low enough. */
 780         if( h->pixf.ssd[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE ) < thresh )
 781             continue;
 782
 783         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 784
 785         /* calculate dct DC */
 786         dct2x2dc( dct2x2, dct4x4 );
 787         h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4PC][i_qp][0]>>1, h->quant4_bias[CQM_4PC][i_qp][0]<<1 );
 788         if( array_non_zero(dct2x2) )
 789             return 0;
 790
 791         /* calculate dct coeffs */
 792         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 793         {
 794             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 795             if( !array_non_zero(dct4x4[i4x4]) )
 796                 continue;
 797             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 798             i_decimate_mb += h->quantf.decimate_score15( dctscan );
 799             if( i_decimate_mb >= 7 )
 800                 return 0;
 801         }
 802     }
 803
 804     h->mb.b_skip_mc = 1;
 805     return 1;
 806 }
 807
 808 /****************************************************************************
 809  * DCT-domain noise reduction / adaptive deadzone
 810  * from libavcodec
 811  ****************************************************************************/
 812
 813 void x264_noise_reduction_update( x264_t *h )
 814 {
 815     int cat, i;
 816     for( cat = 0; cat < 2; cat++ )
 817     {
 818         int size = cat ? 64 : 16;
 819         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
 820
 821         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
 822         {
 823             for( i = 0; i < size; i++ )
 824                 h->nr_residual_sum[cat][i] >>= 1;
 825             h->nr_count[cat] >>= 1;
 826         }
 827
 828         for( i = 0; i < size; i++ )
 829             h->nr_offset[cat][i] =
 830                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
 831                  + h->nr_residual_sum[cat][i]/2)
 832               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
 833     }
 834 }
 835
 836 /*****************************************************************************
 837  * RD only; 4 calls to this do not make up for one macroblock_encode.
 838  * doesn't transform chroma dc.
 839  *****************************************************************************/
 840 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
 841 {
 842     int i_qp = h->mb.i_qp;
 843     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
 844     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
 845     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 846     int nnz8x8 = 0;
 847     int ch;
 848
 849     x264_mb_mc_8x8( h, i8 );
 850
 851     if( h->mb.b_lossless )
 852     {
 853         int i4;
 854         if( h->mb.b_transform_8x8 )
 855         {
 856             h->zigzagf.sub_8x8( h->dct.luma8x8[i8], p_fenc, p_fdec );
 857             nnz8x8 = array_non_zero( h->dct.luma8x8[i8] );
 858             *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+0]] = 0x0101 * nnz8x8;
 859             *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+2]] = 0x0101 * nnz8x8;
 860         }
 861         else
 862         {
 863             for( i4 = i8*4; i4 < i8*4+4; i4++ )
 864             {
 865                 int nz;
 866                 h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
 867                                     h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
 868                                     h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
 869                 nz = array_non_zero( h->dct.luma4x4[i4] );
 870                 h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
 871                 nnz8x8 |= nz;
 872             }
 873         }
 874         for( ch = 0; ch < 2; ch++ )
 875         {
 876             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
 877             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
 878             h->zigzagf.sub_4x4( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec );
 879             h->dct.luma4x4[16+i8+ch*4][0] = 0;
 880         }
 881         h->mb.cache.non_zero_count[x264_scan8[16+i8]] = array_non_zero( h->dct.luma4x4[16+i8] );
 882         h->mb.cache.non_zero_count[x264_scan8[20+i8]] = array_non_zero( h->dct.luma4x4[20+i8] );
 883     }
 884     else
 885     {
 886         if( h->mb.b_transform_8x8 )
 887         {
 888             DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 889             h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
 890             x264_quant_8x8( h, dct8x8, i_qp, 0, i8 );
 891             h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
 892
 893             if( b_decimate && !h->mb.b_trellis )
 894                 nnz8x8 = 4 <= h->quantf.decimate_score64( h->dct.luma8x8[i8] );
 895             else
 896                 nnz8x8 = array_non_zero( dct8x8 );
 897
 898             if( nnz8x8 )
 899             {
 900                 h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
 901                 h->dctf.add8x8_idct8( p_fdec, dct8x8 );
 902                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+0]] = 0x0101;
 903                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+2]] = 0x0101;
 904             }
 905             else
 906             {
 907                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+0]] = 0;
 908                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+2]] = 0;
 909             }
 910         }
 911         else
 912         {
 913             int i4;
 914             DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 915             h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
 916             for( i4 = 0; i4 < 4; i4++ )
 917                 x264_quant_4x4( h, dct4x4[i4], i_qp, DCT_LUMA_4x4, 0, i8*4+i4 );
 918
 919             for( i4 = 0; i4 < 4; i4++ )
 920                 h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
 921
 922             if( b_decimate )
 923             {
 924                 int i_decimate_8x8 = 0;
 925                 for( i4 = 0; i4 < 4 && i_decimate_8x8 < 4; i4++ )
 926                     i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[i8*4+i4] );
 927                 nnz8x8 = 4 <= i_decimate_8x8;
 928             }
 929             else
 930                 nnz8x8 = array_non_zero( dct4x4 );
 931
 932             if( nnz8x8 )
 933             {
 934                 for( i4 = 0; i4 < 4; i4++ )
 935                 {
 936                     if( array_non_zero( dct4x4[i4] ) )
 937                     {
 938                         h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
 939                         h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = 1;
 940                     }
 941                     else
 942                         h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = 0;
 943                 }
 944                 h->dctf.add8x8_idct( p_fdec, dct4x4 );
 945             }
 946             else
 947             {
 948                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+0]] = 0;
 949                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+2]] = 0;
 950             }
 951         }
 952
 953         i_qp = h->mb.i_chroma_qp;
 954
 955         for( ch = 0; ch < 2; ch++ )
 956         {
 957             DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
 958             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
 959             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
 960
 961             h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
 962             dct4x4[0][0] = 0;
 963
 964             if( h->mb.b_trellis )
 965                 x264_quant_4x4_trellis( h, dct4x4, CQM_4PC, i_qp, DCT_CHROMA_AC, 0, 0 );
 966             else
 967                 h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 968
 969             if( array_non_zero( dct4x4 ) )
 970             {
 971                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
 972                 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
 973                 h->dctf.add4x4_idct( p_fdec, dct4x4 );
 974                 h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = 1;
 975             }
 976             else
 977                 h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = 0;
 978         }
 979     }
 980     h->mb.i_cbp_luma &= ~(1 << i8);
 981     h->mb.i_cbp_luma |= nnz8x8 << i8;
 982     h->mb.i_cbp_chroma = 0x02;
 983 }
 984
 985 /*****************************************************************************
 986  * RD only, luma only
 987  *****************************************************************************/
 988 void x264_macroblock_encode_p4x4( x264_t *h, int i4 )
 989 {
 990     int i_qp = h->mb.i_qp;
 991     uint8_t *p_fenc = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[i4]];
 992     uint8_t *p_fdec = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
 993     const int i_ref = h->mb.cache.ref[0][x264_scan8[i4]];
 994     const int mvx   = x264_clip3( h->mb.cache.mv[0][x264_scan8[i4]][0], h->mb.mv_min[0], h->mb.mv_max[0] );
 995     const int mvy   = x264_clip3( h->mb.cache.mv[0][x264_scan8[i4]][1], h->mb.mv_min[1], h->mb.mv_max[1] );
 996
 997     h->mc.mc_luma( p_fdec, FDEC_STRIDE, h->mb.pic.p_fref[0][i_ref], h->mb.pic.i_stride[0], mvx + 4*4*block_idx_x[i4], mvy + 4*4*block_idx_y[i4], 4, 4 );
 998
 999     if( h->mb.b_lossless )
1000     {
1001         h->zigzagf.sub_4x4( h->dct.luma4x4[i4], p_fenc, p_fdec );
1002         h->mb.cache.non_zero_count[x264_scan8[i4]] = array_non_zero( h->dct.luma4x4[i4] );
1003     }
1004     else
1005     {
1006         DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
1007         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1008         x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 0, i4 );
1009         if( array_non_zero( dct4x4 ) )
1010         {
1011             h->zigzagf.scan_4x4( h->dct.luma4x4[i4], dct4x4 );
1012             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PY], i_qp );
1013             h->dctf.add4x4_idct( p_fdec, dct4x4 );
1014             h->mb.cache.non_zero_count[x264_scan8[i4]] = 1;
1015         }
1016         else
1017             h->mb.cache.non_zero_count[x264_scan8[i4]] = 0;
1018     }
1019 }