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 #define IDCT_DEQUANT_START \
  41     int d0 = dct[0][0] + dct[0][1]; \
  42     int d1 = dct[1][0] + dct[1][1]; \
  43     int d2 = dct[0][0] - dct[0][1]; \
  44     int d3 = dct[1][0] - dct[1][1]; \
  45     int dmf = dequant_mf[i_qp%6][0][0]; \
  46     int qbits = i_qp/6 - 5; \
  47     if( qbits > 0 ) \
  48     { \
  49         dmf <<= qbits; \
  50         qbits = 0; \
  51     }
  52
  53 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 )
  54 {
  55     IDCT_DEQUANT_START
  56     dct4x4[0][0][0] = (d0 + d1) * dmf >> -qbits;
  57     dct4x4[1][0][0] = (d0 - d1) * dmf >> -qbits;
  58     dct4x4[2][0][0] = (d2 + d3) * dmf >> -qbits;
  59     dct4x4[3][0][0] = (d2 - d3) * dmf >> -qbits;
  60 }
  61
  62 static inline void idct_dequant_2x2_dconly( int16_t dct[2][2], int dequant_mf[6][4][4], int i_qp )
  63 {
  64     IDCT_DEQUANT_START
  65     dct[0][0] = (d0 + d1) * dmf >> -qbits;
  66     dct[0][1] = (d0 - d1) * dmf >> -qbits;
  67     dct[1][0] = (d2 + d3) * dmf >> -qbits;
  68     dct[1][1] = (d2 - d3) * dmf >> -qbits;
  69 }
  70
  71 static inline void dct2x2dc( int16_t d[2][2], int16_t dct4x4[4][4][4] )
  72 {
  73     int d0 = dct4x4[0][0][0] + dct4x4[1][0][0];
  74     int d1 = dct4x4[2][0][0] + dct4x4[3][0][0];
  75     int d2 = dct4x4[0][0][0] - dct4x4[1][0][0];
  76     int d3 = dct4x4[2][0][0] - dct4x4[3][0][0];
  77     d[0][0] = d0 + d1;
  78     d[1][0] = d2 + d3;
  79     d[0][1] = d0 - d1;
  80     d[1][1] = d2 - d3;
  81     dct4x4[0][0][0] = 0;
  82     dct4x4[1][0][0] = 0;
  83     dct4x4[2][0][0] = 0;
  84     dct4x4[3][0][0] = 0;
  85 }
  86
  87 static inline void dct2x2dc_dconly( int16_t d[2][2] )
  88 {
  89     int d0 = d[0][0] + d[0][1];
  90     int d1 = d[1][0] + d[1][1];
  91     int d2 = d[0][0] - d[0][1];
  92     int d3 = d[1][0] - d[1][1];
  93     d[0][0] = d0 + d1;
  94     d[1][0] = d2 + d3;
  95     d[0][1] = d0 - d1;
  96     d[1][1] = d2 - d3;
  97 }
  98
  99 static ALWAYS_INLINE int x264_quant_4x4( x264_t *h, int16_t dct[4][4], int i_qp, int i_ctxBlockCat, int b_intra, int idx )
 100 {
 101     int i_quant_cat = b_intra ? CQM_4IY : CQM_4PY;
 102     if( h->mb.b_trellis )
 103         return x264_quant_4x4_trellis( h, dct, i_quant_cat, i_qp, i_ctxBlockCat, b_intra, 0, idx );
 104     else
 105         return h->quantf.quant_4x4( dct, h->quant4_mf[i_quant_cat][i_qp], h->quant4_bias[i_quant_cat][i_qp] );
 106 }
 107
 108 static ALWAYS_INLINE int x264_quant_8x8( x264_t *h, int16_t dct[8][8], int i_qp, int b_intra, int idx )
 109 {
 110     int i_quant_cat = b_intra ? CQM_8IY : CQM_8PY;
 111     if( h->mb.b_trellis )
 112         return x264_quant_8x8_trellis( h, dct, i_quant_cat, i_qp, b_intra, idx );
 113     else
 114         return h->quantf.quant_8x8( dct, h->quant8_mf[i_quant_cat][i_qp], h->quant8_bias[i_quant_cat][i_qp] );
 115 }
 116
 117 /* All encoding functions must output the correct CBP and NNZ values.
 118  * The entropy coding functions will check CBP first, then NNZ, before
 119  * actually reading the DCT coefficients.  NNZ still must be correct even
 120  * if CBP is zero because of the use of NNZ values for context selection.
 121  * "NNZ" need only be 0 or 1 rather than the exact coefficient count because
 122  * that is only needed in CAVLC, and will be calculated by CAVLC's residual
 123  * coding and stored as necessary. */
 124
 125 /* This means that decimation can be done merely by adjusting the CBP and NNZ
 126  * rather than memsetting the coefficients. */
 127
 128 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qp )
 129 {
 130     int nz;
 131     uint8_t *p_src = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[idx]];
 132     uint8_t *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[idx]];
 133     DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
 134
 135     if( h->mb.b_lossless )
 136     {
 137         nz = h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
 138         h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 139         h->mb.i_cbp_luma |= nz<<(idx>>2);
 140         return;
 141     }
 142
 143     h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
 144
 145     nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 1, idx );
 146     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 147     if( nz )
 148     {
 149         h->mb.i_cbp_luma |= 1<<(idx>>2);
 150         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4 );
 151         h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qp );
 152         h->dctf.add4x4_idct( p_dst, dct4x4 );
 153     }
 154 }
 155
 156 #define STORE_8x8_NNZ(idx,nz)\
 157 {\
 158     *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[idx*4+0]] = nz * 0x0101;\
 159     *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[idx*4+2]] = nz * 0x0101;\
 160 }
 161
 162 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qp )
 163 {
 164     int x = 8 * (idx&1);
 165     int y = 8 * (idx>>1);
 166     int nz;
 167     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
 168     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
 169     DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 170
 171     if( h->mb.b_lossless )
 172     {
 173         nz = h->zigzagf.sub_8x8( h->dct.luma8x8[idx], p_src, p_dst );
 174         STORE_8x8_NNZ(idx,nz);
 175         h->mb.i_cbp_luma |= nz<<idx;
 176         return;
 177     }
 178
 179     h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
 180
 181     nz = x264_quant_8x8( h, dct8x8, i_qp, 1, idx );
 182     if( nz )
 183     {
 184         h->mb.i_cbp_luma |= 1<<idx;
 185         h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
 186         h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qp );
 187         h->dctf.add8x8_idct8( p_dst, dct8x8 );
 188         STORE_8x8_NNZ(idx,1);
 189     }
 190     else
 191         STORE_8x8_NNZ(idx,0);
 192 }
 193
 194 static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
 195 {
 196     uint8_t  *p_src = h->mb.pic.p_fenc[0];
 197     uint8_t  *p_dst = h->mb.pic.p_fdec[0];
 198
 199     DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 200     DECLARE_ALIGNED_16( int16_t dct_dc4x4[4][4] );
 201
 202     int i, nz;
 203     int b_decimate = h->sh.i_type == SLICE_TYPE_B || (h->param.analyse.b_dct_decimate && h->sh.i_type == SLICE_TYPE_P);
 204     int decimate_score = b_decimate ? 0 : 9;
 205
 206     if( h->mb.b_lossless )
 207     {
 208         for( i = 0; i < 16; i++ )
 209         {
 210             int oe = block_idx_xy_fenc[i];
 211             int od = block_idx_xy_fdec[i];
 212             nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[i], p_src+oe, p_dst+od, &dct_dc4x4[0][block_idx_yx_1d[i]] );
 213             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 214             h->mb.i_cbp_luma |= nz;
 215         }
 216         h->mb.i_cbp_luma *= 0xf;
 217         h->mb.cache.non_zero_count[x264_scan8[24]] = array_non_zero( dct_dc4x4 );
 218         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 219         return;
 220     }
 221
 222     h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
 223
 224     for( i = 0; i < 16; i++ )
 225     {
 226         /* copy dc coeff */
 227         dct_dc4x4[0][block_idx_xy_1d[i]] = dct4x4[i][0][0];
 228         dct4x4[i][0][0] = 0;
 229
 230         /* quant/scan/dequant */
 231         nz = x264_quant_4x4( h, dct4x4[i], i_qp, DCT_LUMA_AC, 1, i );
 232         h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 233         if( nz )
 234         {
 235             h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[i] );
 236             h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IY], i_qp );
 237             if( decimate_score < 6 ) decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[i] );
 238             h->mb.i_cbp_luma = 0xf;
 239         }
 240     }
 241
 242     /* Writing the 16 CBFs in an i16x16 block is quite costly, so decimation can save many bits. */
 243     /* More useful with CAVLC, but still useful with CABAC. */
 244     if( decimate_score < 6 )
 245     {
 246         h->mb.i_cbp_luma = 0;
 247         *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = 0;
 248         *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = 0;
 249         *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = 0;
 250         *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = 0;
 251     }
 252
 253     h->dctf.dct4x4dc( dct_dc4x4 );
 254     if( h->mb.b_trellis )
 255         nz = x264_quant_dc_trellis( h, (int16_t*)dct_dc4x4, CQM_4IY, i_qp, DCT_LUMA_DC, 1, 0 );
 256     else
 257         nz = 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 );
 258
 259     h->mb.cache.non_zero_count[x264_scan8[24]] = nz;
 260     if( nz )
 261     {
 262         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 263
 264         /* output samples to fdec */
 265         h->dctf.idct4x4dc( dct_dc4x4 );
 266         h->quantf.dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[CQM_4IY], i_qp );  /* XXX not inversed */
 267         if( h->mb.i_cbp_luma )
 268             for( i = 0; i < 16; i++ )
 269                 dct4x4[i][0][0] = dct_dc4x4[0][block_idx_xy_1d[i]];
 270     }
 271
 272     /* put pixels to fdec */
 273     if( h->mb.i_cbp_luma )
 274         h->dctf.add16x16_idct( p_dst, dct4x4 );
 275     else if( nz )
 276         h->dctf.add16x16_idct_dc( p_dst, dct_dc4x4 );
 277 }
 278
 279 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
 280 {
 281     int i, ch, nz, nz_dc;
 282     int b_decimate = b_inter && (h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate);
 283     DECLARE_ALIGNED_16( int16_t dct2x2[2][2]  );
 284     h->mb.i_cbp_chroma = 0;
 285
 286     /* Early termination: check variance of chroma residual before encoding.
 287      * Don't bother trying early termination at low QPs.
 288      * Values are experimentally derived. */
 289     if( b_decimate && i_qp >= (h->mb.b_trellis ? 12 : 18) )
 290     {
 291         int thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 292         int ssd[2];
 293         int score  = h->pixf.var2_8x8( h->mb.pic.p_fenc[1], FENC_STRIDE, h->mb.pic.p_fdec[1], FDEC_STRIDE, &ssd[0] );
 294             score += h->pixf.var2_8x8( h->mb.pic.p_fenc[2], FENC_STRIDE, h->mb.pic.p_fdec[2], FDEC_STRIDE, &ssd[1] );
 295         if( score < thresh*4 )
 296         {
 297             h->mb.cache.non_zero_count[x264_scan8[16]] = 0;
 298             h->mb.cache.non_zero_count[x264_scan8[17]] = 0;
 299             h->mb.cache.non_zero_count[x264_scan8[18]] = 0;
 300             h->mb.cache.non_zero_count[x264_scan8[19]] = 0;
 301             h->mb.cache.non_zero_count[x264_scan8[20]] = 0;
 302             h->mb.cache.non_zero_count[x264_scan8[21]] = 0;
 303             h->mb.cache.non_zero_count[x264_scan8[22]] = 0;
 304             h->mb.cache.non_zero_count[x264_scan8[23]] = 0;
 305             h->mb.cache.non_zero_count[x264_scan8[25]] = 0;
 306             h->mb.cache.non_zero_count[x264_scan8[26]] = 0;
 307             for( ch = 0; ch < 2; ch++ )
 308             {
 309                 if( ssd[ch] > thresh )
 310                 {
 311                     h->dctf.sub8x8_dct_dc( dct2x2, h->mb.pic.p_fenc[1+ch], h->mb.pic.p_fdec[1+ch] );
 312                     dct2x2dc_dconly( dct2x2 );
 313                     if( h->mb.b_trellis )
 314                         nz_dc = x264_quant_dc_trellis( h, (int16_t*)dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 315                     else
 316                         nz_dc = 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]<<
 317     1 );
 318                     if( nz_dc )
 319                     {
 320                         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 1;
 321                         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 322                         idct_dequant_2x2_dconly( dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 323                         h->dctf.add8x8_idct_dc( h->mb.pic.p_fdec[1+ch], dct2x2 );
 324                         h->mb.i_cbp_chroma = 1;
 325                     }
 326                 }
 327             }
 328             return;
 329         }
 330     }
 331
 332     for( ch = 0; ch < 2; ch++ )
 333     {
 334         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 335         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 336         int i_decimate_score = 0;
 337         int nz_ac = 0;
 338
 339         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 340
 341         if( h->mb.b_lossless )
 342         {
 343             for( i = 0; i < 4; i++ )
 344             {
 345                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 346                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 347                 nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od, &h->dct.chroma_dc[ch][i] );
 348                 h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 349                 h->mb.i_cbp_chroma |= nz;
 350             }
 351             h->mb.cache.non_zero_count[x264_scan8[25]+ch] = array_non_zero( h->dct.chroma_dc[ch] );
 352             continue;
 353         }
 354
 355         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 356         dct2x2dc( dct2x2, dct4x4 );
 357         /* calculate dct coeffs */
 358         for( i = 0; i < 4; i++ )
 359         {
 360             if( h->mb.b_trellis )
 361                 nz = x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 1, 0 );
 362             else
 363                 nz = h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[CQM_4IC+b_inter][i_qp], h->quant4_bias[CQM_4IC+b_inter][i_qp] );
 364             h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 365             if( nz )
 366             {
 367                 nz_ac = 1;
 368                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 369                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 370                 if( b_decimate )
 371                     i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[16+i+ch*4] );
 372             }
 373         }
 374
 375         if( h->mb.b_trellis )
 376             nz_dc = x264_quant_dc_trellis( h, (int16_t*)dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 377         else
 378             nz_dc = 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 );
 379
 380         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = nz_dc;
 381
 382         if( (b_decimate && i_decimate_score < 7) || !nz_ac )
 383         {
 384             /* Decimate the block */
 385             h->mb.cache.non_zero_count[x264_scan8[16+0]+24*ch] = 0;
 386             h->mb.cache.non_zero_count[x264_scan8[16+1]+24*ch] = 0;
 387             h->mb.cache.non_zero_count[x264_scan8[16+2]+24*ch] = 0;
 388             h->mb.cache.non_zero_count[x264_scan8[16+3]+24*ch] = 0;
 389             if( !nz_dc ) /* Whole block is empty */
 390                 continue;
 391             /* DC-only */
 392             zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 393             idct_dequant_2x2_dconly( dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 394             h->dctf.add8x8_idct_dc( p_dst, dct2x2 );
 395         }
 396         else
 397         {
 398             h->mb.i_cbp_chroma = 1;
 399             if( nz_dc )
 400             {
 401                 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 402                 idct_dequant_2x2_dc( dct2x2, dct4x4, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 403             }
 404             h->dctf.add8x8_idct( p_dst, dct4x4 );
 405         }
 406     }
 407
 408     if( h->mb.i_cbp_chroma )
 409         h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
 410     else if( h->mb.cache.non_zero_count[x264_scan8[25]] |
 411              h->mb.cache.non_zero_count[x264_scan8[26]] )
 412         h->mb.i_cbp_chroma = 1;    /* dc only */
 413 }
 414
 415 static void x264_macroblock_encode_skip( x264_t *h )
 416 {
 417     h->mb.i_cbp_luma = 0x00;
 418     h->mb.i_cbp_chroma = 0x00;
 419     memset( h->mb.cache.non_zero_count, 0, X264_SCAN8_SIZE );
 420     /* store cbp */
 421     h->mb.cbp[h->mb.i_mb_xy] = 0;
 422 }
 423
 424 /*****************************************************************************
 425  * x264_macroblock_encode_pskip:
 426  *  Encode an already marked skip block
 427  *****************************************************************************/
 428 static void x264_macroblock_encode_pskip( x264_t *h )
 429 {
 430     const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 431                                 h->mb.mv_min[0], h->mb.mv_max[0] );
 432     const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 433                                 h->mb.mv_min[1], h->mb.mv_max[1] );
 434
 435     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 436     if( !h->mb.b_skip_mc )
 437     {
 438         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 439                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 440                        mvx, mvy, 16, 16 );
 441
 442         h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 443                          h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 444                          mvx, mvy, 8, 8 );
 445
 446         h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 447                          h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 448                          mvx, mvy, 8, 8 );
 449     }
 450
 451     x264_macroblock_encode_skip( h );
 452 }
 453
 454 /*****************************************************************************
 455  * Intra prediction for predictive lossless mode.
 456  *****************************************************************************/
 457
 458 /* Note that these functions take a shortcut (mc.copy instead of actual pixel prediction) which assumes
 459  * that the edge pixels of the reconstructed frame are the same as that of the source frame.  This means
 460  * they will only work correctly if the neighboring blocks are losslessly coded.  In practice, this means
 461  * lossless mode cannot be mixed with lossy mode within a frame. */
 462 /* This can be resolved by explicitly copying the edge pixels after doing the mc.copy, but this doesn't
 463  * need to be done unless we decide to allow mixing lossless and lossy compression. */
 464
 465 void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
 466 {
 467     int stride = h->fenc->i_stride[1] << h->mb.b_interlaced;
 468     if( i_mode == I_PRED_CHROMA_V )
 469     {
 470         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-stride, stride, 8 );
 471         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-stride, stride, 8 );
 472     }
 473     else if( i_mode == I_PRED_CHROMA_H )
 474     {
 475         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-1, stride, 8 );
 476         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-1, stride, 8 );
 477     }
 478     else
 479     {
 480         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 481         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 482     }
 483 }
 484
 485 void x264_predict_lossless_4x4( x264_t *h, uint8_t *p_dst, int idx, int i_mode )
 486 {
 487     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 488     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
 489
 490     if( i_mode == I_PRED_4x4_V )
 491         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
 492     else if( i_mode == I_PRED_4x4_H )
 493         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-1, stride, 4 );
 494     else
 495         h->predict_4x4[i_mode]( p_dst );
 496 }
 497
 498 void x264_predict_lossless_8x8( x264_t *h, uint8_t *p_dst, int idx, int i_mode, uint8_t edge[33] )
 499 {
 500     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 501     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
 502
 503     if( i_mode == I_PRED_8x8_V )
 504         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
 505     else if( i_mode == I_PRED_8x8_H )
 506         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-1, stride, 8 );
 507     else
 508         h->predict_8x8[i_mode]( p_dst, edge );
 509 }
 510
 511 void x264_predict_lossless_16x16( x264_t *h, int i_mode )
 512 {
 513     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 514     if( i_mode == I_PRED_16x16_V )
 515         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-stride, stride, 16 );
 516     else if( i_mode == I_PRED_16x16_H )
 517         h->mc.copy_16x16_unaligned( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-1, stride, 16 );
 518     else
 519         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 520 }
 521
 522 /*****************************************************************************
 523  * x264_macroblock_encode:
 524  *****************************************************************************/
 525 void x264_macroblock_encode( x264_t *h )
 526 {
 527     int i_cbp_dc = 0;
 528     int i_qp = h->mb.i_qp;
 529     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 530     int b_force_no_skip = 0;
 531     int i,idx,nz;
 532     h->mb.i_cbp_luma = 0;
 533     h->mb.cache.non_zero_count[x264_scan8[24]] = 0;
 534
 535     if( h->sh.b_mbaff
 536         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 537         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 538     {
 539         /* The first skip is predicted to be a frame mb pair.
 540          * We don't yet support the aff part of mbaff, so force it to non-skip
 541          * so that we can pick the aff flag. */
 542         b_force_no_skip = 1;
 543         if( IS_SKIP(h->mb.i_type) )
 544         {
 545             if( h->mb.i_type == P_SKIP )
 546                 h->mb.i_type = P_L0;
 547             else if( h->mb.i_type == B_SKIP )
 548                 h->mb.i_type = B_DIRECT;
 549         }
 550     }
 551
 552     if( h->mb.i_type == P_SKIP )
 553     {
 554         /* A bit special */
 555         x264_macroblock_encode_pskip( h );
 556         return;
 557     }
 558     if( h->mb.i_type == B_SKIP )
 559     {
 560         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 561         if( !h->mb.b_skip_mc )
 562             x264_mb_mc( h );
 563         x264_macroblock_encode_skip( h );
 564         return;
 565     }
 566
 567     if( h->mb.i_type == I_16x16 )
 568     {
 569         const int i_mode = h->mb.i_intra16x16_pred_mode;
 570         h->mb.b_transform_8x8 = 0;
 571
 572         if( h->mb.b_lossless )
 573             x264_predict_lossless_16x16( h, i_mode );
 574         else
 575             h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 576
 577         /* encode the 16x16 macroblock */
 578         x264_mb_encode_i16x16( h, i_qp );
 579     }
 580     else if( h->mb.i_type == I_8x8 )
 581     {
 582         DECLARE_ALIGNED_16( uint8_t edge[33] );
 583         h->mb.b_transform_8x8 = 1;
 584         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 585         if( h->mb.i_skip_intra )
 586         {
 587             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 588             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = h->mb.pic.i8x8_nnz_buf[0];
 589             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = h->mb.pic.i8x8_nnz_buf[1];
 590             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = h->mb.pic.i8x8_nnz_buf[2];
 591             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = h->mb.pic.i8x8_nnz_buf[3];
 592             h->mb.i_cbp_luma = h->mb.pic.i8x8_cbp;
 593             /* In RD mode, restore the now-overwritten DCT data. */
 594             if( h->mb.i_skip_intra == 2 )
 595                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 596         }
 597         for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 598         {
 599             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 600             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 601             h->predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 602
 603             if( h->mb.b_lossless )
 604                 x264_predict_lossless_8x8( h, p_dst, i, i_mode, edge );
 605             else
 606                 h->predict_8x8[i_mode]( p_dst, edge );
 607
 608             x264_mb_encode_i8x8( h, i, i_qp );
 609         }
 610     }
 611     else if( h->mb.i_type == I_4x4 )
 612     {
 613         h->mb.b_transform_8x8 = 0;
 614         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 615         if( h->mb.i_skip_intra )
 616         {
 617             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 618             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = h->mb.pic.i4x4_nnz_buf[0];
 619             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = h->mb.pic.i4x4_nnz_buf[1];
 620             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = h->mb.pic.i4x4_nnz_buf[2];
 621             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = h->mb.pic.i4x4_nnz_buf[3];
 622             h->mb.i_cbp_luma = h->mb.pic.i4x4_cbp;
 623             /* In RD mode, restore the now-overwritten DCT data. */
 624             if( h->mb.i_skip_intra == 2 )
 625                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 626         }
 627         for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 628         {
 629             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
 630             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 631
 632             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 633                 /* emulate missing topright samples */
 634                 *(uint32_t*) &p_dst[4-FDEC_STRIDE] = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 635
 636             if( h->mb.b_lossless )
 637                 x264_predict_lossless_4x4( h, p_dst, i, i_mode );
 638             else
 639                 h->predict_4x4[i_mode]( p_dst );
 640             x264_mb_encode_i4x4( h, i, i_qp );
 641         }
 642     }
 643     else    /* Inter MB */
 644     {
 645         int i8x8, i4x4;
 646         int i_decimate_mb = 0;
 647
 648         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
 649         if( !h->mb.b_skip_mc )
 650             x264_mb_mc( h );
 651
 652         if( h->mb.b_lossless )
 653         {
 654             if( h->mb.b_transform_8x8 )
 655                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 656                 {
 657                     int x = 8*(i8x8&1);
 658                     int y = 8*(i8x8>>1);
 659                     nz = h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8],
 660                                         h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
 661                                         h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
 662                     STORE_8x8_NNZ(i8x8,nz);
 663                     h->mb.i_cbp_luma |= nz << i8x8;
 664                 }
 665             else
 666                 for( i4x4 = 0; i4x4 < 16; i4x4++ )
 667                 {
 668                     nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 669                                         h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
 670                                         h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4x4] );
 671                     h->mb.cache.non_zero_count[x264_scan8[i4x4]] = nz;
 672                     h->mb.i_cbp_luma |= nz << (i4x4>>2);
 673                 }
 674         }
 675         else if( h->mb.b_transform_8x8 )
 676         {
 677             DECLARE_ALIGNED_16( int16_t dct8x8[4][8][8] );
 678             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 679             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 680             h->nr_count[1] += h->mb.b_noise_reduction * 4;
 681
 682             for( idx = 0; idx < 4; idx++ )
 683             {
 684                 if( h->mb.b_noise_reduction )
 685                     h->quantf.denoise_dct( *dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
 686                 nz = x264_quant_8x8( h, dct8x8[idx], i_qp, 0, idx );
 687
 688                 if( nz )
 689                 {
 690                     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 691                     if( b_decimate )
 692                     {
 693                         int i_decimate_8x8 = h->quantf.decimate_score64( h->dct.luma8x8[idx] );
 694                         i_decimate_mb += i_decimate_8x8;
 695                         if( i_decimate_8x8 >= 4 )
 696                             h->mb.i_cbp_luma |= 1<<idx;
 697                     }
 698                     else
 699                         h->mb.i_cbp_luma |= 1<<idx;
 700                 }
 701             }
 702
 703             if( i_decimate_mb < 6 && b_decimate )
 704             {
 705                 h->mb.i_cbp_luma = 0;
 706                 *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = 0;
 707                 *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = 0;
 708                 *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = 0;
 709                 *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = 0;
 710             }
 711             else
 712             {
 713                 for( idx = 0; idx < 4; idx++ )
 714                 {
 715                     if( h->mb.i_cbp_luma&(1<<idx) )
 716                     {
 717                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 718                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 719                         STORE_8x8_NNZ(idx,1);
 720                     }
 721                     else
 722                         STORE_8x8_NNZ(idx,0);
 723                 }
 724             }
 725         }
 726         else
 727         {
 728             DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 729             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 730             h->nr_count[0] += h->mb.b_noise_reduction * 16;
 731
 732             for( i8x8 = 0; i8x8 < 4; i8x8++ )
 733             {
 734                 int i_decimate_8x8 = 0;
 735                 int cbp = 0;
 736
 737                 /* encode one 4x4 block */
 738                 for( i4x4 = 0; i4x4 < 4; i4x4++ )
 739                 {
 740                     idx = i8x8 * 4 + i4x4;
 741
 742                     if( h->mb.b_noise_reduction )
 743                         h->quantf.denoise_dct( *dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
 744                     nz = x264_quant_4x4( h, dct4x4[idx], i_qp, DCT_LUMA_4x4, 0, idx );
 745                     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 746
 747                     if( nz )
 748                     {
 749                         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 750                         h->quantf.dequant_4x4( dct4x4[idx], h->dequant4_mf[CQM_4PY], i_qp );
 751                         if( b_decimate && i_decimate_8x8 < 6 )
 752                             i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[idx] );
 753                         cbp = 1;
 754                     }
 755                 }
 756
 757                 /* decimate this 8x8 block */
 758                 i_decimate_mb += i_decimate_8x8;
 759                 if( b_decimate )
 760                 {
 761                     if( i_decimate_8x8 < 4 )
 762                         STORE_8x8_NNZ(i8x8,0)
 763                     else
 764                         h->mb.i_cbp_luma |= 1<<i8x8;
 765                 }
 766                 else if( cbp )
 767                 {
 768                     h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 769                     h->mb.i_cbp_luma |= 1<<i8x8;
 770                 }
 771             }
 772
 773             if( b_decimate )
 774             {
 775                 if( i_decimate_mb < 6 )
 776                 {
 777                     h->mb.i_cbp_luma = 0;
 778                     *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = 0;
 779                     *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = 0;
 780                     *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = 0;
 781                     *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = 0;
 782                 }
 783                 else
 784                 {
 785                     for( i8x8 = 0; i8x8 < 4; i8x8++ )
 786                         if( h->mb.i_cbp_luma&(1<<i8x8) )
 787                             h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 788                 }
 789             }
 790         }
 791     }
 792
 793     /* encode chroma */
 794     if( IS_INTRA( h->mb.i_type ) )
 795     {
 796         const int i_mode = h->mb.i_chroma_pred_mode;
 797         if( h->mb.b_lossless )
 798             x264_predict_lossless_8x8_chroma( h, i_mode );
 799         else
 800         {
 801             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 802             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 803         }
 804     }
 805
 806     /* encode the 8x8 blocks */
 807     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 808
 809     if( h->param.b_cabac )
 810     {
 811         i_cbp_dc = h->mb.cache.non_zero_count[x264_scan8[24]]
 812                  | h->mb.cache.non_zero_count[x264_scan8[25]] << 1
 813                  | h->mb.cache.non_zero_count[x264_scan8[26]] << 2;
 814     }
 815
 816     /* store cbp */
 817     h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
 818
 819     /* Check for P_SKIP
 820      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 821      *      (if multiple mv give same result)*/
 822     if( !b_force_no_skip )
 823     {
 824         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 825             !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
 826             *(uint32_t*)h->mb.cache.mv[0][x264_scan8[0]] == *(uint32_t*)h->mb.cache.pskip_mv
 827             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 828         {
 829             h->mb.i_type = P_SKIP;
 830         }
 831
 832         /* Check for B_SKIP */
 833         if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
 834         {
 835             h->mb.i_type = B_SKIP;
 836         }
 837     }
 838 }
 839
 840 /*****************************************************************************
 841  * x264_macroblock_probe_skip:
 842  *  Check if the current MB could be encoded as a [PB]_SKIP (it supposes you use
 843  *  the previous QP
 844  *****************************************************************************/
 845 int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
 846 {
 847     DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 848     DECLARE_ALIGNED_16( int16_t dct2x2[2][2] );
 849     DECLARE_ALIGNED_16( int16_t dctscan[16] );
 850
 851     int i_qp = h->mb.i_qp;
 852     int mvp[2];
 853     int ch, thresh, ssd;
 854
 855     int i8x8, i4x4;
 856     int i_decimate_mb;
 857
 858     if( !b_bidir )
 859     {
 860         /* Get the MV */
 861         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 862         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 863
 864         /* Motion compensation */
 865         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 866                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 867                        mvp[0], mvp[1], 16, 16 );
 868     }
 869
 870     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 871     {
 872         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
 873         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
 874         /* get luma diff */
 875         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
 876                                     h->mb.pic.p_fdec[0] + fdec_offset );
 877         /* encode one 4x4 block */
 878         for( i4x4 = 0; i4x4 < 4; i4x4++ )
 879         {
 880             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] ) )
 881                 continue;
 882             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 883             i_decimate_mb += h->quantf.decimate_score16( dctscan );
 884             if( i_decimate_mb >= 6 )
 885                 return 0;
 886         }
 887     }
 888
 889     /* encode chroma */
 890     i_qp = h->mb.i_chroma_qp;
 891     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 892
 893     for( ch = 0; ch < 2; ch++ )
 894     {
 895         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 896         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 897
 898         if( !b_bidir )
 899         {
 900             h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 901                              h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 902                              mvp[0], mvp[1], 8, 8 );
 903         }
 904
 905         /* there is almost never a termination during chroma, but we can't avoid the check entirely */
 906         /* so instead we check SSD and skip the actual check if the score is low enough. */
 907         ssd = h->pixf.ssd[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE );
 908         if( ssd < thresh )
 909             continue;
 910
 911         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 912
 913         /* calculate dct DC */
 914         dct2x2dc( dct2x2, dct4x4 );
 915         if( h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4PC][i_qp][0]>>1, h->quant4_bias[CQM_4PC][i_qp][0]<<1 ) )
 916             return 0;
 917
 918         /* If there wasn't a termination in DC, we can check against a much higher threshold. */
 919         if( ssd < thresh*4 )
 920             continue;
 921
 922         /* calculate dct coeffs */
 923         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 924         {
 925             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] ) )
 926                 continue;
 927             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 928             i_decimate_mb += h->quantf.decimate_score15( dctscan );
 929             if( i_decimate_mb >= 7 )
 930                 return 0;
 931         }
 932     }
 933
 934     h->mb.b_skip_mc = 1;
 935     return 1;
 936 }
 937
 938 /****************************************************************************
 939  * DCT-domain noise reduction / adaptive deadzone
 940  * from libavcodec
 941  ****************************************************************************/
 942
 943 void x264_noise_reduction_update( x264_t *h )
 944 {
 945     int cat, i;
 946     for( cat = 0; cat < 2; cat++ )
 947     {
 948         int size = cat ? 64 : 16;
 949         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
 950
 951         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
 952         {
 953             for( i = 0; i < size; i++ )
 954                 h->nr_residual_sum[cat][i] >>= 1;
 955             h->nr_count[cat] >>= 1;
 956         }
 957
 958         for( i = 0; i < size; i++ )
 959             h->nr_offset[cat][i] =
 960                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
 961                  + h->nr_residual_sum[cat][i]/2)
 962               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
 963     }
 964 }
 965
 966 /*****************************************************************************
 967  * RD only; 4 calls to this do not make up for one macroblock_encode.
 968  * doesn't transform chroma dc.
 969  *****************************************************************************/
 970 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
 971 {
 972     int i_qp = h->mb.i_qp;
 973     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
 974     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
 975     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 976     int nnz8x8 = 0;
 977     int ch, nz;
 978
 979     x264_mb_mc_8x8( h, i8 );
 980
 981     if( h->mb.b_lossless )
 982     {
 983         int i4;
 984         if( h->mb.b_transform_8x8 )
 985         {
 986             nnz8x8 = h->zigzagf.sub_8x8( h->dct.luma8x8[i8], p_fenc, p_fdec );
 987             STORE_8x8_NNZ(i8,nnz8x8);
 988         }
 989         else
 990         {
 991             for( i4 = i8*4; i4 < i8*4+4; i4++ )
 992             {
 993                 int nz;
 994                 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
 995                                     h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
 996                                     h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
 997                 h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
 998                 nnz8x8 |= nz;
 999             }
1000         }
1001         for( ch = 0; ch < 2; ch++ )
1002         {
1003             int16_t dc;
1004             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1005             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1006             nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec, &dc );
1007             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1008         }
1009     }
1010     else
1011     {
1012         if( h->mb.b_transform_8x8 )
1013         {
1014             DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
1015             h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
1016             nnz8x8 = x264_quant_8x8( h, dct8x8, i_qp, 0, i8 );
1017             if( nnz8x8 )
1018             {
1019                 h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
1020
1021                 if( b_decimate && !h->mb.b_trellis )
1022                     nnz8x8 = 4 <= h->quantf.decimate_score64( h->dct.luma8x8[i8] );
1023
1024                 if( nnz8x8 )
1025                 {
1026                     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
1027                     h->dctf.add8x8_idct8( p_fdec, dct8x8 );
1028                     STORE_8x8_NNZ(i8,1);
1029                 }
1030                 else
1031                     STORE_8x8_NNZ(i8,0);
1032             }
1033             else
1034                 STORE_8x8_NNZ(i8,0);
1035         }
1036         else
1037         {
1038             int i4;
1039             int i_decimate_8x8 = 0;
1040             DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
1041             h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
1042             for( i4 = 0; i4 < 4; i4++ )
1043             {
1044                 nz = x264_quant_4x4( h, dct4x4[i4], i_qp, DCT_LUMA_4x4, 0, i8*4+i4 );
1045                 h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = nz;
1046                 if( nz )
1047                 {
1048                     h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
1049                     h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
1050                     if( b_decimate )
1051                         i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[i8*4+i4] );
1052                     nnz8x8 = 1;
1053                 }
1054             }
1055
1056             if( b_decimate && i_decimate_8x8 < 4 )
1057                 nnz8x8 = 0;
1058
1059             if( nnz8x8 )
1060                 h->dctf.add8x8_idct( p_fdec, dct4x4 );
1061             else
1062                 STORE_8x8_NNZ(i8,0);
1063         }
1064
1065         i_qp = h->mb.i_chroma_qp;
1066
1067         for( ch = 0; ch < 2; ch++ )
1068         {
1069             DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
1070             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1071             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1072
1073             h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1074             dct4x4[0][0] = 0;
1075
1076             if( h->mb.b_trellis )
1077                 nz = x264_quant_4x4_trellis( h, dct4x4, CQM_4PC, i_qp, DCT_CHROMA_AC, 0, 1, 0 );
1078             else
1079                 nz = h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
1080
1081             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1082             if( nz )
1083             {
1084                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
1085                 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
1086                 h->dctf.add4x4_idct( p_fdec, dct4x4 );
1087             }
1088         }
1089     }
1090     h->mb.i_cbp_luma &= ~(1 << i8);
1091     h->mb.i_cbp_luma |= nnz8x8 << i8;
1092     h->mb.i_cbp_chroma = 0x02;
1093 }
1094
1095 /*****************************************************************************
1096  * RD only, luma only
1097  *****************************************************************************/
1098 void x264_macroblock_encode_p4x4( x264_t *h, int i4 )
1099 {
1100     int i_qp = h->mb.i_qp;
1101     uint8_t *p_fenc = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[i4]];
1102     uint8_t *p_fdec = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
1103     const int i_ref = h->mb.cache.ref[0][x264_scan8[i4]];
1104     const int mvx   = x264_clip3( h->mb.cache.mv[0][x264_scan8[i4]][0], h->mb.mv_min[0], h->mb.mv_max[0] );
1105     const int mvy   = x264_clip3( h->mb.cache.mv[0][x264_scan8[i4]][1], h->mb.mv_min[1], h->mb.mv_max[1] );
1106     int nz;
1107
1108     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 );
1109
1110     if( h->mb.b_lossless )
1111     {
1112         nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4], p_fenc, p_fdec );
1113         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1114     }
1115     else
1116     {
1117         DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
1118         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1119         nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 0, i4 );
1120         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1121         if( nz )
1122         {
1123             h->zigzagf.scan_4x4( h->dct.luma4x4[i4], dct4x4 );
1124             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PY], i_qp );
1125             h->dctf.add4x4_idct( p_fdec, dct4x4 );
1126         }
1127     }
1128 }