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*2+y];
  31 static inline void zigzag_scan_2x2_dc( int16_t level[4], int16_t dct[4] )
  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] + dct[1]; \
  42     int d1 = dct[2] + dct[3]; \
  43     int d2 = dct[0] - dct[1]; \
  44     int d3 = dct[2] - dct[3]; \
  45     int dmf = dequant_mf[i_qp%6][0] << i_qp/6;
  46
  47 static inline void idct_dequant_2x2_dc( int16_t dct[4], int16_t dct4x4[4][16], int dequant_mf[6][16], int i_qp )
  48 {
  49     IDCT_DEQUANT_START
  50     dct4x4[0][0] = (d0 + d1) * dmf >> 5;
  51     dct4x4[1][0] = (d0 - d1) * dmf >> 5;
  52     dct4x4[2][0] = (d2 + d3) * dmf >> 5;
  53     dct4x4[3][0] = (d2 - d3) * dmf >> 5;
  54 }
  55
  56 static inline void idct_dequant_2x2_dconly( int16_t out[4], int16_t dct[4], int dequant_mf[6][16], int i_qp )
  57 {
  58     IDCT_DEQUANT_START
  59     out[0] = (d0 + d1) * dmf >> 5;
  60     out[1] = (d0 - d1) * dmf >> 5;
  61     out[2] = (d2 + d3) * dmf >> 5;
  62     out[3] = (d2 - d3) * dmf >> 5;
  63 }
  64
  65 static inline void dct2x2dc( int16_t d[4], int16_t dct4x4[4][16] )
  66 {
  67     int d0 = dct4x4[0][0] + dct4x4[1][0];
  68     int d1 = dct4x4[2][0] + dct4x4[3][0];
  69     int d2 = dct4x4[0][0] - dct4x4[1][0];
  70     int d3 = dct4x4[2][0] - dct4x4[3][0];
  71     d[0] = d0 + d1;
  72     d[2] = d2 + d3;
  73     d[1] = d0 - d1;
  74     d[3] = d2 - d3;
  75     dct4x4[0][0] = 0;
  76     dct4x4[1][0] = 0;
  77     dct4x4[2][0] = 0;
  78     dct4x4[3][0] = 0;
  79 }
  80
  81 static inline void dct2x2dc_dconly( int16_t d[4] )
  82 {
  83     int d0 = d[0] + d[1];
  84     int d1 = d[2] + d[3];
  85     int d2 = d[0] - d[1];
  86     int d3 = d[2] - d[3];
  87     d[0] = d0 + d1;
  88     d[2] = d2 + d3;
  89     d[1] = d0 - d1;
  90     d[3] = d2 - d3;
  91 }
  92
  93 static ALWAYS_INLINE int x264_quant_4x4( x264_t *h, int16_t dct[16], int i_qp, int i_ctxBlockCat, int b_intra, int idx )
  94 {
  95     int i_quant_cat = b_intra ? CQM_4IY : CQM_4PY;
  96     if( h->mb.b_trellis )
  97         return x264_quant_4x4_trellis( h, dct, i_quant_cat, i_qp, i_ctxBlockCat, b_intra, 0, idx );
  98     else
  99         return h->quantf.quant_4x4( dct, h->quant4_mf[i_quant_cat][i_qp], h->quant4_bias[i_quant_cat][i_qp] );
 100 }
 101
 102 static ALWAYS_INLINE int x264_quant_8x8( x264_t *h, int16_t dct[64], int i_qp, int b_intra, int idx )
 103 {
 104     int i_quant_cat = b_intra ? CQM_8IY : CQM_8PY;
 105     if( h->mb.b_trellis )
 106         return x264_quant_8x8_trellis( h, dct, i_quant_cat, i_qp, b_intra, idx );
 107     else
 108         return h->quantf.quant_8x8( dct, h->quant8_mf[i_quant_cat][i_qp], h->quant8_bias[i_quant_cat][i_qp] );
 109 }
 110
 111 /* All encoding functions must output the correct CBP and NNZ values.
 112  * The entropy coding functions will check CBP first, then NNZ, before
 113  * actually reading the DCT coefficients.  NNZ still must be correct even
 114  * if CBP is zero because of the use of NNZ values for context selection.
 115  * "NNZ" need only be 0 or 1 rather than the exact coefficient count because
 116  * that is only needed in CAVLC, and will be calculated by CAVLC's residual
 117  * coding and stored as necessary. */
 118
 119 /* This means that decimation can be done merely by adjusting the CBP and NNZ
 120  * rather than memsetting the coefficients. */
 121
 122 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qp )
 123 {
 124     int nz;
 125     uint8_t *p_src = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[idx]];
 126     uint8_t *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[idx]];
 127     ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
 128
 129     if( h->mb.b_lossless )
 130     {
 131         nz = h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
 132         h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 133         h->mb.i_cbp_luma |= nz<<(idx>>2);
 134         return;
 135     }
 136
 137     h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
 138
 139     nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 1, idx );
 140     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 141     if( nz )
 142     {
 143         h->mb.i_cbp_luma |= 1<<(idx>>2);
 144         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4 );
 145         h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qp );
 146         h->dctf.add4x4_idct( p_dst, dct4x4 );
 147     }
 148 }
 149
 150 #define STORE_8x8_NNZ(idx,nz)\
 151 {\
 152     M16( &h->mb.cache.non_zero_count[x264_scan8[idx*4+0]] ) = nz * 0x0101;\
 153     M16( &h->mb.cache.non_zero_count[x264_scan8[idx*4+2]] ) = nz * 0x0101;\
 154 }
 155
 156 #define CLEAR_16x16_NNZ \
 157 {\
 158     M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = 0;\
 159     M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = 0;\
 160     M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = 0;\
 161     M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = 0;\
 162 }
 163
 164 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qp )
 165 {
 166     int x = 8 * (idx&1);
 167     int y = 8 * (idx>>1);
 168     int nz;
 169     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
 170     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
 171     ALIGNED_ARRAY_16( int16_t, dct8x8,[64] );
 172
 173     if( h->mb.b_lossless )
 174     {
 175         nz = h->zigzagf.sub_8x8( h->dct.luma8x8[idx], p_src, p_dst );
 176         STORE_8x8_NNZ(idx,nz);
 177         h->mb.i_cbp_luma |= nz<<idx;
 178         return;
 179     }
 180
 181     h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
 182
 183     nz = x264_quant_8x8( h, dct8x8, i_qp, 1, idx );
 184     if( nz )
 185     {
 186         h->mb.i_cbp_luma |= 1<<idx;
 187         h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
 188         h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qp );
 189         h->dctf.add8x8_idct8( p_dst, dct8x8 );
 190         STORE_8x8_NNZ(idx,1);
 191     }
 192     else
 193         STORE_8x8_NNZ(idx,0);
 194 }
 195
 196 static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
 197 {
 198     uint8_t  *p_src = h->mb.pic.p_fenc[0];
 199     uint8_t  *p_dst = h->mb.pic.p_fdec[0];
 200
 201     ALIGNED_ARRAY_16( int16_t, dct4x4,[16],[16] );
 202     ALIGNED_ARRAY_16( int16_t, dct_dc4x4,[16] );
 203
 204     int i, nz;
 205     int decimate_score = h->mb.b_dct_decimate ? 0 : 9;
 206
 207     if( h->mb.b_lossless )
 208     {
 209         for( i = 0; i < 16; i++ )
 210         {
 211             int oe = block_idx_xy_fenc[i];
 212             int od = block_idx_xy_fdec[i];
 213             nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[i], p_src+oe, p_dst+od, &dct_dc4x4[block_idx_yx_1d[i]] );
 214             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 215             h->mb.i_cbp_luma |= nz;
 216         }
 217         h->mb.i_cbp_luma *= 0xf;
 218         h->mb.cache.non_zero_count[x264_scan8[24]] = array_non_zero( dct_dc4x4 );
 219         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 220         return;
 221     }
 222
 223     h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
 224
 225     for( i = 0; i < 16; i++ )
 226     {
 227         /* copy dc coeff */
 228         dct_dc4x4[block_idx_xy_1d[i]] = dct4x4[i][0];
 229         dct4x4[i][0] = 0;
 230
 231         /* quant/scan/dequant */
 232         nz = x264_quant_4x4( h, dct4x4[i], i_qp, DCT_LUMA_AC, 1, i );
 233         h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 234         if( nz )
 235         {
 236             h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[i] );
 237             h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IY], i_qp );
 238             if( decimate_score < 6 ) decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[i] );
 239             h->mb.i_cbp_luma = 0xf;
 240         }
 241     }
 242
 243     /* Writing the 16 CBFs in an i16x16 block is quite costly, so decimation can save many bits. */
 244     /* More useful with CAVLC, but still useful with CABAC. */
 245     if( decimate_score < 6 )
 246     {
 247         h->mb.i_cbp_luma = 0;
 248         CLEAR_16x16_NNZ
 249     }
 250
 251     h->dctf.dct4x4dc( dct_dc4x4 );
 252     if( h->mb.b_trellis )
 253         nz = x264_quant_dc_trellis( h, dct_dc4x4, CQM_4IY, i_qp, DCT_LUMA_DC, 1, 0 );
 254     else
 255         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 );
 256
 257     h->mb.cache.non_zero_count[x264_scan8[24]] = nz;
 258     if( nz )
 259     {
 260         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 261
 262         /* output samples to fdec */
 263         h->dctf.idct4x4dc( dct_dc4x4 );
 264         h->quantf.dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[CQM_4IY], i_qp );  /* XXX not inversed */
 265         if( h->mb.i_cbp_luma )
 266             for( i = 0; i < 16; i++ )
 267                 dct4x4[i][0] = dct_dc4x4[block_idx_xy_1d[i]];
 268     }
 269
 270     /* put pixels to fdec */
 271     if( h->mb.i_cbp_luma )
 272         h->dctf.add16x16_idct( p_dst, dct4x4 );
 273     else if( nz )
 274         h->dctf.add16x16_idct_dc( p_dst, dct_dc4x4 );
 275 }
 276
 277 static inline int idct_dequant_round_2x2_dc( int16_t ref[4], int16_t dct[4], int dequant_mf[6][16], int i_qp )
 278 {
 279     int16_t out[4];
 280     idct_dequant_2x2_dconly( out, dct, dequant_mf, i_qp );
 281     return ((ref[0] ^ (out[0]+32))
 282           | (ref[1] ^ (out[1]+32))
 283           | (ref[2] ^ (out[2]+32))
 284           | (ref[3] ^ (out[3]+32))) >> 6;
 285 }
 286
 287 /* Round down coefficients losslessly in DC-only chroma blocks.
 288  * Unlike luma blocks, this can't be done with a lookup table or
 289  * other shortcut technique because of the interdependencies
 290  * between the coefficients due to the chroma DC transform. */
 291 static inline int x264_mb_optimize_chroma_dc( x264_t *h, int b_inter, int i_qp, int16_t dct2x2[4] )
 292 {
 293     int16_t dct2x2_orig[4];
 294     int coeff;
 295     int nz = 0;
 296
 297     /* If the QP is too high, there's no benefit to rounding optimization. */
 298     if( h->dequant4_mf[CQM_4IC + b_inter][i_qp%6][0] << (i_qp/6) > 32*64 )
 299         return 1;
 300
 301     idct_dequant_2x2_dconly( dct2x2_orig, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 302     dct2x2_orig[0] += 32;
 303     dct2x2_orig[1] += 32;
 304     dct2x2_orig[2] += 32;
 305     dct2x2_orig[3] += 32;
 306
 307     /* If the DC coefficients already round to zero, terminate early. */
 308     if( !((dct2x2_orig[0]|dct2x2_orig[1]|dct2x2_orig[2]|dct2x2_orig[3])>>6) )
 309         return 0;
 310
 311     /* Start with the highest frequency coefficient... is this the best option? */
 312     for( coeff = 3; coeff >= 0; coeff-- )
 313     {
 314         int sign = dct2x2[coeff] < 0 ? -1 : 1;
 315         int level = dct2x2[coeff];
 316
 317         if( !level )
 318             continue;
 319
 320         while( level )
 321         {
 322             dct2x2[coeff] = level - sign;
 323             if( idct_dequant_round_2x2_dc( dct2x2_orig, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp ) )
 324                 break;
 325             level -= sign;
 326         }
 327
 328         nz |= level;
 329         dct2x2[coeff] = level;
 330     }
 331
 332     return !!nz;
 333 }
 334
 335 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
 336 {
 337     int i, ch, nz, nz_dc;
 338     int b_decimate = b_inter && h->mb.b_dct_decimate;
 339     ALIGNED_ARRAY_16( int16_t, dct2x2,[4] );
 340     h->mb.i_cbp_chroma = 0;
 341
 342     /* Early termination: check variance of chroma residual before encoding.
 343      * Don't bother trying early termination at low QPs.
 344      * Values are experimentally derived. */
 345     if( b_decimate && i_qp >= (h->mb.b_trellis ? 12 : 18) )
 346     {
 347         int thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 348         int ssd[2];
 349         int score  = h->pixf.var2_8x8( h->mb.pic.p_fenc[1], FENC_STRIDE, h->mb.pic.p_fdec[1], FDEC_STRIDE, &ssd[0] );
 350             score += h->pixf.var2_8x8( h->mb.pic.p_fenc[2], FENC_STRIDE, h->mb.pic.p_fdec[2], FDEC_STRIDE, &ssd[1] );
 351         if( score < thresh*4 )
 352         {
 353             h->mb.cache.non_zero_count[x264_scan8[16]] = 0;
 354             h->mb.cache.non_zero_count[x264_scan8[17]] = 0;
 355             h->mb.cache.non_zero_count[x264_scan8[18]] = 0;
 356             h->mb.cache.non_zero_count[x264_scan8[19]] = 0;
 357             h->mb.cache.non_zero_count[x264_scan8[20]] = 0;
 358             h->mb.cache.non_zero_count[x264_scan8[21]] = 0;
 359             h->mb.cache.non_zero_count[x264_scan8[22]] = 0;
 360             h->mb.cache.non_zero_count[x264_scan8[23]] = 0;
 361             h->mb.cache.non_zero_count[x264_scan8[25]] = 0;
 362             h->mb.cache.non_zero_count[x264_scan8[26]] = 0;
 363             for( ch = 0; ch < 2; ch++ )
 364             {
 365                 if( ssd[ch] > thresh )
 366                 {
 367                     h->dctf.sub8x8_dct_dc( dct2x2, h->mb.pic.p_fenc[1+ch], h->mb.pic.p_fdec[1+ch] );
 368                     if( h->mb.b_trellis )
 369                         nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 370                     else
 371                         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]<<
 372     1 );
 373
 374                     if( nz_dc )
 375                     {
 376                         if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
 377                             continue;
 378                         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 1;
 379                         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 380                         idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 381                         h->dctf.add8x8_idct_dc( h->mb.pic.p_fdec[1+ch], dct2x2 );
 382                         h->mb.i_cbp_chroma = 1;
 383                     }
 384                 }
 385             }
 386             return;
 387         }
 388     }
 389
 390     for( ch = 0; ch < 2; ch++ )
 391     {
 392         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 393         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 394         int i_decimate_score = 0;
 395         int nz_ac = 0;
 396
 397         ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
 398
 399         if( h->mb.b_lossless )
 400         {
 401             for( i = 0; i < 4; i++ )
 402             {
 403                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 404                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 405                 nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od, &h->dct.chroma_dc[ch][i] );
 406                 h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 407                 h->mb.i_cbp_chroma |= nz;
 408             }
 409             h->mb.cache.non_zero_count[x264_scan8[25]+ch] = array_non_zero( h->dct.chroma_dc[ch] );
 410             continue;
 411         }
 412
 413         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 414         dct2x2dc( dct2x2, dct4x4 );
 415         /* calculate dct coeffs */
 416         for( i = 0; i < 4; i++ )
 417         {
 418             if( h->mb.b_trellis )
 419                 nz = x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 1, 0 );
 420             else
 421                 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] );
 422             h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 423             if( nz )
 424             {
 425                 nz_ac = 1;
 426                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 427                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 428                 if( b_decimate )
 429                     i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[16+i+ch*4] );
 430             }
 431         }
 432
 433         if( h->mb.b_trellis )
 434             nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 435         else
 436             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 );
 437
 438         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = nz_dc;
 439
 440         if( (b_decimate && i_decimate_score < 7) || !nz_ac )
 441         {
 442             /* Decimate the block */
 443             h->mb.cache.non_zero_count[x264_scan8[16+0]+24*ch] = 0;
 444             h->mb.cache.non_zero_count[x264_scan8[16+1]+24*ch] = 0;
 445             h->mb.cache.non_zero_count[x264_scan8[16+2]+24*ch] = 0;
 446             h->mb.cache.non_zero_count[x264_scan8[16+3]+24*ch] = 0;
 447             if( !nz_dc ) /* Whole block is empty */
 448                 continue;
 449             if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
 450             {
 451                 h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 0;
 452                 continue;
 453             }
 454             /* DC-only */
 455             zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 456             idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 457             h->dctf.add8x8_idct_dc( p_dst, dct2x2 );
 458         }
 459         else
 460         {
 461             h->mb.i_cbp_chroma = 1;
 462             if( nz_dc )
 463             {
 464                 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 465                 idct_dequant_2x2_dc( dct2x2, dct4x4, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 466             }
 467             h->dctf.add8x8_idct( p_dst, dct4x4 );
 468         }
 469     }
 470
 471     if( h->mb.i_cbp_chroma )
 472         h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
 473     else if( h->mb.cache.non_zero_count[x264_scan8[25]] |
 474              h->mb.cache.non_zero_count[x264_scan8[26]] )
 475         h->mb.i_cbp_chroma = 1;    /* dc only */
 476 }
 477
 478 static void x264_macroblock_encode_skip( x264_t *h )
 479 {
 480     h->mb.i_cbp_luma = 0x00;
 481     h->mb.i_cbp_chroma = 0x00;
 482     memset( h->mb.cache.non_zero_count, 0, X264_SCAN8_SIZE );
 483     /* store cbp */
 484     h->mb.cbp[h->mb.i_mb_xy] = 0;
 485 }
 486
 487 /*****************************************************************************
 488  * x264_macroblock_encode_pskip:
 489  *  Encode an already marked skip block
 490  *****************************************************************************/
 491 static void x264_macroblock_encode_pskip( x264_t *h )
 492 {
 493     const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 494                                 h->mb.mv_min[0], h->mb.mv_max[0] );
 495     const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 496                                 h->mb.mv_min[1], h->mb.mv_max[1] );
 497
 498     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 499     if( !h->mb.b_skip_mc )
 500     {
 501         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 502                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 503                        mvx, mvy, 16, 16, &h->sh.weight[0][0] );
 504
 505         h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 506                          h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 507                          mvx, mvy, 8, 8 );
 508
 509         if( h->sh.weight[0][1].weightfn )
 510             h->sh.weight[0][1].weightfn[8>>2]( h->mb.pic.p_fdec[1], FDEC_STRIDE,
 511                                                h->mb.pic.p_fdec[1], FDEC_STRIDE,
 512                                                &h->sh.weight[0][1], 8 );
 513
 514         h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 515                          h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 516                          mvx, mvy, 8, 8 );
 517
 518         if( h->sh.weight[0][2].weightfn )
 519             h->sh.weight[0][2].weightfn[8>>2]( h->mb.pic.p_fdec[2], FDEC_STRIDE,
 520                                                h->mb.pic.p_fdec[2], FDEC_STRIDE,
 521                                                &h->sh.weight[0][2], 8 );
 522     }
 523
 524     x264_macroblock_encode_skip( h );
 525 }
 526
 527 /*****************************************************************************
 528  * Intra prediction for predictive lossless mode.
 529  *****************************************************************************/
 530
 531 /* Note that these functions take a shortcut (mc.copy instead of actual pixel prediction) which assumes
 532  * that the edge pixels of the reconstructed frame are the same as that of the source frame.  This means
 533  * they will only work correctly if the neighboring blocks are losslessly coded.  In practice, this means
 534  * lossless mode cannot be mixed with lossy mode within a frame. */
 535 /* This can be resolved by explicitly copying the edge pixels after doing the mc.copy, but this doesn't
 536  * need to be done unless we decide to allow mixing lossless and lossy compression. */
 537
 538 void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
 539 {
 540     int stride = h->fenc->i_stride[1] << h->mb.b_interlaced;
 541     if( i_mode == I_PRED_CHROMA_V )
 542     {
 543         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-stride, stride, 8 );
 544         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-stride, stride, 8 );
 545     }
 546     else if( i_mode == I_PRED_CHROMA_H )
 547     {
 548         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-1, stride, 8 );
 549         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-1, stride, 8 );
 550     }
 551     else
 552     {
 553         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 554         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 555     }
 556 }
 557
 558 void x264_predict_lossless_4x4( x264_t *h, uint8_t *p_dst, int idx, int i_mode )
 559 {
 560     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 561     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
 562
 563     if( i_mode == I_PRED_4x4_V )
 564         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
 565     else if( i_mode == I_PRED_4x4_H )
 566         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-1, stride, 4 );
 567     else
 568         h->predict_4x4[i_mode]( p_dst );
 569 }
 570
 571 void x264_predict_lossless_8x8( x264_t *h, uint8_t *p_dst, int idx, int i_mode, uint8_t edge[33] )
 572 {
 573     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 574     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
 575
 576     if( i_mode == I_PRED_8x8_V )
 577         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
 578     else if( i_mode == I_PRED_8x8_H )
 579         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-1, stride, 8 );
 580     else
 581         h->predict_8x8[i_mode]( p_dst, edge );
 582 }
 583
 584 void x264_predict_lossless_16x16( x264_t *h, int i_mode )
 585 {
 586     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 587     if( i_mode == I_PRED_16x16_V )
 588         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-stride, stride, 16 );
 589     else if( i_mode == I_PRED_16x16_H )
 590         h->mc.copy_16x16_unaligned( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-1, stride, 16 );
 591     else
 592         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 593 }
 594
 595 /*****************************************************************************
 596  * x264_macroblock_encode:
 597  *****************************************************************************/
 598 void x264_macroblock_encode( x264_t *h )
 599 {
 600     int i_cbp_dc = 0;
 601     int i_qp = h->mb.i_qp;
 602     int b_decimate = h->mb.b_dct_decimate;
 603     int b_force_no_skip = 0;
 604     int i,idx,nz;
 605     h->mb.i_cbp_luma = 0;
 606     h->mb.cache.non_zero_count[x264_scan8[24]] = 0;
 607
 608     if( h->sh.b_mbaff
 609         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 610         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 611     {
 612         /* The first skip is predicted to be a frame mb pair.
 613          * We don't yet support the aff part of mbaff, so force it to non-skip
 614          * so that we can pick the aff flag. */
 615         b_force_no_skip = 1;
 616         if( IS_SKIP(h->mb.i_type) )
 617         {
 618             if( h->mb.i_type == P_SKIP )
 619                 h->mb.i_type = P_L0;
 620             else if( h->mb.i_type == B_SKIP )
 621                 h->mb.i_type = B_DIRECT;
 622         }
 623     }
 624
 625     if( h->mb.i_type == P_SKIP )
 626     {
 627         /* A bit special */
 628         x264_macroblock_encode_pskip( h );
 629         return;
 630     }
 631     if( h->mb.i_type == B_SKIP )
 632     {
 633         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 634         if( !h->mb.b_skip_mc )
 635             x264_mb_mc( h );
 636         x264_macroblock_encode_skip( h );
 637         return;
 638     }
 639
 640     if( h->mb.i_type == I_16x16 )
 641     {
 642         const int i_mode = h->mb.i_intra16x16_pred_mode;
 643         h->mb.b_transform_8x8 = 0;
 644
 645         if( h->mb.b_lossless )
 646             x264_predict_lossless_16x16( h, i_mode );
 647         else
 648             h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 649
 650         /* encode the 16x16 macroblock */
 651         x264_mb_encode_i16x16( h, i_qp );
 652     }
 653     else if( h->mb.i_type == I_8x8 )
 654     {
 655         ALIGNED_ARRAY_16( uint8_t, edge,[33] );
 656         h->mb.b_transform_8x8 = 1;
 657         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 658         if( h->mb.i_skip_intra )
 659         {
 660             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 661             M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i8x8_nnz_buf[0];
 662             M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i8x8_nnz_buf[1];
 663             M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i8x8_nnz_buf[2];
 664             M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i8x8_nnz_buf[3];
 665             h->mb.i_cbp_luma = h->mb.pic.i8x8_cbp;
 666             /* In RD mode, restore the now-overwritten DCT data. */
 667             if( h->mb.i_skip_intra == 2 )
 668                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 669         }
 670         for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 671         {
 672             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 673             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 674             h->predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 675
 676             if( h->mb.b_lossless )
 677                 x264_predict_lossless_8x8( h, p_dst, i, i_mode, edge );
 678             else
 679                 h->predict_8x8[i_mode]( p_dst, edge );
 680
 681             x264_mb_encode_i8x8( h, i, i_qp );
 682         }
 683     }
 684     else if( h->mb.i_type == I_4x4 )
 685     {
 686         h->mb.b_transform_8x8 = 0;
 687         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 688         if( h->mb.i_skip_intra )
 689         {
 690             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 691             M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i4x4_nnz_buf[0];
 692             M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i4x4_nnz_buf[1];
 693             M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i4x4_nnz_buf[2];
 694             M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i4x4_nnz_buf[3];
 695             h->mb.i_cbp_luma = h->mb.pic.i4x4_cbp;
 696             /* In RD mode, restore the now-overwritten DCT data. */
 697             if( h->mb.i_skip_intra == 2 )
 698                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 699         }
 700         for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 701         {
 702             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
 703             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 704
 705             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 706                 /* emulate missing topright samples */
 707                 M32( &p_dst[4-FDEC_STRIDE] ) = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 708
 709             if( h->mb.b_lossless )
 710                 x264_predict_lossless_4x4( h, p_dst, i, i_mode );
 711             else
 712                 h->predict_4x4[i_mode]( p_dst );
 713             x264_mb_encode_i4x4( h, i, i_qp );
 714         }
 715     }
 716     else    /* Inter MB */
 717     {
 718         int i8x8, i4x4;
 719         int i_decimate_mb = 0;
 720
 721         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
 722         if( !h->mb.b_skip_mc )
 723             x264_mb_mc( h );
 724
 725         if( h->mb.b_lossless )
 726         {
 727             if( h->mb.b_transform_8x8 )
 728                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 729                 {
 730                     int x = 8*(i8x8&1);
 731                     int y = 8*(i8x8>>1);
 732                     nz = h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8],
 733                                         h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
 734                                         h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
 735                     STORE_8x8_NNZ(i8x8,nz);
 736                     h->mb.i_cbp_luma |= nz << i8x8;
 737                 }
 738             else
 739                 for( i4x4 = 0; i4x4 < 16; i4x4++ )
 740                 {
 741                     nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 742                                         h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
 743                                         h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4x4] );
 744                     h->mb.cache.non_zero_count[x264_scan8[i4x4]] = nz;
 745                     h->mb.i_cbp_luma |= nz << (i4x4>>2);
 746                 }
 747         }
 748         else if( h->mb.b_transform_8x8 )
 749         {
 750             ALIGNED_ARRAY_16( int16_t, dct8x8,[4],[64] );
 751             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 752             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 753             h->nr_count[1] += h->mb.b_noise_reduction * 4;
 754
 755             for( idx = 0; idx < 4; idx++ )
 756             {
 757                 if( h->mb.b_noise_reduction )
 758                     h->quantf.denoise_dct( dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
 759                 nz = x264_quant_8x8( h, dct8x8[idx], i_qp, 0, idx );
 760
 761                 if( nz )
 762                 {
 763                     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 764                     if( b_decimate )
 765                     {
 766                         int i_decimate_8x8 = h->quantf.decimate_score64( h->dct.luma8x8[idx] );
 767                         i_decimate_mb += i_decimate_8x8;
 768                         if( i_decimate_8x8 >= 4 )
 769                             h->mb.i_cbp_luma |= 1<<idx;
 770                     }
 771                     else
 772                         h->mb.i_cbp_luma |= 1<<idx;
 773                 }
 774             }
 775
 776             if( i_decimate_mb < 6 && b_decimate )
 777             {
 778                 h->mb.i_cbp_luma = 0;
 779                 CLEAR_16x16_NNZ
 780             }
 781             else
 782             {
 783                 for( idx = 0; idx < 4; idx++ )
 784                 {
 785                     if( h->mb.i_cbp_luma&(1<<idx) )
 786                     {
 787                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 788                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 789                         STORE_8x8_NNZ(idx,1);
 790                     }
 791                     else
 792                         STORE_8x8_NNZ(idx,0);
 793                 }
 794             }
 795         }
 796         else
 797         {
 798             ALIGNED_ARRAY_16( int16_t, dct4x4,[16],[16] );
 799             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 800             h->nr_count[0] += h->mb.b_noise_reduction * 16;
 801
 802             for( i8x8 = 0; i8x8 < 4; i8x8++ )
 803             {
 804                 int i_decimate_8x8 = 0;
 805                 int cbp = 0;
 806
 807                 /* encode one 4x4 block */
 808                 for( i4x4 = 0; i4x4 < 4; i4x4++ )
 809                 {
 810                     idx = i8x8 * 4 + i4x4;
 811
 812                     if( h->mb.b_noise_reduction )
 813                         h->quantf.denoise_dct( dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
 814                     nz = x264_quant_4x4( h, dct4x4[idx], i_qp, DCT_LUMA_4x4, 0, idx );
 815                     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 816
 817                     if( nz )
 818                     {
 819                         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 820                         h->quantf.dequant_4x4( dct4x4[idx], h->dequant4_mf[CQM_4PY], i_qp );
 821                         if( b_decimate && i_decimate_8x8 < 6 )
 822                             i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[idx] );
 823                         cbp = 1;
 824                     }
 825                 }
 826
 827                 /* decimate this 8x8 block */
 828                 i_decimate_mb += i_decimate_8x8;
 829                 if( b_decimate )
 830                 {
 831                     if( i_decimate_8x8 < 4 )
 832                         STORE_8x8_NNZ(i8x8,0)
 833                     else
 834                         h->mb.i_cbp_luma |= 1<<i8x8;
 835                 }
 836                 else if( cbp )
 837                 {
 838                     h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 839                     h->mb.i_cbp_luma |= 1<<i8x8;
 840                 }
 841             }
 842
 843             if( b_decimate )
 844             {
 845                 if( i_decimate_mb < 6 )
 846                 {
 847                     h->mb.i_cbp_luma = 0;
 848                     CLEAR_16x16_NNZ
 849                 }
 850                 else
 851                 {
 852                     for( i8x8 = 0; i8x8 < 4; i8x8++ )
 853                         if( h->mb.i_cbp_luma&(1<<i8x8) )
 854                             h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 855                 }
 856             }
 857         }
 858     }
 859
 860     /* encode chroma */
 861     if( IS_INTRA( h->mb.i_type ) )
 862     {
 863         const int i_mode = h->mb.i_chroma_pred_mode;
 864         if( h->mb.b_lossless )
 865             x264_predict_lossless_8x8_chroma( h, i_mode );
 866         else
 867         {
 868             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 869             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 870         }
 871     }
 872
 873     /* encode the 8x8 blocks */
 874     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 875
 876     if( h->param.b_cabac )
 877     {
 878         i_cbp_dc = h->mb.cache.non_zero_count[x264_scan8[24]]
 879                  | h->mb.cache.non_zero_count[x264_scan8[25]] << 1
 880                  | h->mb.cache.non_zero_count[x264_scan8[26]] << 2;
 881     }
 882
 883     /* store cbp */
 884     h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
 885
 886     /* Check for P_SKIP
 887      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 888      *      (if multiple mv give same result)*/
 889     if( !b_force_no_skip )
 890     {
 891         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 892             !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
 893             M32( h->mb.cache.mv[0][x264_scan8[0]] ) == M32( h->mb.cache.pskip_mv )
 894             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 895         {
 896             h->mb.i_type = P_SKIP;
 897         }
 898
 899         /* Check for B_SKIP */
 900         if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
 901         {
 902             h->mb.i_type = B_SKIP;
 903         }
 904     }
 905 }
 906
 907 /*****************************************************************************
 908  * x264_macroblock_probe_skip:
 909  *  Check if the current MB could be encoded as a [PB]_SKIP
 910  *****************************************************************************/
 911 int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
 912 {
 913     ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
 914     ALIGNED_ARRAY_16( int16_t, dct2x2,[4] );
 915     ALIGNED_ARRAY_16( int16_t, dctscan,[16] );
 916
 917     int i_qp = h->mb.i_qp;
 918     int mvp[2];
 919     int ch, thresh, ssd;
 920
 921     int i8x8, i4x4;
 922     int i_decimate_mb;
 923
 924     if( !b_bidir )
 925     {
 926         /* Get the MV */
 927         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 928         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 929
 930         /* Motion compensation */
 931         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 932                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 933                        mvp[0], mvp[1], 16, 16, &h->sh.weight[0][0] );
 934     }
 935
 936     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 937     {
 938         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
 939         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
 940         /* get luma diff */
 941         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
 942                                     h->mb.pic.p_fdec[0] + fdec_offset );
 943         /* encode one 4x4 block */
 944         for( i4x4 = 0; i4x4 < 4; i4x4++ )
 945         {
 946             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] ) )
 947                 continue;
 948             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 949             i_decimate_mb += h->quantf.decimate_score16( dctscan );
 950             if( i_decimate_mb >= 6 )
 951                 return 0;
 952         }
 953     }
 954
 955     /* encode chroma */
 956     i_qp = h->mb.i_chroma_qp;
 957     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 958
 959     for( ch = 0; ch < 2; ch++ )
 960     {
 961         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 962         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 963
 964         if( !b_bidir )
 965         {
 966             h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 967                              h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 968                              mvp[0], mvp[1], 8, 8 );
 969
 970             if( h->sh.weight[0][1+ch].weightfn )
 971                 h->sh.weight[0][1+ch].weightfn[8>>2]( h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
 972                                                       h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
 973                                                       &h->sh.weight[0][1+ch], 8 );
 974         }
 975
 976         /* there is almost never a termination during chroma, but we can't avoid the check entirely */
 977         /* so instead we check SSD and skip the actual check if the score is low enough. */
 978         ssd = h->pixf.ssd[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE );
 979         if( ssd < thresh )
 980             continue;
 981
 982         /* The vast majority of chroma checks will terminate during the DC check or the higher
 983          * threshold check, so we can save time by doing a DC-only DCT. */
 984         h->dctf.sub8x8_dct_dc( dct2x2, p_src, p_dst );
 985
 986         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 ) )
 987             return 0;
 988
 989         /* If there wasn't a termination in DC, we can check against a much higher threshold. */
 990         if( ssd < thresh*4 )
 991             continue;
 992
 993         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 994
 995         /* calculate dct coeffs */
 996         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 997         {
 998             /* We don't need to zero the DC coefficient before quantization because we already
 999              * checked that all the DCs were zero above at twice the precision that quant4x4
1000              * uses.  This applies even though the DC here is being quantized before the 2x2
1001              * transform. */
1002             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] ) )
1003                 continue;
1004             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
1005             i_decimate_mb += h->quantf.decimate_score15( dctscan );
1006             if( i_decimate_mb >= 7 )
1007                 return 0;
1008         }
1009     }
1010
1011     h->mb.b_skip_mc = 1;
1012     return 1;
1013 }
1014
1015 /****************************************************************************
1016  * DCT-domain noise reduction / adaptive deadzone
1017  * from libavcodec
1018  ****************************************************************************/
1019
1020 void x264_noise_reduction_update( x264_t *h )
1021 {
1022     int cat, i;
1023     for( cat = 0; cat < 2; cat++ )
1024     {
1025         int size = cat ? 64 : 16;
1026         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
1027
1028         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
1029         {
1030             for( i = 0; i < size; i++ )
1031                 h->nr_residual_sum[cat][i] >>= 1;
1032             h->nr_count[cat] >>= 1;
1033         }
1034
1035         for( i = 0; i < size; i++ )
1036             h->nr_offset[cat][i] =
1037                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
1038                  + h->nr_residual_sum[cat][i]/2)
1039               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
1040     }
1041 }
1042
1043 /*****************************************************************************
1044  * RD only; 4 calls to this do not make up for one macroblock_encode.
1045  * doesn't transform chroma dc.
1046  *****************************************************************************/
1047 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
1048 {
1049     int i_qp = h->mb.i_qp;
1050     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
1051     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
1052     int b_decimate = h->mb.b_dct_decimate;
1053     int nnz8x8 = 0;
1054     int ch, nz;
1055
1056     if( !h->mb.b_skip_mc )
1057         x264_mb_mc_8x8( h, i8 );
1058
1059     if( h->mb.b_lossless )
1060     {
1061         int i4;
1062         if( h->mb.b_transform_8x8 )
1063         {
1064             nnz8x8 = h->zigzagf.sub_8x8( h->dct.luma8x8[i8], p_fenc, p_fdec );
1065             STORE_8x8_NNZ(i8,nnz8x8);
1066         }
1067         else
1068         {
1069             for( i4 = i8*4; i4 < i8*4+4; i4++ )
1070             {
1071                 int nz;
1072                 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
1073                                     h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
1074                                     h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
1075                 h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1076                 nnz8x8 |= nz;
1077             }
1078         }
1079         for( ch = 0; ch < 2; ch++ )
1080         {
1081             int16_t dc;
1082             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1083             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1084             nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec, &dc );
1085             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1086         }
1087     }
1088     else
1089     {
1090         if( h->mb.b_transform_8x8 )
1091         {
1092             ALIGNED_ARRAY_16( int16_t, dct8x8,[64] );
1093             h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
1094             nnz8x8 = x264_quant_8x8( h, dct8x8, i_qp, 0, i8 );
1095             if( nnz8x8 )
1096             {
1097                 h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
1098
1099                 if( b_decimate && !h->mb.b_trellis )
1100                     nnz8x8 = 4 <= h->quantf.decimate_score64( h->dct.luma8x8[i8] );
1101
1102                 if( nnz8x8 )
1103                 {
1104                     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
1105                     h->dctf.add8x8_idct8( p_fdec, dct8x8 );
1106                     STORE_8x8_NNZ(i8,1);
1107                 }
1108                 else
1109                     STORE_8x8_NNZ(i8,0);
1110             }
1111             else
1112                 STORE_8x8_NNZ(i8,0);
1113         }
1114         else
1115         {
1116             int i4;
1117             int i_decimate_8x8 = 0;
1118             ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
1119             h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
1120             for( i4 = 0; i4 < 4; i4++ )
1121             {
1122                 nz = x264_quant_4x4( h, dct4x4[i4], i_qp, DCT_LUMA_4x4, 0, i8*4+i4 );
1123                 h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = nz;
1124                 if( nz )
1125                 {
1126                     h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
1127                     h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
1128                     if( b_decimate )
1129                         i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[i8*4+i4] );
1130                     nnz8x8 = 1;
1131                 }
1132             }
1133
1134             if( b_decimate && i_decimate_8x8 < 4 )
1135                 nnz8x8 = 0;
1136
1137             if( nnz8x8 )
1138                 h->dctf.add8x8_idct( p_fdec, dct4x4 );
1139             else
1140                 STORE_8x8_NNZ(i8,0);
1141         }
1142
1143         i_qp = h->mb.i_chroma_qp;
1144
1145         for( ch = 0; ch < 2; ch++ )
1146         {
1147             ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
1148             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1149             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1150
1151             h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1152             dct4x4[0] = 0;
1153
1154             if( h->mb.b_trellis )
1155                 nz = x264_quant_4x4_trellis( h, dct4x4, CQM_4PC, i_qp, DCT_CHROMA_AC, 0, 1, 0 );
1156             else
1157                 nz = h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
1158
1159             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1160             if( nz )
1161             {
1162                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
1163                 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
1164                 h->dctf.add4x4_idct( p_fdec, dct4x4 );
1165             }
1166         }
1167     }
1168     h->mb.i_cbp_luma &= ~(1 << i8);
1169     h->mb.i_cbp_luma |= nnz8x8 << i8;
1170     h->mb.i_cbp_chroma = 0x02;
1171 }
1172
1173 /*****************************************************************************
1174  * RD only, luma only
1175  *****************************************************************************/
1176 void x264_macroblock_encode_p4x4( x264_t *h, int i4 )
1177 {
1178     int i_qp = h->mb.i_qp;
1179     uint8_t *p_fenc = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[i4]];
1180     uint8_t *p_fdec = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
1181     int nz;
1182
1183     /* Don't need motion compensation as this function is only used in qpel-RD, which caches pixel data. */
1184
1185     if( h->mb.b_lossless )
1186     {
1187         nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4], p_fenc, p_fdec );
1188         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1189     }
1190     else
1191     {
1192         ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
1193         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1194         nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 0, i4 );
1195         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1196         if( nz )
1197         {
1198             h->zigzagf.scan_4x4( h->dct.luma4x4[i4], dct4x4 );
1199             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PY], i_qp );
1200             h->dctf.add4x4_idct( p_fdec, dct4x4 );
1201         }
1202     }
1203 }