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                     dct2x2dc_dconly( dct2x2 );
 369                     if( h->mb.b_trellis )
 370                         nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 371                     else
 372                         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]<<
 373     1 );
 374
 375                     if( nz_dc )
 376                     {
 377                         if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
 378                             continue;
 379                         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 1;
 380                         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 381                         idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 382                         h->dctf.add8x8_idct_dc( h->mb.pic.p_fdec[1+ch], dct2x2 );
 383                         h->mb.i_cbp_chroma = 1;
 384                     }
 385                 }
 386             }
 387             return;
 388         }
 389     }
 390
 391     for( ch = 0; ch < 2; ch++ )
 392     {
 393         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 394         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 395         int i_decimate_score = 0;
 396         int nz_ac = 0;
 397
 398         ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
 399
 400         if( h->mb.b_lossless )
 401         {
 402             for( i = 0; i < 4; i++ )
 403             {
 404                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 405                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 406                 nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od, &h->dct.chroma_dc[ch][i] );
 407                 h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 408                 h->mb.i_cbp_chroma |= nz;
 409             }
 410             h->mb.cache.non_zero_count[x264_scan8[25]+ch] = array_non_zero( h->dct.chroma_dc[ch] );
 411             continue;
 412         }
 413
 414         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 415         dct2x2dc( dct2x2, dct4x4 );
 416         /* calculate dct coeffs */
 417         for( i = 0; i < 4; i++ )
 418         {
 419             if( h->mb.b_trellis )
 420                 nz = x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 1, 0 );
 421             else
 422                 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] );
 423             h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 424             if( nz )
 425             {
 426                 nz_ac = 1;
 427                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 428                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 429                 if( b_decimate )
 430                     i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[16+i+ch*4] );
 431             }
 432         }
 433
 434         if( h->mb.b_trellis )
 435             nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 436         else
 437             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 );
 438
 439         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = nz_dc;
 440
 441         if( (b_decimate && i_decimate_score < 7) || !nz_ac )
 442         {
 443             /* Decimate the block */
 444             h->mb.cache.non_zero_count[x264_scan8[16+0]+24*ch] = 0;
 445             h->mb.cache.non_zero_count[x264_scan8[16+1]+24*ch] = 0;
 446             h->mb.cache.non_zero_count[x264_scan8[16+2]+24*ch] = 0;
 447             h->mb.cache.non_zero_count[x264_scan8[16+3]+24*ch] = 0;
 448             if( !nz_dc ) /* Whole block is empty */
 449                 continue;
 450             if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
 451             {
 452                 h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 0;
 453                 continue;
 454             }
 455             /* DC-only */
 456             zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 457             idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 458             h->dctf.add8x8_idct_dc( p_dst, dct2x2 );
 459         }
 460         else
 461         {
 462             h->mb.i_cbp_chroma = 1;
 463             if( nz_dc )
 464             {
 465                 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 466                 idct_dequant_2x2_dc( dct2x2, dct4x4, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 467             }
 468             h->dctf.add8x8_idct( p_dst, dct4x4 );
 469         }
 470     }
 471
 472     if( h->mb.i_cbp_chroma )
 473         h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
 474     else if( h->mb.cache.non_zero_count[x264_scan8[25]] |
 475              h->mb.cache.non_zero_count[x264_scan8[26]] )
 476         h->mb.i_cbp_chroma = 1;    /* dc only */
 477 }
 478
 479 static void x264_macroblock_encode_skip( x264_t *h )
 480 {
 481     h->mb.i_cbp_luma = 0x00;
 482     h->mb.i_cbp_chroma = 0x00;
 483     memset( h->mb.cache.non_zero_count, 0, X264_SCAN8_SIZE );
 484     /* store cbp */
 485     h->mb.cbp[h->mb.i_mb_xy] = 0;
 486 }
 487
 488 /*****************************************************************************
 489  * x264_macroblock_encode_pskip:
 490  *  Encode an already marked skip block
 491  *****************************************************************************/
 492 static void x264_macroblock_encode_pskip( x264_t *h )
 493 {
 494     const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 495                                 h->mb.mv_min[0], h->mb.mv_max[0] );
 496     const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 497                                 h->mb.mv_min[1], h->mb.mv_max[1] );
 498
 499     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 500     if( !h->mb.b_skip_mc )
 501     {
 502         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 503                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 504                        mvx, mvy, 16, 16, &h->sh.weight[0][0] );
 505
 506         h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 507                          h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 508                          mvx, mvy, 8, 8 );
 509
 510         if( h->sh.weight[0][1].weightfn )
 511             h->sh.weight[0][1].weightfn[8>>2]( h->mb.pic.p_fdec[1], FDEC_STRIDE,
 512                                                h->mb.pic.p_fdec[1], FDEC_STRIDE,
 513                                                &h->sh.weight[0][1], 8 );
 514
 515         h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 516                          h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 517                          mvx, mvy, 8, 8 );
 518
 519         if( h->sh.weight[0][2].weightfn )
 520             h->sh.weight[0][2].weightfn[8>>2]( h->mb.pic.p_fdec[2], FDEC_STRIDE,
 521                                                h->mb.pic.p_fdec[2], FDEC_STRIDE,
 522                                                &h->sh.weight[0][2], 8 );
 523     }
 524
 525     x264_macroblock_encode_skip( h );
 526 }
 527
 528 /*****************************************************************************
 529  * Intra prediction for predictive lossless mode.
 530  *****************************************************************************/
 531
 532 /* Note that these functions take a shortcut (mc.copy instead of actual pixel prediction) which assumes
 533  * that the edge pixels of the reconstructed frame are the same as that of the source frame.  This means
 534  * they will only work correctly if the neighboring blocks are losslessly coded.  In practice, this means
 535  * lossless mode cannot be mixed with lossy mode within a frame. */
 536 /* This can be resolved by explicitly copying the edge pixels after doing the mc.copy, but this doesn't
 537  * need to be done unless we decide to allow mixing lossless and lossy compression. */
 538
 539 void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
 540 {
 541     int stride = h->fenc->i_stride[1] << h->mb.b_interlaced;
 542     if( i_mode == I_PRED_CHROMA_V )
 543     {
 544         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-stride, stride, 8 );
 545         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-stride, stride, 8 );
 546     }
 547     else if( i_mode == I_PRED_CHROMA_H )
 548     {
 549         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-1, stride, 8 );
 550         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-1, stride, 8 );
 551     }
 552     else
 553     {
 554         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 555         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 556     }
 557 }
 558
 559 void x264_predict_lossless_4x4( x264_t *h, uint8_t *p_dst, int idx, int i_mode )
 560 {
 561     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 562     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
 563
 564     if( i_mode == I_PRED_4x4_V )
 565         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
 566     else if( i_mode == I_PRED_4x4_H )
 567         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-1, stride, 4 );
 568     else
 569         h->predict_4x4[i_mode]( p_dst );
 570 }
 571
 572 void x264_predict_lossless_8x8( x264_t *h, uint8_t *p_dst, int idx, int i_mode, uint8_t edge[33] )
 573 {
 574     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 575     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
 576
 577     if( i_mode == I_PRED_8x8_V )
 578         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
 579     else if( i_mode == I_PRED_8x8_H )
 580         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-1, stride, 8 );
 581     else
 582         h->predict_8x8[i_mode]( p_dst, edge );
 583 }
 584
 585 void x264_predict_lossless_16x16( x264_t *h, int i_mode )
 586 {
 587     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 588     if( i_mode == I_PRED_16x16_V )
 589         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-stride, stride, 16 );
 590     else if( i_mode == I_PRED_16x16_H )
 591         h->mc.copy_16x16_unaligned( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-1, stride, 16 );
 592     else
 593         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 594 }
 595
 596 /*****************************************************************************
 597  * x264_macroblock_encode:
 598  *****************************************************************************/
 599 void x264_macroblock_encode( x264_t *h )
 600 {
 601     int i_cbp_dc = 0;
 602     int i_qp = h->mb.i_qp;
 603     int b_decimate = h->mb.b_dct_decimate;
 604     int b_force_no_skip = 0;
 605     int i,idx,nz;
 606     h->mb.i_cbp_luma = 0;
 607     h->mb.cache.non_zero_count[x264_scan8[24]] = 0;
 608
 609     if( h->sh.b_mbaff
 610         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 611         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 612     {
 613         /* The first skip is predicted to be a frame mb pair.
 614          * We don't yet support the aff part of mbaff, so force it to non-skip
 615          * so that we can pick the aff flag. */
 616         b_force_no_skip = 1;
 617         if( IS_SKIP(h->mb.i_type) )
 618         {
 619             if( h->mb.i_type == P_SKIP )
 620                 h->mb.i_type = P_L0;
 621             else if( h->mb.i_type == B_SKIP )
 622                 h->mb.i_type = B_DIRECT;
 623         }
 624     }
 625
 626     if( h->mb.i_type == P_SKIP )
 627     {
 628         /* A bit special */
 629         x264_macroblock_encode_pskip( h );
 630         return;
 631     }
 632     if( h->mb.i_type == B_SKIP )
 633     {
 634         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 635         if( !h->mb.b_skip_mc )
 636             x264_mb_mc( h );
 637         x264_macroblock_encode_skip( h );
 638         return;
 639     }
 640
 641     if( h->mb.i_type == I_16x16 )
 642     {
 643         const int i_mode = h->mb.i_intra16x16_pred_mode;
 644         h->mb.b_transform_8x8 = 0;
 645
 646         if( h->mb.b_lossless )
 647             x264_predict_lossless_16x16( h, i_mode );
 648         else
 649             h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 650
 651         /* encode the 16x16 macroblock */
 652         x264_mb_encode_i16x16( h, i_qp );
 653     }
 654     else if( h->mb.i_type == I_8x8 )
 655     {
 656         ALIGNED_ARRAY_16( uint8_t, edge,[33] );
 657         h->mb.b_transform_8x8 = 1;
 658         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 659         if( h->mb.i_skip_intra )
 660         {
 661             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 662             M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i8x8_nnz_buf[0];
 663             M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i8x8_nnz_buf[1];
 664             M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i8x8_nnz_buf[2];
 665             M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i8x8_nnz_buf[3];
 666             h->mb.i_cbp_luma = h->mb.pic.i8x8_cbp;
 667             /* In RD mode, restore the now-overwritten DCT data. */
 668             if( h->mb.i_skip_intra == 2 )
 669                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 670         }
 671         for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 672         {
 673             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 674             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 675             h->predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 676
 677             if( h->mb.b_lossless )
 678                 x264_predict_lossless_8x8( h, p_dst, i, i_mode, edge );
 679             else
 680                 h->predict_8x8[i_mode]( p_dst, edge );
 681
 682             x264_mb_encode_i8x8( h, i, i_qp );
 683         }
 684     }
 685     else if( h->mb.i_type == I_4x4 )
 686     {
 687         h->mb.b_transform_8x8 = 0;
 688         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 689         if( h->mb.i_skip_intra )
 690         {
 691             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 692             M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i4x4_nnz_buf[0];
 693             M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i4x4_nnz_buf[1];
 694             M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i4x4_nnz_buf[2];
 695             M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i4x4_nnz_buf[3];
 696             h->mb.i_cbp_luma = h->mb.pic.i4x4_cbp;
 697             /* In RD mode, restore the now-overwritten DCT data. */
 698             if( h->mb.i_skip_intra == 2 )
 699                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 700         }
 701         for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 702         {
 703             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
 704             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 705
 706             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 707                 /* emulate missing topright samples */
 708                 M32( &p_dst[4-FDEC_STRIDE] ) = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 709
 710             if( h->mb.b_lossless )
 711                 x264_predict_lossless_4x4( h, p_dst, i, i_mode );
 712             else
 713                 h->predict_4x4[i_mode]( p_dst );
 714             x264_mb_encode_i4x4( h, i, i_qp );
 715         }
 716     }
 717     else    /* Inter MB */
 718     {
 719         int i8x8, i4x4;
 720         int i_decimate_mb = 0;
 721
 722         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
 723         if( !h->mb.b_skip_mc )
 724             x264_mb_mc( h );
 725
 726         if( h->mb.b_lossless )
 727         {
 728             if( h->mb.b_transform_8x8 )
 729                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 730                 {
 731                     int x = 8*(i8x8&1);
 732                     int y = 8*(i8x8>>1);
 733                     nz = h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8],
 734                                         h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
 735                                         h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
 736                     STORE_8x8_NNZ(i8x8,nz);
 737                     h->mb.i_cbp_luma |= nz << i8x8;
 738                 }
 739             else
 740                 for( i4x4 = 0; i4x4 < 16; i4x4++ )
 741                 {
 742                     nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 743                                         h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
 744                                         h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4x4] );
 745                     h->mb.cache.non_zero_count[x264_scan8[i4x4]] = nz;
 746                     h->mb.i_cbp_luma |= nz << (i4x4>>2);
 747                 }
 748         }
 749         else if( h->mb.b_transform_8x8 )
 750         {
 751             ALIGNED_ARRAY_16( int16_t, dct8x8,[4],[64] );
 752             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 753             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 754             h->nr_count[1] += h->mb.b_noise_reduction * 4;
 755
 756             for( idx = 0; idx < 4; idx++ )
 757             {
 758                 if( h->mb.b_noise_reduction )
 759                     h->quantf.denoise_dct( dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
 760                 nz = x264_quant_8x8( h, dct8x8[idx], i_qp, 0, idx );
 761
 762                 if( nz )
 763                 {
 764                     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 765                     if( b_decimate )
 766                     {
 767                         int i_decimate_8x8 = h->quantf.decimate_score64( h->dct.luma8x8[idx] );
 768                         i_decimate_mb += i_decimate_8x8;
 769                         if( i_decimate_8x8 >= 4 )
 770                             h->mb.i_cbp_luma |= 1<<idx;
 771                     }
 772                     else
 773                         h->mb.i_cbp_luma |= 1<<idx;
 774                 }
 775             }
 776
 777             if( i_decimate_mb < 6 && b_decimate )
 778             {
 779                 h->mb.i_cbp_luma = 0;
 780                 CLEAR_16x16_NNZ
 781             }
 782             else
 783             {
 784                 for( idx = 0; idx < 4; idx++ )
 785                 {
 786                     if( h->mb.i_cbp_luma&(1<<idx) )
 787                     {
 788                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 789                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 790                         STORE_8x8_NNZ(idx,1);
 791                     }
 792                     else
 793                         STORE_8x8_NNZ(idx,0);
 794                 }
 795             }
 796         }
 797         else
 798         {
 799             ALIGNED_ARRAY_16( int16_t, dct4x4,[16],[16] );
 800             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 801             h->nr_count[0] += h->mb.b_noise_reduction * 16;
 802
 803             for( i8x8 = 0; i8x8 < 4; i8x8++ )
 804             {
 805                 int i_decimate_8x8 = 0;
 806                 int cbp = 0;
 807
 808                 /* encode one 4x4 block */
 809                 for( i4x4 = 0; i4x4 < 4; i4x4++ )
 810                 {
 811                     idx = i8x8 * 4 + i4x4;
 812
 813                     if( h->mb.b_noise_reduction )
 814                         h->quantf.denoise_dct( dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
 815                     nz = x264_quant_4x4( h, dct4x4[idx], i_qp, DCT_LUMA_4x4, 0, idx );
 816                     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 817
 818                     if( nz )
 819                     {
 820                         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 821                         h->quantf.dequant_4x4( dct4x4[idx], h->dequant4_mf[CQM_4PY], i_qp );
 822                         if( b_decimate && i_decimate_8x8 < 6 )
 823                             i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[idx] );
 824                         cbp = 1;
 825                     }
 826                 }
 827
 828                 /* decimate this 8x8 block */
 829                 i_decimate_mb += i_decimate_8x8;
 830                 if( b_decimate )
 831                 {
 832                     if( i_decimate_8x8 < 4 )
 833                         STORE_8x8_NNZ(i8x8,0)
 834                     else
 835                         h->mb.i_cbp_luma |= 1<<i8x8;
 836                 }
 837                 else if( cbp )
 838                 {
 839                     h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 840                     h->mb.i_cbp_luma |= 1<<i8x8;
 841                 }
 842             }
 843
 844             if( b_decimate )
 845             {
 846                 if( i_decimate_mb < 6 )
 847                 {
 848                     h->mb.i_cbp_luma = 0;
 849                     CLEAR_16x16_NNZ
 850                 }
 851                 else
 852                 {
 853                     for( i8x8 = 0; i8x8 < 4; i8x8++ )
 854                         if( h->mb.i_cbp_luma&(1<<i8x8) )
 855                             h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 856                 }
 857             }
 858         }
 859     }
 860
 861     /* encode chroma */
 862     if( IS_INTRA( h->mb.i_type ) )
 863     {
 864         const int i_mode = h->mb.i_chroma_pred_mode;
 865         if( h->mb.b_lossless )
 866             x264_predict_lossless_8x8_chroma( h, i_mode );
 867         else
 868         {
 869             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 870             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 871         }
 872     }
 873
 874     /* encode the 8x8 blocks */
 875     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 876
 877     if( h->param.b_cabac )
 878     {
 879         i_cbp_dc = h->mb.cache.non_zero_count[x264_scan8[24]]
 880                  | h->mb.cache.non_zero_count[x264_scan8[25]] << 1
 881                  | h->mb.cache.non_zero_count[x264_scan8[26]] << 2;
 882     }
 883
 884     /* store cbp */
 885     h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
 886
 887     /* Check for P_SKIP
 888      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 889      *      (if multiple mv give same result)*/
 890     if( !b_force_no_skip )
 891     {
 892         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 893             !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
 894             M32( h->mb.cache.mv[0][x264_scan8[0]] ) == M32( h->mb.cache.pskip_mv )
 895             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 896         {
 897             h->mb.i_type = P_SKIP;
 898         }
 899
 900         /* Check for B_SKIP */
 901         if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
 902         {
 903             h->mb.i_type = B_SKIP;
 904         }
 905     }
 906 }
 907
 908 /*****************************************************************************
 909  * x264_macroblock_probe_skip:
 910  *  Check if the current MB could be encoded as a [PB]_SKIP
 911  *****************************************************************************/
 912 int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
 913 {
 914     ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
 915     ALIGNED_ARRAY_16( int16_t, dct2x2,[4] );
 916     ALIGNED_ARRAY_16( int16_t, dctscan,[16] );
 917
 918     int i_qp = h->mb.i_qp;
 919     int mvp[2];
 920     int ch, thresh, ssd;
 921
 922     int i8x8, i4x4;
 923     int i_decimate_mb;
 924
 925     if( !b_bidir )
 926     {
 927         /* Get the MV */
 928         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 929         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 930
 931         /* Motion compensation */
 932         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 933                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 934                        mvp[0], mvp[1], 16, 16, &h->sh.weight[0][0] );
 935     }
 936
 937     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 938     {
 939         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
 940         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
 941         /* get luma diff */
 942         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
 943                                     h->mb.pic.p_fdec[0] + fdec_offset );
 944         /* encode one 4x4 block */
 945         for( i4x4 = 0; i4x4 < 4; i4x4++ )
 946         {
 947             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] ) )
 948                 continue;
 949             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 950             i_decimate_mb += h->quantf.decimate_score16( dctscan );
 951             if( i_decimate_mb >= 6 )
 952                 return 0;
 953         }
 954     }
 955
 956     /* encode chroma */
 957     i_qp = h->mb.i_chroma_qp;
 958     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 959
 960     for( ch = 0; ch < 2; ch++ )
 961     {
 962         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 963         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 964
 965         if( !b_bidir )
 966         {
 967             h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 968                              h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 969                              mvp[0], mvp[1], 8, 8 );
 970
 971             if( h->sh.weight[0][1+ch].weightfn )
 972                 h->sh.weight[0][1+ch].weightfn[8>>2]( h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
 973                                                       h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
 974                                                       &h->sh.weight[0][1+ch], 8 );
 975         }
 976
 977         /* there is almost never a termination during chroma, but we can't avoid the check entirely */
 978         /* so instead we check SSD and skip the actual check if the score is low enough. */
 979         ssd = h->pixf.ssd[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE );
 980         if( ssd < thresh )
 981             continue;
 982
 983         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 984
 985         /* calculate dct DC */
 986         dct2x2dc( dct2x2, dct4x4 );
 987         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 ) )
 988             return 0;
 989
 990         /* If there wasn't a termination in DC, we can check against a much higher threshold. */
 991         if( ssd < thresh*4 )
 992             continue;
 993
 994         /* calculate dct coeffs */
 995         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 996         {
 997             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] ) )
 998                 continue;
 999             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
1000             i_decimate_mb += h->quantf.decimate_score15( dctscan );
1001             if( i_decimate_mb >= 7 )
1002                 return 0;
1003         }
1004     }
1005
1006     h->mb.b_skip_mc = 1;
1007     return 1;
1008 }
1009
1010 /****************************************************************************
1011  * DCT-domain noise reduction / adaptive deadzone
1012  * from libavcodec
1013  ****************************************************************************/
1014
1015 void x264_noise_reduction_update( x264_t *h )
1016 {
1017     int cat, i;
1018     for( cat = 0; cat < 2; cat++ )
1019     {
1020         int size = cat ? 64 : 16;
1021         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
1022
1023         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
1024         {
1025             for( i = 0; i < size; i++ )
1026                 h->nr_residual_sum[cat][i] >>= 1;
1027             h->nr_count[cat] >>= 1;
1028         }
1029
1030         for( i = 0; i < size; i++ )
1031             h->nr_offset[cat][i] =
1032                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
1033                  + h->nr_residual_sum[cat][i]/2)
1034               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
1035     }
1036 }
1037
1038 /*****************************************************************************
1039  * RD only; 4 calls to this do not make up for one macroblock_encode.
1040  * doesn't transform chroma dc.
1041  *****************************************************************************/
1042 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
1043 {
1044     int i_qp = h->mb.i_qp;
1045     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
1046     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
1047     int b_decimate = h->mb.b_dct_decimate;
1048     int nnz8x8 = 0;
1049     int ch, nz;
1050
1051     if( !h->mb.b_skip_mc )
1052         x264_mb_mc_8x8( h, i8 );
1053
1054     if( h->mb.b_lossless )
1055     {
1056         int i4;
1057         if( h->mb.b_transform_8x8 )
1058         {
1059             nnz8x8 = h->zigzagf.sub_8x8( h->dct.luma8x8[i8], p_fenc, p_fdec );
1060             STORE_8x8_NNZ(i8,nnz8x8);
1061         }
1062         else
1063         {
1064             for( i4 = i8*4; i4 < i8*4+4; i4++ )
1065             {
1066                 int nz;
1067                 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
1068                                     h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
1069                                     h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
1070                 h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1071                 nnz8x8 |= nz;
1072             }
1073         }
1074         for( ch = 0; ch < 2; ch++ )
1075         {
1076             int16_t dc;
1077             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1078             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1079             nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec, &dc );
1080             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1081         }
1082     }
1083     else
1084     {
1085         if( h->mb.b_transform_8x8 )
1086         {
1087             ALIGNED_ARRAY_16( int16_t, dct8x8,[64] );
1088             h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
1089             nnz8x8 = x264_quant_8x8( h, dct8x8, i_qp, 0, i8 );
1090             if( nnz8x8 )
1091             {
1092                 h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
1093
1094                 if( b_decimate && !h->mb.b_trellis )
1095                     nnz8x8 = 4 <= h->quantf.decimate_score64( h->dct.luma8x8[i8] );
1096
1097                 if( nnz8x8 )
1098                 {
1099                     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
1100                     h->dctf.add8x8_idct8( p_fdec, dct8x8 );
1101                     STORE_8x8_NNZ(i8,1);
1102                 }
1103                 else
1104                     STORE_8x8_NNZ(i8,0);
1105             }
1106             else
1107                 STORE_8x8_NNZ(i8,0);
1108         }
1109         else
1110         {
1111             int i4;
1112             int i_decimate_8x8 = 0;
1113             ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
1114             h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
1115             for( i4 = 0; i4 < 4; i4++ )
1116             {
1117                 nz = x264_quant_4x4( h, dct4x4[i4], i_qp, DCT_LUMA_4x4, 0, i8*4+i4 );
1118                 h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = nz;
1119                 if( nz )
1120                 {
1121                     h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
1122                     h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
1123                     if( b_decimate )
1124                         i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[i8*4+i4] );
1125                     nnz8x8 = 1;
1126                 }
1127             }
1128
1129             if( b_decimate && i_decimate_8x8 < 4 )
1130                 nnz8x8 = 0;
1131
1132             if( nnz8x8 )
1133                 h->dctf.add8x8_idct( p_fdec, dct4x4 );
1134             else
1135                 STORE_8x8_NNZ(i8,0);
1136         }
1137
1138         i_qp = h->mb.i_chroma_qp;
1139
1140         for( ch = 0; ch < 2; ch++ )
1141         {
1142             ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
1143             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1144             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1145
1146             h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1147             dct4x4[0] = 0;
1148
1149             if( h->mb.b_trellis )
1150                 nz = x264_quant_4x4_trellis( h, dct4x4, CQM_4PC, i_qp, DCT_CHROMA_AC, 0, 1, 0 );
1151             else
1152                 nz = h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
1153
1154             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1155             if( nz )
1156             {
1157                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
1158                 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
1159                 h->dctf.add4x4_idct( p_fdec, dct4x4 );
1160             }
1161         }
1162     }
1163     h->mb.i_cbp_luma &= ~(1 << i8);
1164     h->mb.i_cbp_luma |= nnz8x8 << i8;
1165     h->mb.i_cbp_chroma = 0x02;
1166 }
1167
1168 /*****************************************************************************
1169  * RD only, luma only
1170  *****************************************************************************/
1171 void x264_macroblock_encode_p4x4( x264_t *h, int i4 )
1172 {
1173     int i_qp = h->mb.i_qp;
1174     uint8_t *p_fenc = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[i4]];
1175     uint8_t *p_fdec = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
1176     int nz;
1177
1178     /* Don't need motion compensation as this function is only used in qpel-RD, which caches pixel data. */
1179
1180     if( h->mb.b_lossless )
1181     {
1182         nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4], p_fenc, p_fdec );
1183         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1184     }
1185     else
1186     {
1187         ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
1188         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1189         nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 0, i4 );
1190         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1191         if( nz )
1192         {
1193             h->zigzagf.scan_4x4( h->dct.luma4x4[i4], dct4x4 );
1194             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PY], i_qp );
1195             h->dctf.add4x4_idct( p_fdec, dct4x4 );
1196         }
1197     }
1198 }