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( dctcoef level[4], dctcoef 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( dctcoef dct[4], dctcoef 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( dctcoef out[4], dctcoef 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( dctcoef d[4], dctcoef 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 ALWAYS_INLINE int x264_quant_4x4( x264_t *h, dctcoef dct[16], int i_qp, int i_ctxBlockCat, int b_intra, int idx )
  82 {
  83     int i_quant_cat = b_intra ? CQM_4IY : CQM_4PY;
  84     if( h->mb.b_trellis )
  85         return x264_quant_4x4_trellis( h, dct, i_quant_cat, i_qp, i_ctxBlockCat, b_intra, 0, idx );
  86     else
  87         return h->quantf.quant_4x4( dct, h->quant4_mf[i_quant_cat][i_qp], h->quant4_bias[i_quant_cat][i_qp] );
  88 }
  89
  90 static ALWAYS_INLINE int x264_quant_8x8( x264_t *h, dctcoef dct[64], int i_qp, int b_intra, int idx )
  91 {
  92     int i_quant_cat = b_intra ? CQM_8IY : CQM_8PY;
  93     if( h->mb.b_trellis )
  94         return x264_quant_8x8_trellis( h, dct, i_quant_cat, i_qp, b_intra, idx );
  95     else
  96         return h->quantf.quant_8x8( dct, h->quant8_mf[i_quant_cat][i_qp], h->quant8_bias[i_quant_cat][i_qp] );
  97 }
  98
  99 /* All encoding functions must output the correct CBP and NNZ values.
 100  * The entropy coding functions will check CBP first, then NNZ, before
 101  * actually reading the DCT coefficients.  NNZ still must be correct even
 102  * if CBP is zero because of the use of NNZ values for context selection.
 103  * "NNZ" need only be 0 or 1 rather than the exact coefficient count because
 104  * that is only needed in CAVLC, and will be calculated by CAVLC's residual
 105  * coding and stored as necessary. */
 106
 107 /* This means that decimation can be done merely by adjusting the CBP and NNZ
 108  * rather than memsetting the coefficients. */
 109
 110 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qp )
 111 {
 112     int nz;
 113     pixel *p_src = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[idx]];
 114     pixel *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[idx]];
 115     ALIGNED_ARRAY_16( dctcoef, dct4x4,[16] );
 116
 117     if( h->mb.b_lossless )
 118     {
 119         nz = h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
 120         h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 121         h->mb.i_cbp_luma |= nz<<(idx>>2);
 122         return;
 123     }
 124
 125     h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
 126
 127     nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 1, idx );
 128     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 129     if( nz )
 130     {
 131         h->mb.i_cbp_luma |= 1<<(idx>>2);
 132         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4 );
 133         h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qp );
 134         h->dctf.add4x4_idct( p_dst, dct4x4 );
 135     }
 136 }
 137
 138 #define STORE_8x8_NNZ( s8, nz )\
 139 do\
 140 {\
 141     M16( &h->mb.cache.non_zero_count[(s8) + 0*8] ) = (nz) * 0x0101;\
 142     M16( &h->mb.cache.non_zero_count[(s8) + 1*8] ) = (nz) * 0x0101;\
 143 } while(0)
 144
 145 #define CLEAR_16x16_NNZ \
 146 {\
 147     M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = 0;\
 148     M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = 0;\
 149     M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = 0;\
 150     M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = 0;\
 151 }
 152
 153 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qp )
 154 {
 155     int x = idx&1;
 156     int y = idx>>1;
 157     int s8 = X264_SCAN8_0 + 2*x + 16*y;
 158     int nz;
 159     pixel *p_src = &h->mb.pic.p_fenc[0][8*x + 8*y*FENC_STRIDE];
 160     pixel *p_dst = &h->mb.pic.p_fdec[0][8*x + 8*y*FDEC_STRIDE];
 161     ALIGNED_ARRAY_16( dctcoef, dct8x8,[64] );
 162
 163     if( h->mb.b_lossless )
 164     {
 165         nz = h->zigzagf.sub_8x8( h->dct.luma8x8[idx], p_src, p_dst );
 166         STORE_8x8_NNZ( s8, nz );
 167         h->mb.i_cbp_luma |= nz<<idx;
 168         return;
 169     }
 170
 171     h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
 172
 173     nz = x264_quant_8x8( h, dct8x8, i_qp, 1, idx );
 174     if( nz )
 175     {
 176         h->mb.i_cbp_luma |= 1<<idx;
 177         h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
 178         h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qp );
 179         h->dctf.add8x8_idct8( p_dst, dct8x8 );
 180         STORE_8x8_NNZ( s8, 1 );
 181     }
 182     else
 183         STORE_8x8_NNZ( s8, 0 );
 184 }
 185
 186 static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
 187 {
 188     pixel *p_src = h->mb.pic.p_fenc[0];
 189     pixel *p_dst = h->mb.pic.p_fdec[0];
 190
 191     ALIGNED_ARRAY_16( dctcoef, dct4x4,[16],[16] );
 192     ALIGNED_ARRAY_16( dctcoef, dct_dc4x4,[16] );
 193
 194     int nz;
 195     int decimate_score = h->mb.b_dct_decimate ? 0 : 9;
 196
 197     if( h->mb.b_lossless )
 198     {
 199         for( int i = 0; i < 16; i++ )
 200         {
 201             int oe = block_idx_xy_fenc[i];
 202             int od = block_idx_xy_fdec[i];
 203             nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[i], p_src+oe, p_dst+od, &dct_dc4x4[block_idx_yx_1d[i]] );
 204             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 205             h->mb.i_cbp_luma |= nz;
 206         }
 207         h->mb.i_cbp_luma *= 0xf;
 208         h->mb.cache.non_zero_count[x264_scan8[24]] = array_non_zero( dct_dc4x4 );
 209         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 210         return;
 211     }
 212
 213     h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
 214
 215     for( int i = 0; i < 16; i++ )
 216     {
 217         /* copy dc coeff */
 218         dct_dc4x4[block_idx_xy_1d[i]] = dct4x4[i][0];
 219         dct4x4[i][0] = 0;
 220
 221         /* quant/scan/dequant */
 222         nz = x264_quant_4x4( h, dct4x4[i], i_qp, DCT_LUMA_AC, 1, i );
 223         h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 224         if( nz )
 225         {
 226             h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[i] );
 227             h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IY], i_qp );
 228             if( decimate_score < 6 ) decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[i] );
 229             h->mb.i_cbp_luma = 0xf;
 230         }
 231     }
 232
 233     /* Writing the 16 CBFs in an i16x16 block is quite costly, so decimation can save many bits. */
 234     /* More useful with CAVLC, but still useful with CABAC. */
 235     if( decimate_score < 6 )
 236     {
 237         h->mb.i_cbp_luma = 0;
 238         CLEAR_16x16_NNZ
 239     }
 240
 241     h->dctf.dct4x4dc( dct_dc4x4 );
 242     if( h->mb.b_trellis )
 243         nz = x264_quant_dc_trellis( h, dct_dc4x4, CQM_4IY, i_qp, DCT_LUMA_DC, 1, 0 );
 244     else
 245         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 );
 246
 247     h->mb.cache.non_zero_count[x264_scan8[24]] = nz;
 248     if( nz )
 249     {
 250         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 251
 252         /* output samples to fdec */
 253         h->dctf.idct4x4dc( dct_dc4x4 );
 254         h->quantf.dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[CQM_4IY], i_qp );  /* XXX not inversed */
 255         if( h->mb.i_cbp_luma )
 256             for( int i = 0; i < 16; i++ )
 257                 dct4x4[i][0] = dct_dc4x4[block_idx_xy_1d[i]];
 258     }
 259
 260     /* put pixels to fdec */
 261     if( h->mb.i_cbp_luma )
 262         h->dctf.add16x16_idct( p_dst, dct4x4 );
 263     else if( nz )
 264         h->dctf.add16x16_idct_dc( p_dst, dct_dc4x4 );
 265 }
 266
 267 static inline int idct_dequant_round_2x2_dc( dctcoef ref[4], dctcoef dct[4], int dequant_mf[6][16], int i_qp )
 268 {
 269     dctcoef out[4];
 270     idct_dequant_2x2_dconly( out, dct, dequant_mf, i_qp );
 271     return ((ref[0] ^ (out[0]+32))
 272           | (ref[1] ^ (out[1]+32))
 273           | (ref[2] ^ (out[2]+32))
 274           | (ref[3] ^ (out[3]+32))) >> 6;
 275 }
 276
 277 /* Round down coefficients losslessly in DC-only chroma blocks.
 278  * Unlike luma blocks, this can't be done with a lookup table or
 279  * other shortcut technique because of the interdependencies
 280  * between the coefficients due to the chroma DC transform. */
 281 static inline int x264_mb_optimize_chroma_dc( x264_t *h, int b_inter, int i_qp, dctcoef dct2x2[4] )
 282 {
 283     dctcoef dct2x2_orig[4];
 284     int coeff, nz;
 285
 286     /* If the QP is too high, there's no benefit to rounding optimization. */
 287     if( h->dequant4_mf[CQM_4IC + b_inter][i_qp%6][0] << (i_qp/6) > 32*64 )
 288         return 1;
 289
 290     idct_dequant_2x2_dconly( dct2x2_orig, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 291     dct2x2_orig[0] += 32;
 292     dct2x2_orig[1] += 32;
 293     dct2x2_orig[2] += 32;
 294     dct2x2_orig[3] += 32;
 295
 296     /* If the DC coefficients already round to zero, terminate early. */
 297     if( !((dct2x2_orig[0]|dct2x2_orig[1]|dct2x2_orig[2]|dct2x2_orig[3])>>6) )
 298         return 0;
 299
 300     /* Start with the highest frequency coefficient... is this the best option? */
 301     for( nz = 0, coeff = h->quantf.coeff_last[DCT_CHROMA_DC]( dct2x2 ); coeff >= 0; coeff-- )
 302     {
 303         int level = dct2x2[coeff];
 304         int sign = level>>31 | 1; /* dct2x2[coeff] < 0 ? -1 : 1 */
 305
 306         while( level )
 307         {
 308             dct2x2[coeff] = level - sign;
 309             if( idct_dequant_round_2x2_dc( dct2x2_orig, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp ) )
 310             {
 311                 nz = 1;
 312                 dct2x2[coeff] = level;
 313                 break;
 314             }
 315             level -= sign;
 316         }
 317     }
 318
 319     return nz;
 320 }
 321
 322 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
 323 {
 324     int nz, nz_dc;
 325     int b_decimate = b_inter && h->mb.b_dct_decimate;
 326     ALIGNED_ARRAY_16( dctcoef, dct2x2,[4] );
 327     h->mb.i_cbp_chroma = 0;
 328
 329     /* Early termination: check variance of chroma residual before encoding.
 330      * Don't bother trying early termination at low QPs.
 331      * Values are experimentally derived. */
 332     if( b_decimate && i_qp >= (h->mb.b_trellis ? 12 : 18) )
 333     {
 334         int thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 335         int ssd[2];
 336         int score = h->pixf.var2_8x8( h->mb.pic.p_fenc[1], FENC_STRIDE, h->mb.pic.p_fdec[1], FDEC_STRIDE, &ssd[0] );
 337         if( score < thresh*4 )
 338             score += h->pixf.var2_8x8( h->mb.pic.p_fenc[2], FENC_STRIDE, h->mb.pic.p_fdec[2], FDEC_STRIDE, &ssd[1] );
 339         if( score < thresh*4 )
 340         {
 341             h->mb.cache.non_zero_count[x264_scan8[16]] = 0;
 342             h->mb.cache.non_zero_count[x264_scan8[17]] = 0;
 343             h->mb.cache.non_zero_count[x264_scan8[18]] = 0;
 344             h->mb.cache.non_zero_count[x264_scan8[19]] = 0;
 345             h->mb.cache.non_zero_count[x264_scan8[20]] = 0;
 346             h->mb.cache.non_zero_count[x264_scan8[21]] = 0;
 347             h->mb.cache.non_zero_count[x264_scan8[22]] = 0;
 348             h->mb.cache.non_zero_count[x264_scan8[23]] = 0;
 349             M16( &h->mb.cache.non_zero_count[x264_scan8[25]] ) = 0;
 350
 351             for( int ch = 0; ch < 2; ch++ )
 352             {
 353                 if( ssd[ch] > thresh )
 354                 {
 355                     h->dctf.sub8x8_dct_dc( dct2x2, h->mb.pic.p_fenc[1+ch], h->mb.pic.p_fdec[1+ch] );
 356                     if( h->mb.b_trellis )
 357                         nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 358                     else
 359                         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 );
 360
 361                     if( nz_dc )
 362                     {
 363                         if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
 364                             continue;
 365                         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 1;
 366                         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 367                         idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 368                         h->dctf.add8x8_idct_dc( h->mb.pic.p_fdec[1+ch], dct2x2 );
 369                         h->mb.i_cbp_chroma = 1;
 370                     }
 371                 }
 372             }
 373             return;
 374         }
 375     }
 376
 377     for( int ch = 0; ch < 2; ch++ )
 378     {
 379         pixel *p_src = h->mb.pic.p_fenc[1+ch];
 380         pixel *p_dst = h->mb.pic.p_fdec[1+ch];
 381         int i_decimate_score = 0;
 382         int nz_ac = 0;
 383
 384         ALIGNED_ARRAY_16( dctcoef, dct4x4,[4],[16] );
 385
 386         if( h->mb.b_lossless )
 387         {
 388             for( int i = 0; i < 4; i++ )
 389             {
 390                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 391                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 392                 nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od, &h->dct.chroma_dc[ch][i] );
 393                 h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 394                 h->mb.i_cbp_chroma |= nz;
 395             }
 396             h->mb.cache.non_zero_count[x264_scan8[25]+ch] = array_non_zero( h->dct.chroma_dc[ch] );
 397             continue;
 398         }
 399
 400         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 401         dct2x2dc( dct2x2, dct4x4 );
 402         /* calculate dct coeffs */
 403         for( int i = 0; i < 4; i++ )
 404         {
 405             if( h->mb.b_trellis )
 406                 nz = x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 1, 0 );
 407             else
 408                 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] );
 409             h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 410             if( nz )
 411             {
 412                 nz_ac = 1;
 413                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 414                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 415                 if( b_decimate )
 416                     i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[16+i+ch*4] );
 417             }
 418         }
 419
 420         if( h->mb.b_trellis )
 421             nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 422         else
 423             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 );
 424
 425         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = nz_dc;
 426
 427         if( (b_decimate && i_decimate_score < 7) || !nz_ac )
 428         {
 429             /* Decimate the block */
 430             h->mb.cache.non_zero_count[x264_scan8[16+0]+24*ch] = 0;
 431             h->mb.cache.non_zero_count[x264_scan8[16+1]+24*ch] = 0;
 432             h->mb.cache.non_zero_count[x264_scan8[16+2]+24*ch] = 0;
 433             h->mb.cache.non_zero_count[x264_scan8[16+3]+24*ch] = 0;
 434             if( !nz_dc ) /* Whole block is empty */
 435                 continue;
 436             if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
 437             {
 438                 h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 0;
 439                 continue;
 440             }
 441             /* DC-only */
 442             zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 443             idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 444             h->dctf.add8x8_idct_dc( p_dst, dct2x2 );
 445         }
 446         else
 447         {
 448             h->mb.i_cbp_chroma = 1;
 449             if( nz_dc )
 450             {
 451                 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 452                 idct_dequant_2x2_dc( dct2x2, dct4x4, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 453             }
 454             h->dctf.add8x8_idct( p_dst, dct4x4 );
 455         }
 456     }
 457
 458     /* 0 = none, 1 = DC only, 2 = DC+AC */
 459     h->mb.i_cbp_chroma = ((!!M16( &h->mb.cache.non_zero_count[x264_scan8[25]] )) | h->mb.i_cbp_chroma) + h->mb.i_cbp_chroma;
 460 }
 461
 462 static void x264_macroblock_encode_skip( x264_t *h )
 463 {
 464     M32( &h->mb.cache.non_zero_count[x264_scan8[0]+0*8] ) = 0;
 465     M32( &h->mb.cache.non_zero_count[x264_scan8[0]+1*8] ) = 0;
 466     M32( &h->mb.cache.non_zero_count[x264_scan8[0]+2*8] ) = 0;
 467     M32( &h->mb.cache.non_zero_count[x264_scan8[0]+3*8] ) = 0;
 468     for( int i = 16; i < 24; i++ )
 469         h->mb.cache.non_zero_count[x264_scan8[i]] = 0;
 470     h->mb.i_cbp_luma = 0;
 471     h->mb.i_cbp_chroma = 0;
 472     h->mb.cbp[h->mb.i_mb_xy] = 0;
 473 }
 474
 475 /*****************************************************************************
 476  * x264_macroblock_encode_pskip:
 477  *  Encode an already marked skip block
 478  *****************************************************************************/
 479 static void x264_macroblock_encode_pskip( x264_t *h )
 480 {
 481     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 482     if( !h->mb.b_skip_mc )
 483     {
 484         int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 485                               h->mb.mv_min[0], h->mb.mv_max[0] );
 486         int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 487                               h->mb.mv_min[1], h->mb.mv_max[1] );
 488
 489         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 490                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 491                        mvx, mvy, 16, 16, &h->sh.weight[0][0] );
 492
 493         /* Special case for mv0, which is (of course) very common in P-skip mode. */
 494         if( mvx | mvy )
 495             h->mc.mc_chroma( h->mb.pic.p_fdec[1], h->mb.pic.p_fdec[2], FDEC_STRIDE,
 496                              h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 497                              mvx, mvy, 8, 8 );
 498         else
 499             h->mc.load_deinterleave_8x8x2_fdec( h->mb.pic.p_fdec[1], h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1] );
 500
 501         if( h->sh.weight[0][1].weightfn )
 502             h->sh.weight[0][1].weightfn[8>>2]( h->mb.pic.p_fdec[1], FDEC_STRIDE,
 503                                                h->mb.pic.p_fdec[1], FDEC_STRIDE,
 504                                                &h->sh.weight[0][1], 8 );
 505         if( h->sh.weight[0][2].weightfn )
 506             h->sh.weight[0][2].weightfn[8>>2]( h->mb.pic.p_fdec[2], FDEC_STRIDE,
 507                                                h->mb.pic.p_fdec[2], FDEC_STRIDE,
 508                                                &h->sh.weight[0][2], 8 );
 509     }
 510
 511     x264_macroblock_encode_skip( h );
 512 }
 513
 514 /*****************************************************************************
 515  * Intra prediction for predictive lossless mode.
 516  *****************************************************************************/
 517
 518 void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
 519 {
 520     if( i_mode == I_PRED_CHROMA_V )
 521     {
 522         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc[1]-FENC_STRIDE, FENC_STRIDE, 8 );
 523         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc[2]-FENC_STRIDE, FENC_STRIDE, 8 );
 524         memcpy( h->mb.pic.p_fdec[1], h->mb.pic.p_fdec[1]-FDEC_STRIDE, 8*sizeof(pixel) );
 525         memcpy( h->mb.pic.p_fdec[2], h->mb.pic.p_fdec[2]-FDEC_STRIDE, 8*sizeof(pixel) );
 526     }
 527     else if( i_mode == I_PRED_CHROMA_H )
 528     {
 529         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc[1]-1, FENC_STRIDE, 8 );
 530         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc[2]-1, FENC_STRIDE, 8 );
 531         x264_copy_column8( h->mb.pic.p_fdec[1]+4*FDEC_STRIDE, h->mb.pic.p_fdec[1]+4*FDEC_STRIDE-1 );
 532         x264_copy_column8( h->mb.pic.p_fdec[2]+4*FDEC_STRIDE, h->mb.pic.p_fdec[2]+4*FDEC_STRIDE-1 );
 533     }
 534     else
 535     {
 536         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 537         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 538     }
 539 }
 540
 541 void x264_predict_lossless_4x4( x264_t *h, pixel *p_dst, int idx, int i_mode )
 542 {
 543     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 544     pixel *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
 545
 546     if( i_mode == I_PRED_4x4_V )
 547         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
 548     else if( i_mode == I_PRED_4x4_H )
 549         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-1, stride, 4 );
 550     else
 551         h->predict_4x4[i_mode]( p_dst );
 552 }
 553
 554 void x264_predict_lossless_8x8( x264_t *h, pixel *p_dst, int idx, int i_mode, pixel edge[33] )
 555 {
 556     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 557     pixel *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
 558
 559     if( i_mode == I_PRED_8x8_V )
 560         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
 561     else if( i_mode == I_PRED_8x8_H )
 562         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-1, stride, 8 );
 563     else
 564         h->predict_8x8[i_mode]( p_dst, edge );
 565 }
 566
 567 void x264_predict_lossless_16x16( x264_t *h, int i_mode )
 568 {
 569     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 570     if( i_mode == I_PRED_16x16_V )
 571         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-stride, stride, 16 );
 572     else if( i_mode == I_PRED_16x16_H )
 573         h->mc.copy_16x16_unaligned( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-1, stride, 16 );
 574     else
 575         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 576 }
 577
 578 /*****************************************************************************
 579  * x264_macroblock_encode:
 580  *****************************************************************************/
 581 void x264_macroblock_encode( x264_t *h )
 582 {
 583     int i_qp = h->mb.i_qp;
 584     int b_decimate = h->mb.b_dct_decimate;
 585     int b_force_no_skip = 0;
 586     int nz;
 587     h->mb.i_cbp_luma = 0;
 588     h->mb.cache.non_zero_count[x264_scan8[24]] = 0;
 589
 590     if( h->mb.i_type == I_PCM )
 591     {
 592         /* if PCM is chosen, we need to store reconstructed frame data */
 593         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE, 16 );
 594         h->mc.copy[PIXEL_8x8]  ( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc[1], FENC_STRIDE, 8 );
 595         h->mc.copy[PIXEL_8x8]  ( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc[2], FENC_STRIDE, 8 );
 596         return;
 597     }
 598
 599     if( h->sh.b_mbaff
 600         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 601         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 602     {
 603         /* The first skip is predicted to be a frame mb pair.
 604          * We don't yet support the aff part of mbaff, so force it to non-skip
 605          * so that we can pick the aff flag. */
 606         b_force_no_skip = 1;
 607         if( IS_SKIP(h->mb.i_type) )
 608         {
 609             if( h->mb.i_type == P_SKIP )
 610                 h->mb.i_type = P_L0;
 611             else if( h->mb.i_type == B_SKIP )
 612                 h->mb.i_type = B_DIRECT;
 613         }
 614     }
 615
 616     if( h->mb.i_type == P_SKIP )
 617     {
 618         /* A bit special */
 619         x264_macroblock_encode_pskip( h );
 620         return;
 621     }
 622     if( h->mb.i_type == B_SKIP )
 623     {
 624         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 625         if( !h->mb.b_skip_mc )
 626             x264_mb_mc( h );
 627         x264_macroblock_encode_skip( h );
 628         return;
 629     }
 630
 631     if( h->mb.i_type == I_16x16 )
 632     {
 633         const int i_mode = h->mb.i_intra16x16_pred_mode;
 634         h->mb.b_transform_8x8 = 0;
 635
 636         if( h->mb.b_lossless )
 637             x264_predict_lossless_16x16( h, i_mode );
 638         else
 639             h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 640
 641         /* encode the 16x16 macroblock */
 642         x264_mb_encode_i16x16( h, i_qp );
 643     }
 644     else if( h->mb.i_type == I_8x8 )
 645     {
 646         ALIGNED_ARRAY_16( pixel, edge,[33] );
 647         h->mb.b_transform_8x8 = 1;
 648         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 649         if( h->mb.i_skip_intra )
 650         {
 651             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 652             M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i8x8_nnz_buf[0];
 653             M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i8x8_nnz_buf[1];
 654             M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i8x8_nnz_buf[2];
 655             M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i8x8_nnz_buf[3];
 656             h->mb.i_cbp_luma = h->mb.pic.i8x8_cbp;
 657             /* In RD mode, restore the now-overwritten DCT data. */
 658             if( h->mb.i_skip_intra == 2 )
 659                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 660         }
 661         for( int i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 662         {
 663             pixel *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 664             int i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 665             h->predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 666
 667             if( h->mb.b_lossless )
 668                 x264_predict_lossless_8x8( h, p_dst, i, i_mode, edge );
 669             else
 670                 h->predict_8x8[i_mode]( p_dst, edge );
 671
 672             x264_mb_encode_i8x8( h, i, i_qp );
 673         }
 674     }
 675     else if( h->mb.i_type == I_4x4 )
 676     {
 677         h->mb.b_transform_8x8 = 0;
 678         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 679         if( h->mb.i_skip_intra )
 680         {
 681             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 682             M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i4x4_nnz_buf[0];
 683             M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i4x4_nnz_buf[1];
 684             M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i4x4_nnz_buf[2];
 685             M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i4x4_nnz_buf[3];
 686             h->mb.i_cbp_luma = h->mb.pic.i4x4_cbp;
 687             /* In RD mode, restore the now-overwritten DCT data. */
 688             if( h->mb.i_skip_intra == 2 )
 689                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 690         }
 691         for( int i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 692         {
 693             pixel *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
 694             int i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 695
 696             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 697                 /* emulate missing topright samples */
 698                 MPIXEL_X4( &p_dst[4-FDEC_STRIDE] ) = PIXEL_SPLAT_X4( p_dst[3-FDEC_STRIDE] );
 699
 700             if( h->mb.b_lossless )
 701                 x264_predict_lossless_4x4( h, p_dst, i, i_mode );
 702             else
 703                 h->predict_4x4[i_mode]( p_dst );
 704             x264_mb_encode_i4x4( h, i, i_qp );
 705         }
 706     }
 707     else    /* Inter MB */
 708     {
 709         int i_decimate_mb = 0;
 710
 711         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
 712         if( !h->mb.b_skip_mc )
 713             x264_mb_mc( h );
 714
 715         if( h->mb.b_lossless )
 716         {
 717             if( h->mb.b_transform_8x8 )
 718                 for( int i8x8 = 0; i8x8 < 4; i8x8++ )
 719                 {
 720                     int x = i8x8&1;
 721                     int y = i8x8>>1;
 722                     int s8 = X264_SCAN8_0 + 2*x + 16*y;
 723
 724                     nz = h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8], h->mb.pic.p_fenc[0] + 8*x + 8*y*FENC_STRIDE,
 725                                                                    h->mb.pic.p_fdec[0] + 8*x + 8*y*FDEC_STRIDE );
 726                     STORE_8x8_NNZ( s8, nz );
 727                     h->mb.i_cbp_luma |= nz << i8x8;
 728                 }
 729             else
 730                 for( int i4x4 = 0; i4x4 < 16; i4x4++ )
 731                 {
 732                     nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 733                                         h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
 734                                         h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4x4] );
 735                     h->mb.cache.non_zero_count[x264_scan8[i4x4]] = nz;
 736                     h->mb.i_cbp_luma |= nz << (i4x4>>2);
 737                 }
 738         }
 739         else if( h->mb.b_transform_8x8 )
 740         {
 741             ALIGNED_ARRAY_16( dctcoef, dct8x8,[4],[64] );
 742             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 743             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 744             h->nr_count[1] += h->mb.b_noise_reduction * 4;
 745
 746             for( int idx = 0; idx < 4; idx++ )
 747             {
 748                 if( h->mb.b_noise_reduction )
 749                     h->quantf.denoise_dct( dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
 750                 nz = x264_quant_8x8( h, dct8x8[idx], i_qp, 0, idx );
 751
 752                 if( nz )
 753                 {
 754                     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 755                     if( b_decimate )
 756                     {
 757                         int i_decimate_8x8 = h->quantf.decimate_score64( h->dct.luma8x8[idx] );
 758                         i_decimate_mb += i_decimate_8x8;
 759                         if( i_decimate_8x8 >= 4 )
 760                             h->mb.i_cbp_luma |= 1<<idx;
 761                     }
 762                     else
 763                         h->mb.i_cbp_luma |= 1<<idx;
 764                 }
 765             }
 766
 767             if( i_decimate_mb < 6 && b_decimate )
 768             {
 769                 h->mb.i_cbp_luma = 0;
 770                 CLEAR_16x16_NNZ
 771             }
 772             else
 773             {
 774                 for( int idx = 0; idx < 4; idx++ )
 775                 {
 776                     int x = idx&1;
 777                     int y = idx>>1;
 778                     int s8 = X264_SCAN8_0 + 2*x + 16*y;
 779
 780                     if( h->mb.i_cbp_luma&(1<<idx) )
 781                     {
 782                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 783                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][8*x + 8*y*FDEC_STRIDE], dct8x8[idx] );
 784                         STORE_8x8_NNZ( s8, 1 );
 785                     }
 786                     else
 787                         STORE_8x8_NNZ( s8, 0 );
 788                 }
 789             }
 790         }
 791         else
 792         {
 793             ALIGNED_ARRAY_16( dctcoef, dct4x4,[16],[16] );
 794             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 795             h->nr_count[0] += h->mb.b_noise_reduction * 16;
 796
 797             for( int i8x8 = 0; i8x8 < 4; i8x8++ )
 798             {
 799                 int i_decimate_8x8 = 0;
 800                 int cbp = 0;
 801
 802                 /* encode one 4x4 block */
 803                 for( int i4x4 = 0; i4x4 < 4; i4x4++ )
 804                 {
 805                     int idx = i8x8 * 4 + i4x4;
 806
 807                     if( h->mb.b_noise_reduction )
 808                         h->quantf.denoise_dct( dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
 809                     nz = x264_quant_4x4( h, dct4x4[idx], i_qp, DCT_LUMA_4x4, 0, idx );
 810                     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 811
 812                     if( nz )
 813                     {
 814                         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 815                         h->quantf.dequant_4x4( dct4x4[idx], h->dequant4_mf[CQM_4PY], i_qp );
 816                         if( b_decimate && i_decimate_8x8 < 6 )
 817                             i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[idx] );
 818                         cbp = 1;
 819                     }
 820                 }
 821
 822                 int x = i8x8&1;
 823                 int y = i8x8>>1;
 824
 825                 /* decimate this 8x8 block */
 826                 i_decimate_mb += i_decimate_8x8;
 827                 if( b_decimate )
 828                 {
 829                     if( i_decimate_8x8 < 4 )
 830                     {
 831                         int s8 = X264_SCAN8_0 + 2*x + 16*y;
 832                         STORE_8x8_NNZ( s8, 0 );
 833                     }
 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][8*x + 8*y*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( int 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     /* store cbp */
 878     int cbp = h->mb.i_cbp_chroma << 4 | h->mb.i_cbp_luma;
 879     if( h->param.b_cabac )
 880         cbp |= h->mb.cache.non_zero_count[x264_scan8[24]] << 8
 881             |  h->mb.cache.non_zero_count[x264_scan8[25]] << 9
 882             |  h->mb.cache.non_zero_count[x264_scan8[26]] << 10;
 883     h->mb.cbp[h->mb.i_mb_xy] = cbp;
 884
 885     /* Check for P_SKIP
 886      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 887      *      (if multiple mv give same result)*/
 888     if( !b_force_no_skip )
 889     {
 890         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 891             !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
 892             M32( h->mb.cache.mv[0][x264_scan8[0]] ) == M32( h->mb.cache.pskip_mv )
 893             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 894         {
 895             h->mb.i_type = P_SKIP;
 896         }
 897
 898         /* Check for B_SKIP */
 899         if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
 900         {
 901             h->mb.i_type = B_SKIP;
 902         }
 903     }
 904 }
 905
 906 /*****************************************************************************
 907  * x264_macroblock_probe_skip:
 908  *  Check if the current MB could be encoded as a [PB]_SKIP
 909  *****************************************************************************/
 910 int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
 911 {
 912     ALIGNED_ARRAY_16( dctcoef, dct4x4,[4],[16] );
 913     ALIGNED_ARRAY_16( dctcoef, dct2x2,[4] );
 914     ALIGNED_ARRAY_16( dctcoef, dctscan,[16] );
 915     ALIGNED_4( int16_t mvp[2] );
 916
 917     int i_qp = h->mb.i_qp;
 918     int thresh, ssd;
 919
 920     if( !b_bidir )
 921     {
 922         /* Get the MV */
 923         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 924         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 925
 926         /* Motion compensation */
 927         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 928                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 929                        mvp[0], mvp[1], 16, 16, &h->sh.weight[0][0] );
 930     }
 931
 932     for( int i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 933     {
 934         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
 935         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
 936         /* get luma diff */
 937         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
 938                                     h->mb.pic.p_fdec[0] + fdec_offset );
 939         /* encode one 4x4 block */
 940         for( int i4x4 = 0; i4x4 < 4; i4x4++ )
 941         {
 942             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] ) )
 943                 continue;
 944             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 945             i_decimate_mb += h->quantf.decimate_score16( dctscan );
 946             if( i_decimate_mb >= 6 )
 947                 return 0;
 948         }
 949     }
 950
 951     /* encode chroma */
 952     i_qp = h->mb.i_chroma_qp;
 953     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 954
 955     if( !b_bidir )
 956     {
 957         /* Special case for mv0, which is (of course) very common in P-skip mode. */
 958         if( M32( mvp ) )
 959             h->mc.mc_chroma( h->mb.pic.p_fdec[1], h->mb.pic.p_fdec[2], FDEC_STRIDE,
 960                              h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 961                              mvp[0], mvp[1], 8, 8 );
 962         else
 963             h->mc.load_deinterleave_8x8x2_fdec( h->mb.pic.p_fdec[1], h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1] );
 964     }
 965
 966     for( int ch = 0; ch < 2; ch++ )
 967     {
 968         pixel *p_src = h->mb.pic.p_fenc[1+ch];
 969         pixel *p_dst = h->mb.pic.p_fdec[1+ch];
 970
 971         if( !b_bidir && 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         /* 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( int i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 997         {
 998             dct4x4[i4x4][0] = 0;
 999             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] ) )
1000                 continue;
1001             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
1002             i_decimate_mb += h->quantf.decimate_score15( dctscan );
1003             if( i_decimate_mb >= 7 )
1004                 return 0;
1005         }
1006     }
1007
1008     h->mb.b_skip_mc = 1;
1009     return 1;
1010 }
1011
1012 /****************************************************************************
1013  * DCT-domain noise reduction / adaptive deadzone
1014  * from libavcodec
1015  ****************************************************************************/
1016
1017 void x264_noise_reduction_update( x264_t *h )
1018 {
1019     for( int cat = 0; cat < 2; cat++ )
1020     {
1021         int size = cat ? 64 : 16;
1022         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
1023
1024         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
1025         {
1026             for( int i = 0; i < size; i++ )
1027                 h->nr_residual_sum[cat][i] >>= 1;
1028             h->nr_count[cat] >>= 1;
1029         }
1030
1031         for( int i = 0; i < size; i++ )
1032             h->nr_offset[cat][i] =
1033                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
1034                  + h->nr_residual_sum[cat][i]/2)
1035               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
1036     }
1037 }
1038
1039 /*****************************************************************************
1040  * RD only; 4 calls to this do not make up for one macroblock_encode.
1041  * doesn't transform chroma dc.
1042  *****************************************************************************/
1043 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
1044 {
1045     int i_qp = h->mb.i_qp;
1046     int x = i8&1;
1047     int y = i8>>1;
1048     int s8 = X264_SCAN8_0 + 2*x + 16*y;
1049     pixel *p_fenc = h->mb.pic.p_fenc[0] + 8*x + 8*y*FENC_STRIDE;
1050     pixel *p_fdec = h->mb.pic.p_fdec[0] + 8*x + 8*y*FDEC_STRIDE;
1051     int b_decimate = h->mb.b_dct_decimate;
1052     int nnz8x8 = 0;
1053     int nz;
1054
1055     if( !h->mb.b_skip_mc )
1056         x264_mb_mc_8x8( h, i8 );
1057
1058     if( h->mb.b_lossless )
1059     {
1060         if( h->mb.b_transform_8x8 )
1061         {
1062             nnz8x8 = h->zigzagf.sub_8x8( h->dct.luma8x8[i8], p_fenc, p_fdec );
1063             STORE_8x8_NNZ( s8, nnz8x8 );
1064         }
1065         else
1066         {
1067             for( int i4 = i8*4; i4 < i8*4+4; i4++ )
1068             {
1069                 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
1070                                     h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
1071                                     h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
1072                 h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1073                 nnz8x8 |= nz;
1074             }
1075         }
1076         for( int ch = 0; ch < 2; ch++ )
1077         {
1078             dctcoef dc;
1079             p_fenc = h->mb.pic.p_fenc[1+ch] + 4*x + 4*y*FENC_STRIDE;
1080             p_fdec = h->mb.pic.p_fdec[1+ch] + 4*x + 4*y*FDEC_STRIDE;
1081             nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec, &dc );
1082             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1083         }
1084     }
1085     else
1086     {
1087         if( h->mb.b_transform_8x8 )
1088         {
1089             ALIGNED_ARRAY_16( dctcoef, dct8x8,[64] );
1090             h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
1091             nnz8x8 = x264_quant_8x8( h, dct8x8, i_qp, 0, i8 );
1092             if( nnz8x8 )
1093             {
1094                 h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
1095
1096                 if( b_decimate && !h->mb.b_trellis )
1097                     nnz8x8 = 4 <= h->quantf.decimate_score64( h->dct.luma8x8[i8] );
1098
1099                 if( nnz8x8 )
1100                 {
1101                     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
1102                     h->dctf.add8x8_idct8( p_fdec, dct8x8 );
1103                     STORE_8x8_NNZ( s8, 1 );
1104                 }
1105                 else
1106                     STORE_8x8_NNZ( s8, 0 );
1107             }
1108             else
1109                 STORE_8x8_NNZ( s8, 0 );
1110         }
1111         else
1112         {
1113             int i_decimate_8x8 = 0;
1114             ALIGNED_ARRAY_16( dctcoef, dct4x4,[4],[16] );
1115             h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
1116             for( int i4 = 0; i4 < 4; i4++ )
1117             {
1118                 nz = x264_quant_4x4( h, dct4x4[i4], i_qp, DCT_LUMA_4x4, 0, i8*4+i4 );
1119                 h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = nz;
1120                 if( nz )
1121                 {
1122                     h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
1123                     h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
1124                     if( b_decimate )
1125                         i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[i8*4+i4] );
1126                     nnz8x8 = 1;
1127                 }
1128             }
1129
1130             if( b_decimate && i_decimate_8x8 < 4 )
1131                 nnz8x8 = 0;
1132
1133             if( nnz8x8 )
1134                 h->dctf.add8x8_idct( p_fdec, dct4x4 );
1135             else
1136                 STORE_8x8_NNZ( s8, 0 );
1137         }
1138
1139         i_qp = h->mb.i_chroma_qp;
1140
1141         for( int ch = 0; ch < 2; ch++ )
1142         {
1143             ALIGNED_ARRAY_16( dctcoef, dct4x4,[16] );
1144             p_fenc = h->mb.pic.p_fenc[1+ch] + 4*x + 4*y*FENC_STRIDE;
1145             p_fdec = h->mb.pic.p_fdec[1+ch] + 4*x + 4*y*FDEC_STRIDE;
1146
1147             h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1148             dct4x4[0] = 0;
1149
1150             if( h->mb.b_trellis )
1151                 nz = x264_quant_4x4_trellis( h, dct4x4, CQM_4PC, i_qp, DCT_CHROMA_AC, 0, 1, 0 );
1152             else
1153                 nz = h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
1154
1155             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1156             if( nz )
1157             {
1158                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
1159                 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
1160                 h->dctf.add4x4_idct( p_fdec, dct4x4 );
1161             }
1162         }
1163     }
1164     h->mb.i_cbp_luma &= ~(1 << i8);
1165     h->mb.i_cbp_luma |= nnz8x8 << i8;
1166     h->mb.i_cbp_chroma = 0x02;
1167 }
1168
1169 /*****************************************************************************
1170  * RD only, luma only
1171  *****************************************************************************/
1172 void x264_macroblock_encode_p4x4( x264_t *h, int i4 )
1173 {
1174     int i_qp = h->mb.i_qp;
1175     pixel *p_fenc = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[i4]];
1176     pixel *p_fdec = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
1177     int nz;
1178
1179     /* Don't need motion compensation as this function is only used in qpel-RD, which caches pixel data. */
1180
1181     if( h->mb.b_lossless )
1182     {
1183         nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4], p_fenc, p_fdec );
1184         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1185     }
1186     else
1187     {
1188         ALIGNED_ARRAY_16( dctcoef, dct4x4,[16] );
1189         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1190         nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 0, i4 );
1191         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1192         if( nz )
1193         {
1194             h->zigzagf.scan_4x4( h->dct.luma4x4[i4], dct4x4 );
1195             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PY], i_qp );
1196             h->dctf.add4x4_idct( p_fdec, dct4x4 );
1197         }
1198     }
1199 }