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

   1 /*****************************************************************************
   2  * macroblock.c: h264 encoder library
   3  *****************************************************************************
   4  * Copyright (C) 2003-2008 x264 project
   5  *
   6  * Authors: Laurent Aimar <fenrir@via.ecp.fr>
   7  *          Loren Merritt <lorenm@u.washington.edu>
   8  *          Fiona Glaser <fiona@x264.com>
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published by
  12  * the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  * GNU General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * along with this program; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
  23  *****************************************************************************/
  24
  25 #include "common/common.h"
  26 #include "macroblock.h"
  27
  28 /* These chroma DC functions don't have assembly versions and are only used here. */
  29
  30 #define ZIG(i,y,x) level[i] = dct[x][y];
  31 static inline void zigzag_scan_2x2_dc( int16_t level[4], int16_t dct[2][2] )
  32 {
  33     ZIG(0,0,0)
  34     ZIG(1,0,1)
  35     ZIG(2,1,0)
  36     ZIG(3,1,1)
  37 }
  38 #undef ZIG
  39
  40 #define IDCT_DEQUANT_START \
  41     int d0 = dct[0][0] + dct[0][1]; \
  42     int d1 = dct[1][0] + dct[1][1]; \
  43     int d2 = dct[0][0] - dct[0][1]; \
  44     int d3 = dct[1][0] - dct[1][1]; \
  45     int dmf = dequant_mf[i_qp%6][0][0]; \
  46     int qbits = i_qp/6 - 5; \
  47     if( qbits > 0 ) \
  48     { \
  49         dmf <<= qbits; \
  50         qbits = 0; \
  51     }
  52
  53 static inline void idct_dequant_2x2_dc( int16_t dct[2][2], int16_t dct4x4[4][4][4], int dequant_mf[6][4][4], int i_qp )
  54 {
  55     IDCT_DEQUANT_START
  56     dct4x4[0][0][0] = (d0 + d1) * dmf >> -qbits;
  57     dct4x4[1][0][0] = (d0 - d1) * dmf >> -qbits;
  58     dct4x4[2][0][0] = (d2 + d3) * dmf >> -qbits;
  59     dct4x4[3][0][0] = (d2 - d3) * dmf >> -qbits;
  60 }
  61
  62 static inline void idct_dequant_2x2_dconly( int16_t dct[2][2], int dequant_mf[6][4][4], int i_qp )
  63 {
  64     IDCT_DEQUANT_START
  65     dct[0][0] = (d0 + d1) * dmf >> -qbits;
  66     dct[0][1] = (d0 - d1) * dmf >> -qbits;
  67     dct[1][0] = (d2 + d3) * dmf >> -qbits;
  68     dct[1][1] = (d2 - d3) * dmf >> -qbits;
  69 }
  70
  71 static inline void dct2x2dc( int16_t d[2][2], int16_t dct4x4[4][4][4] )
  72 {
  73     int d0 = dct4x4[0][0][0] + dct4x4[1][0][0];
  74     int d1 = dct4x4[2][0][0] + dct4x4[3][0][0];
  75     int d2 = dct4x4[0][0][0] - dct4x4[1][0][0];
  76     int d3 = dct4x4[2][0][0] - dct4x4[3][0][0];
  77     d[0][0] = d0 + d1;
  78     d[1][0] = d2 + d3;
  79     d[0][1] = d0 - d1;
  80     d[1][1] = d2 - d3;
  81     dct4x4[0][0][0] = 0;
  82     dct4x4[1][0][0] = 0;
  83     dct4x4[2][0][0] = 0;
  84     dct4x4[3][0][0] = 0;
  85 }
  86
  87 static ALWAYS_INLINE int x264_quant_4x4( x264_t *h, int16_t dct[4][4], int i_qp, int i_ctxBlockCat, int b_intra, int idx )
  88 {
  89     int i_quant_cat = b_intra ? CQM_4IY : CQM_4PY;
  90     if( h->mb.b_trellis )
  91         return x264_quant_4x4_trellis( h, dct, i_quant_cat, i_qp, i_ctxBlockCat, b_intra, idx );
  92     else
  93         return h->quantf.quant_4x4( dct, h->quant4_mf[i_quant_cat][i_qp], h->quant4_bias[i_quant_cat][i_qp] );
  94 }
  95
  96 static ALWAYS_INLINE int x264_quant_8x8( x264_t *h, int16_t dct[8][8], int i_qp, int b_intra, int idx )
  97 {
  98     int i_quant_cat = b_intra ? CQM_8IY : CQM_8PY;
  99     if( h->mb.b_trellis )
 100         return x264_quant_8x8_trellis( h, dct, i_quant_cat, i_qp, b_intra, idx );
 101     else
 102         return h->quantf.quant_8x8( dct, h->quant8_mf[i_quant_cat][i_qp], h->quant8_bias[i_quant_cat][i_qp] );
 103 }
 104
 105 /* All encoding functions must output the correct CBP and NNZ values.
 106  * The entropy coding functions will check CBP first, then NNZ, before
 107  * actually reading the DCT coefficients.  NNZ still must be correct even
 108  * if CBP is zero because of the use of NNZ values for context selection.
 109  * "NNZ" need only be 0 or 1 rather than the exact coefficient count because
 110  * that is only needed in CAVLC, and will be calculated by CAVLC's residual
 111  * coding and stored as necessary. */
 112
 113 /* This means that decimation can be done merely by adjusting the CBP and NNZ
 114  * rather than memsetting the coefficients. */
 115
 116 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qp )
 117 {
 118     int nz;
 119     uint8_t *p_src = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[idx]];
 120     uint8_t *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[idx]];
 121     DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
 122
 123     if( h->mb.b_lossless )
 124     {
 125         h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
 126         nz = array_non_zero( h->dct.luma4x4[idx] );
 127         h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 128         h->mb.i_cbp_luma |= nz<<(idx>>2);
 129         return;
 130     }
 131
 132     h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
 133
 134     nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 1, idx );
 135     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 136     if( nz )
 137     {
 138         h->mb.i_cbp_luma |= 1<<(idx>>2);
 139         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4 );
 140         h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qp );
 141         h->dctf.add4x4_idct( p_dst, dct4x4 );
 142     }
 143 }
 144
 145 #define STORE_8x8_NNZ(idx,nz)\
 146 {\
 147     *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[idx*4+0]] = nz * 0x0101;\
 148     *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[idx*4+2]] = nz * 0x0101;\
 149 }
 150
 151 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qp )
 152 {
 153     int x = 8 * (idx&1);
 154     int y = 8 * (idx>>1);
 155     int nz;
 156     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
 157     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
 158     DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 159
 160     if( h->mb.b_lossless )
 161     {
 162         h->zigzagf.sub_8x8( h->dct.luma8x8[idx], p_src, p_dst );
 163         nz = array_non_zero( h->dct.luma8x8[idx] );
 164         STORE_8x8_NNZ(idx,nz);
 165         h->mb.i_cbp_luma |= nz<<idx;
 166         return;
 167     }
 168
 169     h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
 170
 171     nz = x264_quant_8x8( h, dct8x8, i_qp, 1, idx );
 172     if( nz )
 173     {
 174         h->mb.i_cbp_luma |= 1<<idx;
 175         h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
 176         h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qp );
 177         h->dctf.add8x8_idct8( p_dst, dct8x8 );
 178         STORE_8x8_NNZ(idx,1);
 179     }
 180     else
 181         STORE_8x8_NNZ(idx,0);
 182 }
 183
 184 static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
 185 {
 186     uint8_t  *p_src = h->mb.pic.p_fenc[0];
 187     uint8_t  *p_dst = h->mb.pic.p_fdec[0];
 188
 189     DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 190     DECLARE_ALIGNED_16( int16_t dct_dc4x4[4][4] );
 191
 192     int i, nz;
 193     int b_decimate = h->sh.i_type == SLICE_TYPE_B || (h->param.analyse.b_dct_decimate && h->sh.i_type == SLICE_TYPE_P);
 194     int decimate_score = b_decimate ? 0 : 9;
 195
 196     if( h->mb.b_lossless )
 197     {
 198         for( i = 0; i < 16; i++ )
 199         {
 200             int oe = block_idx_xy_fenc[i];
 201             int od = block_idx_xy_fdec[i];
 202             h->zigzagf.sub_4x4( h->dct.luma4x4[i], p_src+oe, p_dst+od );
 203             dct_dc4x4[0][block_idx_yx_1d[i]] = h->dct.luma4x4[i][0];
 204             h->dct.luma4x4[i][0] = 0;
 205             nz = array_non_zero( h->dct.luma4x4[i] );
 206             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 207             h->mb.i_cbp_luma |= nz;
 208         }
 209         h->mb.i_cbp_luma *= 0xf;
 210         h->mb.cache.non_zero_count[x264_scan8[24]] = array_non_zero( dct_dc4x4 );
 211         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 212         return;
 213     }
 214
 215     h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
 216
 217     for( i = 0; i < 16; i++ )
 218     {
 219         /* copy dc coeff */
 220         dct_dc4x4[0][block_idx_xy_1d[i]] = dct4x4[i][0][0];
 221         dct4x4[i][0][0] = 0;
 222
 223         /* quant/scan/dequant */
 224         nz = x264_quant_4x4( h, dct4x4[i], i_qp, DCT_LUMA_AC, 1, i );
 225         h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 226         if( nz )
 227         {
 228             h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[i] );
 229             h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IY], i_qp );
 230             if( decimate_score < 6 ) decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[i] );
 231             h->mb.i_cbp_luma = 0xf;
 232         }
 233     }
 234
 235     /* Writing the 16 CBFs in an i16x16 block is quite costly, so decimation can save many bits. */
 236     /* More useful with CAVLC, but still useful with CABAC. */
 237     if( decimate_score < 6 )
 238     {
 239         h->mb.i_cbp_luma = 0;
 240         *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = 0;
 241         *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = 0;
 242         *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = 0;
 243         *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = 0;
 244     }
 245
 246     h->dctf.dct4x4dc( dct_dc4x4 );
 247     if( h->mb.b_trellis )
 248         nz = x264_quant_dc_trellis( h, (int16_t*)dct_dc4x4, CQM_4IY, i_qp, DCT_LUMA_DC, 1);
 249     else
 250         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 );
 251
 252     h->mb.cache.non_zero_count[x264_scan8[24]] = nz;
 253     if( nz )
 254     {
 255         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 256
 257         /* output samples to fdec */
 258         h->dctf.idct4x4dc( dct_dc4x4 );
 259         h->quantf.dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[CQM_4IY], i_qp );  /* XXX not inversed */
 260         if( h->mb.i_cbp_luma )
 261             for( i = 0; i < 16; i++ )
 262                 dct4x4[i][0][0] = dct_dc4x4[0][block_idx_xy_1d[i]];
 263     }
 264
 265     /* put pixels to fdec */
 266     if( h->mb.i_cbp_luma )
 267         h->dctf.add16x16_idct( p_dst, dct4x4 );
 268     else if( nz )
 269         h->dctf.add16x16_idct_dc( p_dst, dct_dc4x4 );
 270 }
 271
 272 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
 273 {
 274     int i, ch, nz, nz_dc;
 275     int b_decimate = b_inter && (h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate);
 276     h->mb.i_cbp_chroma = 0;
 277
 278     for( ch = 0; ch < 2; ch++ )
 279     {
 280         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 281         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 282         int i_decimate_score = 0;
 283         int nz_ac = 0;
 284
 285         DECLARE_ALIGNED_16( int16_t dct2x2[2][2]  );
 286         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 287
 288         if( h->mb.b_lossless )
 289         {
 290             for( i = 0; i < 4; i++ )
 291             {
 292                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 293                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 294                 h->zigzagf.sub_4x4( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od );
 295                 h->dct.chroma_dc[ch][i] = h->dct.luma4x4[16+i+ch*4][0];
 296                 h->dct.luma4x4[16+i+ch*4][0] = 0;
 297                 nz = array_non_zero( h->dct.luma4x4[16+i+ch*4] );
 298                 h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 299                 h->mb.i_cbp_chroma |= nz;
 300             }
 301             h->mb.cache.non_zero_count[x264_scan8[25]+ch] = array_non_zero( h->dct.chroma_dc[ch] );
 302             continue;
 303         }
 304
 305         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 306         dct2x2dc( dct2x2, dct4x4 );
 307         /* calculate dct coeffs */
 308         for( i = 0; i < 4; i++ )
 309         {
 310             if( h->mb.b_trellis )
 311                 nz = x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 0 );
 312             else
 313                 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] );
 314             h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 315             if( nz )
 316             {
 317                 nz_ac = 1;
 318                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 319                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 320                 if( b_decimate )
 321                     i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[16+i+ch*4] );
 322             }
 323         }
 324
 325         if( h->mb.b_trellis )
 326             nz_dc = x264_quant_dc_trellis( h, (int16_t*)dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter );
 327         else
 328             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 );
 329
 330         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = nz_dc;
 331
 332         if( (b_decimate && i_decimate_score < 7) || !nz_ac )
 333         {
 334             /* Decimate the block */
 335             h->mb.cache.non_zero_count[x264_scan8[16+0]+24*ch] = 0;
 336             h->mb.cache.non_zero_count[x264_scan8[16+1]+24*ch] = 0;
 337             h->mb.cache.non_zero_count[x264_scan8[16+2]+24*ch] = 0;
 338             h->mb.cache.non_zero_count[x264_scan8[16+3]+24*ch] = 0;
 339             if( !nz_dc ) /* Whole block is empty */
 340                 continue;
 341             /* DC-only */
 342             zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 343             idct_dequant_2x2_dconly( dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 344             h->dctf.add8x8_idct_dc( p_dst, dct2x2 );
 345         }
 346         else
 347         {
 348             h->mb.i_cbp_chroma = 1;
 349             if( nz_dc )
 350             {
 351                 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 352                 idct_dequant_2x2_dc( dct2x2, dct4x4, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 353             }
 354             h->dctf.add8x8_idct( p_dst, dct4x4 );
 355         }
 356     }
 357
 358     if( h->mb.i_cbp_chroma )
 359         h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
 360     else if( h->mb.cache.non_zero_count[x264_scan8[25]] |
 361              h->mb.cache.non_zero_count[x264_scan8[26]] )
 362         h->mb.i_cbp_chroma = 1;    /* dc only */
 363 }
 364
 365 static void x264_macroblock_encode_skip( x264_t *h )
 366 {
 367     h->mb.i_cbp_luma = 0x00;
 368     h->mb.i_cbp_chroma = 0x00;
 369     memset( h->mb.cache.non_zero_count, 0, X264_SCAN8_SIZE );
 370     /* store cbp */
 371     h->mb.cbp[h->mb.i_mb_xy] = 0;
 372 }
 373
 374 /*****************************************************************************
 375  * x264_macroblock_encode_pskip:
 376  *  Encode an already marked skip block
 377  *****************************************************************************/
 378 static void x264_macroblock_encode_pskip( x264_t *h )
 379 {
 380     const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 381                                 h->mb.mv_min[0], h->mb.mv_max[0] );
 382     const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 383                                 h->mb.mv_min[1], h->mb.mv_max[1] );
 384
 385     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 386     if( !h->mb.b_skip_mc )
 387     {
 388         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 389                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 390                        mvx, mvy, 16, 16 );
 391
 392         h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 393                          h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 394                          mvx, mvy, 8, 8 );
 395
 396         h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 397                          h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 398                          mvx, mvy, 8, 8 );
 399     }
 400
 401     x264_macroblock_encode_skip( h );
 402 }
 403
 404 /*****************************************************************************
 405  * Intra prediction for predictive lossless mode.
 406  *****************************************************************************/
 407
 408 /* Note that these functions take a shortcut (mc.copy instead of actual pixel prediction) which assumes
 409  * that the edge pixels of the reconstructed frame are the same as that of the source frame.  This means
 410  * they will only work correctly if the neighboring blocks are losslessly coded.  In practice, this means
 411  * lossless mode cannot be mixed with lossy mode within a frame. */
 412 /* This can be resolved by explicitly copying the edge pixels after doing the mc.copy, but this doesn't
 413  * need to be done unless we decide to allow mixing lossless and lossy compression. */
 414
 415 void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
 416 {
 417     int stride = h->fenc->i_stride[1] << h->mb.b_interlaced;
 418     if( i_mode == I_PRED_CHROMA_V )
 419     {
 420         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-stride, stride, 8 );
 421         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-stride, stride, 8 );
 422     }
 423     else if( i_mode == I_PRED_CHROMA_H )
 424     {
 425         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-1, stride, 8 );
 426         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-1, stride, 8 );
 427     }
 428     else
 429     {
 430         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 431         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 432     }
 433 }
 434
 435 void x264_predict_lossless_4x4( x264_t *h, uint8_t *p_dst, int idx, int i_mode )
 436 {
 437     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 438     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
 439
 440     if( i_mode == I_PRED_4x4_V )
 441         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
 442     else if( i_mode == I_PRED_4x4_H )
 443         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-1, stride, 4 );
 444     else
 445         h->predict_4x4[i_mode]( p_dst );
 446 }
 447
 448 void x264_predict_lossless_8x8( x264_t *h, uint8_t *p_dst, int idx, int i_mode, uint8_t edge[33] )
 449 {
 450     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 451     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
 452
 453     if( i_mode == I_PRED_8x8_V )
 454         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
 455     else if( i_mode == I_PRED_8x8_H )
 456         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-1, stride, 8 );
 457     else
 458         h->predict_8x8[i_mode]( p_dst, edge );
 459 }
 460
 461 void x264_predict_lossless_16x16( x264_t *h, int i_mode )
 462 {
 463     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 464     if( i_mode == I_PRED_16x16_V )
 465         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-stride, stride, 16 );
 466     else if( i_mode == I_PRED_16x16_H )
 467         h->mc.copy_16x16_unaligned( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-1, stride, 16 );
 468     else
 469         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 470 }
 471
 472 /*****************************************************************************
 473  * x264_macroblock_encode:
 474  *****************************************************************************/
 475 void x264_macroblock_encode( x264_t *h )
 476 {
 477     int i_cbp_dc = 0;
 478     int i_qp = h->mb.i_qp;
 479     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 480     int b_force_no_skip = 0;
 481     int i,idx,nz;
 482     h->mb.i_cbp_luma = 0;
 483     h->mb.cache.non_zero_count[x264_scan8[24]] = 0;
 484
 485     if( h->sh.b_mbaff
 486         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 487         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 488     {
 489         /* The first skip is predicted to be a frame mb pair.
 490          * We don't yet support the aff part of mbaff, so force it to non-skip
 491          * so that we can pick the aff flag. */
 492         b_force_no_skip = 1;
 493         if( IS_SKIP(h->mb.i_type) )
 494         {
 495             if( h->mb.i_type == P_SKIP )
 496                 h->mb.i_type = P_L0;
 497             else if( h->mb.i_type == B_SKIP )
 498                 h->mb.i_type = B_DIRECT;
 499         }
 500     }
 501
 502     if( h->mb.i_type == P_SKIP )
 503     {
 504         /* A bit special */
 505         x264_macroblock_encode_pskip( h );
 506         return;
 507     }
 508     if( h->mb.i_type == B_SKIP )
 509     {
 510         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 511         if( !h->mb.b_skip_mc )
 512             x264_mb_mc( h );
 513         x264_macroblock_encode_skip( h );
 514         return;
 515     }
 516
 517     if( h->mb.i_type == I_16x16 )
 518     {
 519         const int i_mode = h->mb.i_intra16x16_pred_mode;
 520         h->mb.b_transform_8x8 = 0;
 521
 522         if( h->mb.b_lossless )
 523             x264_predict_lossless_16x16( h, i_mode );
 524         else
 525             h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 526
 527         /* encode the 16x16 macroblock */
 528         x264_mb_encode_i16x16( h, i_qp );
 529     }
 530     else if( h->mb.i_type == I_8x8 )
 531     {
 532         DECLARE_ALIGNED_16( uint8_t edge[33] );
 533         h->mb.b_transform_8x8 = 1;
 534         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 535         if( h->mb.i_skip_intra )
 536         {
 537             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 538             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = h->mb.pic.i8x8_nnz_buf[0];
 539             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = h->mb.pic.i8x8_nnz_buf[1];
 540             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = h->mb.pic.i8x8_nnz_buf[2];
 541             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = h->mb.pic.i8x8_nnz_buf[3];
 542             h->mb.i_cbp_luma = h->mb.pic.i8x8_cbp;
 543             /* In RD mode, restore the now-overwritten DCT data. */
 544             if( h->mb.i_skip_intra == 2 )
 545                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 546         }
 547         for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 548         {
 549             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 550             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 551             x264_predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 552
 553             if( h->mb.b_lossless )
 554                 x264_predict_lossless_8x8( h, p_dst, i, i_mode, edge );
 555             else
 556                 h->predict_8x8[i_mode]( p_dst, edge );
 557
 558             x264_mb_encode_i8x8( h, i, i_qp );
 559         }
 560     }
 561     else if( h->mb.i_type == I_4x4 )
 562     {
 563         h->mb.b_transform_8x8 = 0;
 564         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 565         if( h->mb.i_skip_intra )
 566         {
 567             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 568             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = h->mb.pic.i4x4_nnz_buf[0];
 569             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = h->mb.pic.i4x4_nnz_buf[1];
 570             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = h->mb.pic.i4x4_nnz_buf[2];
 571             *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = h->mb.pic.i4x4_nnz_buf[3];
 572             h->mb.i_cbp_luma = h->mb.pic.i4x4_cbp;
 573             /* In RD mode, restore the now-overwritten DCT data. */
 574             if( h->mb.i_skip_intra == 2 )
 575                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 576         }
 577         for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 578         {
 579             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
 580             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 581
 582             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 583                 /* emulate missing topright samples */
 584                 *(uint32_t*) &p_dst[4-FDEC_STRIDE] = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 585
 586             if( h->mb.b_lossless )
 587                 x264_predict_lossless_4x4( h, p_dst, i, i_mode );
 588             else
 589                 h->predict_4x4[i_mode]( p_dst );
 590             x264_mb_encode_i4x4( h, i, i_qp );
 591         }
 592     }
 593     else    /* Inter MB */
 594     {
 595         int i8x8, i4x4;
 596         int i_decimate_mb = 0;
 597
 598         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
 599         if( !h->mb.b_skip_mc )
 600             x264_mb_mc( h );
 601
 602         if( h->mb.b_lossless )
 603         {
 604             if( h->mb.b_transform_8x8 )
 605                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 606                 {
 607                     int x = 8*(i8x8&1);
 608                     int y = 8*(i8x8>>1);
 609                     h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8],
 610                                         h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
 611                                         h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
 612                     nz = array_non_zero( h->dct.luma8x8[i8x8] );
 613                     STORE_8x8_NNZ(i8x8,nz);
 614                     h->mb.i_cbp_luma |= nz << i8x8;
 615                 }
 616             else
 617                 for( i4x4 = 0; i4x4 < 16; i4x4++ )
 618                 {
 619                     h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 620                                         h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
 621                                         h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4x4] );
 622                     nz = array_non_zero( h->dct.luma4x4[i4x4] );
 623                     h->mb.cache.non_zero_count[x264_scan8[i4x4]] = nz;
 624                     h->mb.i_cbp_luma |= nz << (i4x4>>2);
 625                 }
 626         }
 627         else if( h->mb.b_transform_8x8 )
 628         {
 629             DECLARE_ALIGNED_16( int16_t dct8x8[4][8][8] );
 630             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 631             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 632             h->nr_count[1] += h->mb.b_noise_reduction * 4;
 633
 634             for( idx = 0; idx < 4; idx++ )
 635             {
 636                 if( h->mb.b_noise_reduction )
 637                     h->quantf.denoise_dct( *dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
 638                 nz = x264_quant_8x8( h, dct8x8[idx], i_qp, 0, idx );
 639
 640                 if( nz )
 641                 {
 642                     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 643                     if( b_decimate )
 644                     {
 645                         int i_decimate_8x8 = h->quantf.decimate_score64( h->dct.luma8x8[idx] );
 646                         i_decimate_mb += i_decimate_8x8;
 647                         if( i_decimate_8x8 >= 4 )
 648                             h->mb.i_cbp_luma |= 1<<idx;
 649                     }
 650                     else
 651                         h->mb.i_cbp_luma |= 1<<idx;
 652                 }
 653             }
 654
 655             if( i_decimate_mb < 6 && b_decimate )
 656             {
 657                 h->mb.i_cbp_luma = 0;
 658                 *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = 0;
 659                 *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = 0;
 660                 *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = 0;
 661                 *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = 0;
 662             }
 663             else
 664             {
 665                 for( idx = 0; idx < 4; idx++ )
 666                 {
 667                     if( h->mb.i_cbp_luma&(1<<idx) )
 668                     {
 669                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 670                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 671                         STORE_8x8_NNZ(idx,1);
 672                     }
 673                     else
 674                         STORE_8x8_NNZ(idx,0);
 675                 }
 676             }
 677         }
 678         else
 679         {
 680             DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 681             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 682             h->nr_count[0] += h->mb.b_noise_reduction * 16;
 683
 684             for( i8x8 = 0; i8x8 < 4; i8x8++ )
 685             {
 686                 int i_decimate_8x8 = 0;
 687                 int cbp = 0;
 688
 689                 /* encode one 4x4 block */
 690                 for( i4x4 = 0; i4x4 < 4; i4x4++ )
 691                 {
 692                     idx = i8x8 * 4 + i4x4;
 693
 694                     if( h->mb.b_noise_reduction )
 695                         h->quantf.denoise_dct( *dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
 696                     nz = x264_quant_4x4( h, dct4x4[idx], i_qp, DCT_LUMA_4x4, 0, idx );
 697                     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 698
 699                     if( nz )
 700                     {
 701                         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 702                         h->quantf.dequant_4x4( dct4x4[idx], h->dequant4_mf[CQM_4PY], i_qp );
 703                         if( b_decimate && i_decimate_8x8 < 6 )
 704                             i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[idx] );
 705                         cbp = 1;
 706                     }
 707                 }
 708
 709                 /* decimate this 8x8 block */
 710                 i_decimate_mb += i_decimate_8x8;
 711                 if( b_decimate )
 712                 {
 713                     if( i_decimate_8x8 < 4 )
 714                         STORE_8x8_NNZ(i8x8,0)
 715                     else
 716                         h->mb.i_cbp_luma |= 1<<i8x8;
 717                 }
 718                 else if( cbp )
 719                 {
 720                     h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 721                     h->mb.i_cbp_luma |= 1<<i8x8;
 722                 }
 723             }
 724
 725             if( b_decimate )
 726             {
 727                 if( i_decimate_mb < 6 )
 728                 {
 729                     h->mb.i_cbp_luma = 0;
 730                     *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 0]] = 0;
 731                     *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 2]] = 0;
 732                     *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[ 8]] = 0;
 733                     *(uint32_t*)&h->mb.cache.non_zero_count[x264_scan8[10]] = 0;
 734                 }
 735                 else
 736                 {
 737                     for( i8x8 = 0; i8x8 < 4; i8x8++ )
 738                         if( h->mb.i_cbp_luma&(1<<i8x8) )
 739                             h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 740                 }
 741             }
 742         }
 743     }
 744
 745     /* encode chroma */
 746     if( IS_INTRA( h->mb.i_type ) )
 747     {
 748         const int i_mode = h->mb.i_chroma_pred_mode;
 749         if( h->mb.b_lossless )
 750             x264_predict_lossless_8x8_chroma( h, i_mode );
 751         else
 752         {
 753             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 754             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 755         }
 756     }
 757
 758     /* encode the 8x8 blocks */
 759     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 760
 761     if( h->param.b_cabac )
 762     {
 763         i_cbp_dc = h->mb.cache.non_zero_count[x264_scan8[24]]
 764                  | h->mb.cache.non_zero_count[x264_scan8[25]] << 1
 765                  | h->mb.cache.non_zero_count[x264_scan8[26]] << 2;
 766     }
 767
 768     /* store cbp */
 769     h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
 770
 771     /* Check for P_SKIP
 772      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 773      *      (if multiple mv give same result)*/
 774     if( !b_force_no_skip )
 775     {
 776         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 777             !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
 778             *(uint32_t*)h->mb.cache.mv[0][x264_scan8[0]] == *(uint32_t*)h->mb.cache.pskip_mv
 779             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 780         {
 781             h->mb.i_type = P_SKIP;
 782         }
 783
 784         /* Check for B_SKIP */
 785         if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
 786         {
 787             h->mb.i_type = B_SKIP;
 788         }
 789     }
 790 }
 791
 792 /*****************************************************************************
 793  * x264_macroblock_probe_skip:
 794  *  Check if the current MB could be encoded as a [PB]_SKIP (it supposes you use
 795  *  the previous QP
 796  *****************************************************************************/
 797 int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
 798 {
 799     DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 800     DECLARE_ALIGNED_16( int16_t dct2x2[2][2] );
 801     DECLARE_ALIGNED_16( int16_t dctscan[16] );
 802
 803     int i_qp = h->mb.i_qp;
 804     int mvp[2];
 805     int ch, thresh, ssd;
 806
 807     int i8x8, i4x4;
 808     int i_decimate_mb;
 809
 810     if( !b_bidir )
 811     {
 812         /* Get the MV */
 813         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 814         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 815
 816         /* Motion compensation */
 817         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 818                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 819                        mvp[0], mvp[1], 16, 16 );
 820     }
 821
 822     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 823     {
 824         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
 825         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
 826         /* get luma diff */
 827         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
 828                                     h->mb.pic.p_fdec[0] + fdec_offset );
 829         /* encode one 4x4 block */
 830         for( i4x4 = 0; i4x4 < 4; i4x4++ )
 831         {
 832             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] ) )
 833                 continue;
 834             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 835             i_decimate_mb += h->quantf.decimate_score16( dctscan );
 836             if( i_decimate_mb >= 6 )
 837                 return 0;
 838         }
 839     }
 840
 841     /* encode chroma */
 842     i_qp = h->mb.i_chroma_qp;
 843     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 844
 845     for( ch = 0; ch < 2; ch++ )
 846     {
 847         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 848         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 849
 850         if( !b_bidir )
 851         {
 852             h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 853                              h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 854                              mvp[0], mvp[1], 8, 8 );
 855         }
 856
 857         /* there is almost never a termination during chroma, but we can't avoid the check entirely */
 858         /* so instead we check SSD and skip the actual check if the score is low enough. */
 859         ssd = h->pixf.ssd[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE );
 860         if( ssd < thresh )
 861             continue;
 862
 863         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 864
 865         /* calculate dct DC */
 866         dct2x2dc( dct2x2, dct4x4 );
 867         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 ) )
 868             return 0;
 869
 870         /* If there wasn't a termination in DC, we can check against a much higher threshold. */
 871         if( ssd < thresh*4 )
 872             continue;
 873
 874         /* calculate dct coeffs */
 875         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 876         {
 877             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] ) )
 878                 continue;
 879             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 880             i_decimate_mb += h->quantf.decimate_score15( dctscan );
 881             if( i_decimate_mb >= 7 )
 882                 return 0;
 883         }
 884     }
 885
 886     h->mb.b_skip_mc = 1;
 887     return 1;
 888 }
 889
 890 /****************************************************************************
 891  * DCT-domain noise reduction / adaptive deadzone
 892  * from libavcodec
 893  ****************************************************************************/
 894
 895 void x264_noise_reduction_update( x264_t *h )
 896 {
 897     int cat, i;
 898     for( cat = 0; cat < 2; cat++ )
 899     {
 900         int size = cat ? 64 : 16;
 901         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
 902
 903         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
 904         {
 905             for( i = 0; i < size; i++ )
 906                 h->nr_residual_sum[cat][i] >>= 1;
 907             h->nr_count[cat] >>= 1;
 908         }
 909
 910         for( i = 0; i < size; i++ )
 911             h->nr_offset[cat][i] =
 912                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
 913                  + h->nr_residual_sum[cat][i]/2)
 914               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
 915     }
 916 }
 917
 918 /*****************************************************************************
 919  * RD only; 4 calls to this do not make up for one macroblock_encode.
 920  * doesn't transform chroma dc.
 921  *****************************************************************************/
 922 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
 923 {
 924     int i_qp = h->mb.i_qp;
 925     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
 926     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
 927     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 928     int nnz8x8 = 0;
 929     int ch, nz;
 930
 931     x264_mb_mc_8x8( h, i8 );
 932
 933     if( h->mb.b_lossless )
 934     {
 935         int i4;
 936         if( h->mb.b_transform_8x8 )
 937         {
 938             h->zigzagf.sub_8x8( h->dct.luma8x8[i8], p_fenc, p_fdec );
 939             nnz8x8 = array_non_zero( h->dct.luma8x8[i8] );
 940             STORE_8x8_NNZ(i8,nnz8x8);
 941         }
 942         else
 943         {
 944             for( i4 = i8*4; i4 < i8*4+4; i4++ )
 945             {
 946                 int nz;
 947                 h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
 948                                     h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
 949                                     h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
 950                 nz = array_non_zero( h->dct.luma4x4[i4] );
 951                 h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
 952                 nnz8x8 |= nz;
 953             }
 954         }
 955         for( ch = 0; ch < 2; ch++ )
 956         {
 957             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
 958             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
 959             h->zigzagf.sub_4x4( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec );
 960             h->dct.luma4x4[16+i8+ch*4][0] = 0;
 961             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = array_non_zero( h->dct.luma4x4[16+i8+ch*4] );
 962         }
 963     }
 964     else
 965     {
 966         if( h->mb.b_transform_8x8 )
 967         {
 968             DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 969             h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
 970             nnz8x8 = x264_quant_8x8( h, dct8x8, i_qp, 0, i8 );
 971             if( nnz8x8 )
 972             {
 973                 h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
 974
 975                 if( b_decimate && !h->mb.b_trellis )
 976                     nnz8x8 = 4 <= h->quantf.decimate_score64( h->dct.luma8x8[i8] );
 977
 978                 if( nnz8x8 )
 979                 {
 980                     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
 981                     h->dctf.add8x8_idct8( p_fdec, dct8x8 );
 982                     STORE_8x8_NNZ(i8,1);
 983                 }
 984                 else
 985                     STORE_8x8_NNZ(i8,0);
 986             }
 987             else
 988                 STORE_8x8_NNZ(i8,0);
 989         }
 990         else
 991         {
 992             int i4;
 993             int i_decimate_8x8 = 0;
 994             DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 995             h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
 996             for( i4 = 0; i4 < 4; i4++ )
 997             {
 998                 nz = x264_quant_4x4( h, dct4x4[i4], i_qp, DCT_LUMA_4x4, 0, i8*4+i4 );
 999                 h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = nz;
1000                 if( nz )
1001                 {
1002                     h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
1003                     h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
1004                     if( b_decimate )
1005                         i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[i8*4+i4] );
1006                     nnz8x8 = 1;
1007                 }
1008             }
1009
1010             if( b_decimate && i_decimate_8x8 < 4 )
1011                 nnz8x8 = 0;
1012
1013             if( nnz8x8 )
1014                 h->dctf.add8x8_idct( p_fdec, dct4x4 );
1015             else
1016                 STORE_8x8_NNZ(i8,0);
1017         }
1018
1019         i_qp = h->mb.i_chroma_qp;
1020
1021         for( ch = 0; ch < 2; ch++ )
1022         {
1023             DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
1024             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1025             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1026
1027             h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1028             dct4x4[0][0] = 0;
1029
1030             if( h->mb.b_trellis )
1031                 nz = x264_quant_4x4_trellis( h, dct4x4, CQM_4PC, i_qp, DCT_CHROMA_AC, 0, 0 );
1032             else
1033                 nz = h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
1034
1035             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1036             if( nz )
1037             {
1038                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
1039                 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
1040                 h->dctf.add4x4_idct( p_fdec, dct4x4 );
1041             }
1042         }
1043     }
1044     h->mb.i_cbp_luma &= ~(1 << i8);
1045     h->mb.i_cbp_luma |= nnz8x8 << i8;
1046     h->mb.i_cbp_chroma = 0x02;
1047 }
1048
1049 /*****************************************************************************
1050  * RD only, luma only
1051  *****************************************************************************/
1052 void x264_macroblock_encode_p4x4( x264_t *h, int i4 )
1053 {
1054     int i_qp = h->mb.i_qp;
1055     uint8_t *p_fenc = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[i4]];
1056     uint8_t *p_fdec = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
1057     const int i_ref = h->mb.cache.ref[0][x264_scan8[i4]];
1058     const int mvx   = x264_clip3( h->mb.cache.mv[0][x264_scan8[i4]][0], h->mb.mv_min[0], h->mb.mv_max[0] );
1059     const int mvy   = x264_clip3( h->mb.cache.mv[0][x264_scan8[i4]][1], h->mb.mv_min[1], h->mb.mv_max[1] );
1060     int nz;
1061
1062     h->mc.mc_luma( p_fdec, FDEC_STRIDE, h->mb.pic.p_fref[0][i_ref], h->mb.pic.i_stride[0], mvx + 4*4*block_idx_x[i4], mvy + 4*4*block_idx_y[i4], 4, 4 );
1063
1064     if( h->mb.b_lossless )
1065     {
1066         h->zigzagf.sub_4x4( h->dct.luma4x4[i4], p_fenc, p_fdec );
1067         h->mb.cache.non_zero_count[x264_scan8[i4]] = array_non_zero( h->dct.luma4x4[i4] );
1068     }
1069     else
1070     {
1071         DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
1072         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1073         nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 0, i4 );
1074         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1075         if( nz )
1076         {
1077             h->zigzagf.scan_4x4( h->dct.luma4x4[i4], dct4x4 );
1078             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PY], i_qp );
1079             h->dctf.add4x4_idct( p_fdec, dct4x4 );
1080         }
1081     }
1082 }