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 void 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         x264_quant_4x4_trellis( h, dct, i_quant_cat, i_qp, i_ctxBlockCat, b_intra, idx );
  92     else
  93         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 void 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         x264_quant_8x8_trellis( h, dct, i_quant_cat, i_qp, b_intra, idx );
 101     else
 102         h->quantf.quant_8x8( dct, h->quant8_mf[i_quant_cat][i_qp], h->quant8_bias[i_quant_cat][i_qp] );
 103 }
 104
 105 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qp )
 106 {
 107     uint8_t *p_src = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[idx]];
 108     uint8_t *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[idx]];
 109     DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
 110
 111     if( h->mb.b_lossless )
 112     {
 113         h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
 114         return;
 115     }
 116
 117     h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
 118
 119     x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 1, idx );
 120
 121     if( array_non_zero( dct4x4 ) )
 122     {
 123         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4 );
 124         h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qp );
 125
 126         /* output samples to fdec */
 127         h->dctf.add4x4_idct( p_dst, dct4x4 );
 128     }
 129     else
 130         memset( h->dct.luma4x4[idx], 0, sizeof(h->dct.luma4x4[idx]));
 131 }
 132
 133 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qp )
 134 {
 135     int x = 8 * (idx&1);
 136     int y = 8 * (idx>>1);
 137     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
 138     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
 139     DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 140
 141     if( h->mb.b_lossless )
 142     {
 143         h->zigzagf.sub_8x8( h->dct.luma8x8[idx], p_src, p_dst );
 144         return;
 145     }
 146
 147     h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
 148
 149     x264_quant_8x8( h, dct8x8, i_qp, 1, idx );
 150
 151     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
 152     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qp );
 153     h->dctf.add8x8_idct8( p_dst, dct8x8 );
 154 }
 155
 156 static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
 157 {
 158     uint8_t  *p_src = h->mb.pic.p_fenc[0];
 159     uint8_t  *p_dst = h->mb.pic.p_fdec[0];
 160
 161     DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 162     DECLARE_ALIGNED_16( int16_t dct_dc4x4[4][4] );
 163
 164     int i;
 165
 166     if( h->mb.b_lossless )
 167     {
 168         for( i = 0; i < 16; i++ )
 169         {
 170             int oe = block_idx_xy_fenc[i];
 171             int od = block_idx_xy_fdec[i];
 172             h->zigzagf.sub_4x4( h->dct.luma4x4[i], p_src+oe, p_dst+od );
 173             dct_dc4x4[0][block_idx_yx_1d[i]] = h->dct.luma4x4[i][0];
 174             h->dct.luma4x4[i][0] = 0;
 175         }
 176         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 177         return;
 178     }
 179
 180     h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
 181     for( i = 0; i < 16; i++ )
 182     {
 183         /* copy dc coeff */
 184         dct_dc4x4[0][block_idx_xy_1d[i]] = dct4x4[i][0][0];
 185         dct4x4[i][0][0] = 0;
 186
 187         /* quant/scan/dequant */
 188         x264_quant_4x4( h, dct4x4[i], i_qp, DCT_LUMA_AC, 1, i );
 189
 190         h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[i] );
 191         h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IY], i_qp );
 192     }
 193
 194     h->dctf.dct4x4dc( dct_dc4x4 );
 195     if( h->mb.b_trellis )
 196         x264_quant_dc_trellis( h, (int16_t*)dct_dc4x4, CQM_4IY, i_qp, DCT_LUMA_DC, 1);
 197     else
 198         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 );
 199     h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 200
 201     /* output samples to fdec */
 202     h->dctf.idct4x4dc( dct_dc4x4 );
 203     h->quantf.dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[CQM_4IY], i_qp );  /* XXX not inversed */
 204
 205     /* calculate dct coeffs */
 206     for( i = 0; i < 16; i++ )
 207     {
 208         /* copy dc coeff */
 209         dct4x4[i][0][0] = dct_dc4x4[0][block_idx_xy_1d[i]];
 210     }
 211     /* put pixels to fdec */
 212     h->dctf.add16x16_idct( p_dst, dct4x4 );
 213 }
 214
 215 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
 216 {
 217     int i, ch, nz;
 218     int b_decimate = b_inter && (h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate);
 219     h->mb.i_cbp_chroma = 0;
 220
 221     for( ch = 0; ch < 2; ch++ )
 222     {
 223         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 224         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 225         int i_decimate_score = 0;
 226
 227         DECLARE_ALIGNED_16( int16_t dct2x2[2][2]  );
 228         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 229
 230         if( h->mb.b_lossless )
 231         {
 232             for( i = 0; i < 4; i++ )
 233             {
 234                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 235                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 236                 h->zigzagf.sub_4x4( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od );
 237                 h->dct.chroma_dc[ch][i] = h->dct.luma4x4[16+i+ch*4][0];
 238                 h->dct.luma4x4[16+i+ch*4][0] = 0;
 239                 nz = array_non_zero( h->dct.luma4x4[16+i+ch*4] );
 240                 h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 241                 h->mb.i_cbp_chroma |= nz;
 242             }
 243             h->mb.cache.non_zero_count[x264_scan8[25]+ch] = array_non_zero( h->dct.chroma_dc[ch] );
 244             continue;
 245         }
 246
 247         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 248         dct2x2dc( dct2x2, dct4x4 );
 249         /* calculate dct coeffs */
 250         for( i = 0; i < 4; i++ )
 251         {
 252             if( h->mb.b_trellis )
 253                 x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 0 );
 254             else
 255                 h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[CQM_4IC+b_inter][i_qp], h->quant4_bias[CQM_4IC+b_inter][i_qp] );
 256             h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 257
 258             if( b_decimate )
 259                 i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[16+i+ch*4] );
 260         }
 261
 262         if( h->mb.b_trellis )
 263             x264_quant_dc_trellis( h, (int16_t*)dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter );
 264         else
 265             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 );
 266
 267         if( b_decimate && i_decimate_score < 7 )
 268         {
 269             /* Decimate the block */
 270             h->mb.cache.non_zero_count[x264_scan8[16+0]+24*ch] = 0;
 271             h->mb.cache.non_zero_count[x264_scan8[16+1]+24*ch] = 0;
 272             h->mb.cache.non_zero_count[x264_scan8[16+2]+24*ch] = 0;
 273             h->mb.cache.non_zero_count[x264_scan8[16+3]+24*ch] = 0;
 274             if( !array_non_zero( dct2x2 ) ) /* Whole block is empty */
 275             {
 276                 h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 0;
 277                 continue;
 278             }
 279             /* DC-only */
 280             h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 1;
 281             zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 282             idct_dequant_2x2_dconly( dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 283             h->dctf.add8x8_idct_dc( p_dst, dct2x2 );
 284         }
 285         else
 286         {
 287             for( i = 0; i < 4; i++ )
 288             {
 289                 nz = array_non_zero( h->dct.luma4x4[16+ch*4+i] );
 290                 h->mb.cache.non_zero_count[x264_scan8[16+ch*4+i]] = nz;
 291                 h->mb.i_cbp_chroma |= nz;
 292                 if( nz )
 293                     h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 294             }
 295             /* Don't optimize for the AC-only case--it's very rare */
 296             h->mb.cache.non_zero_count[x264_scan8[25]+ch] = array_non_zero( dct2x2 );
 297             zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 298             idct_dequant_2x2_dc( dct2x2, dct4x4, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 299             h->dctf.add8x8_idct( p_dst, dct4x4 );
 300         }
 301     }
 302
 303     if( h->mb.i_cbp_chroma )
 304         h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
 305     else if( h->mb.cache.non_zero_count[x264_scan8[25]] |
 306              h->mb.cache.non_zero_count[x264_scan8[26]] )
 307         h->mb.i_cbp_chroma = 1;    /* dc only */
 308 }
 309
 310 static void x264_macroblock_encode_skip( x264_t *h )
 311 {
 312     h->mb.i_cbp_luma = 0x00;
 313     h->mb.i_cbp_chroma = 0x00;
 314     memset( h->mb.cache.non_zero_count, 0, X264_SCAN8_SIZE );
 315     /* store cbp */
 316     h->mb.cbp[h->mb.i_mb_xy] = 0;
 317 }
 318
 319 /*****************************************************************************
 320  * x264_macroblock_encode_pskip:
 321  *  Encode an already marked skip block
 322  *****************************************************************************/
 323 static void x264_macroblock_encode_pskip( x264_t *h )
 324 {
 325     const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 326                                 h->mb.mv_min[0], h->mb.mv_max[0] );
 327     const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 328                                 h->mb.mv_min[1], h->mb.mv_max[1] );
 329
 330     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 331     if( !h->mb.b_skip_mc )
 332     {
 333         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 334                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 335                        mvx, mvy, 16, 16 );
 336
 337         h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 338                          h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 339                          mvx, mvy, 8, 8 );
 340
 341         h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 342                          h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 343                          mvx, mvy, 8, 8 );
 344     }
 345
 346     x264_macroblock_encode_skip( h );
 347 }
 348
 349 /*****************************************************************************
 350  * Intra prediction for predictive lossless mode.
 351  *****************************************************************************/
 352
 353 /* Note that these functions take a shortcut (mc.copy instead of actual pixel prediction) which assumes
 354  * that the edge pixels of the reconstructed frame are the same as that of the source frame.  This means
 355  * they will only work correctly if the neighboring blocks are losslessly coded.  In practice, this means
 356  * lossless mode cannot be mixed with lossy mode within a frame. */
 357 /* This can be resolved by explicitly copying the edge pixels after doing the mc.copy, but this doesn't
 358  * need to be done unless we decide to allow mixing lossless and lossy compression. */
 359
 360 void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
 361 {
 362     int stride = h->fenc->i_stride[1] << h->mb.b_interlaced;
 363     if( i_mode == I_PRED_CHROMA_V )
 364     {
 365         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-stride, stride, 8 );
 366         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-stride, stride, 8 );
 367     }
 368     else if( i_mode == I_PRED_CHROMA_H )
 369     {
 370         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-1, stride, 8 );
 371         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-1, stride, 8 );
 372     }
 373     else
 374     {
 375         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 376         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 377     }
 378 }
 379
 380 void x264_predict_lossless_4x4( x264_t *h, uint8_t *p_dst, int idx, int i_mode )
 381 {
 382     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 383     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
 384
 385     if( i_mode == I_PRED_4x4_V )
 386         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
 387     else if( i_mode == I_PRED_4x4_H )
 388         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-1, stride, 4 );
 389     else
 390         h->predict_4x4[i_mode]( p_dst );
 391 }
 392
 393 void x264_predict_lossless_8x8( x264_t *h, uint8_t *p_dst, int idx, int i_mode, uint8_t edge[33] )
 394 {
 395     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 396     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
 397
 398     if( i_mode == I_PRED_8x8_V )
 399         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
 400     else if( i_mode == I_PRED_8x8_H )
 401         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-1, stride, 8 );
 402     else
 403         h->predict_8x8[i_mode]( p_dst, edge );
 404 }
 405
 406 void x264_predict_lossless_16x16( x264_t *h, int i_mode )
 407 {
 408     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 409     if( i_mode == I_PRED_16x16_V )
 410         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-stride, stride, 16 );
 411     else if( i_mode == I_PRED_16x16_H )
 412         h->mc.copy_16x16_unaligned( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-1, stride, 16 );
 413     else
 414         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 415 }
 416
 417 /*****************************************************************************
 418  * x264_macroblock_encode:
 419  *****************************************************************************/
 420 void x264_macroblock_encode( x264_t *h )
 421 {
 422     int i_cbp_dc = 0;
 423     int i_qp = h->mb.i_qp;
 424     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 425     int b_force_no_skip = 0;
 426     int i,j,idx;
 427     uint8_t nnz8x8[4] = {1,1,1,1};
 428
 429     if( h->sh.b_mbaff
 430         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 431         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 432     {
 433         /* The first skip is predicted to be a frame mb pair.
 434          * We don't yet support the aff part of mbaff, so force it to non-skip
 435          * so that we can pick the aff flag. */
 436         b_force_no_skip = 1;
 437         if( IS_SKIP(h->mb.i_type) )
 438         {
 439             if( h->mb.i_type == P_SKIP )
 440                 h->mb.i_type = P_L0;
 441             else if( h->mb.i_type == B_SKIP )
 442                 h->mb.i_type = B_DIRECT;
 443         }
 444     }
 445
 446     if( h->mb.i_type == P_SKIP )
 447     {
 448         /* A bit special */
 449         x264_macroblock_encode_pskip( h );
 450         return;
 451     }
 452     if( h->mb.i_type == B_SKIP )
 453     {
 454         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 455         if( !h->mb.b_skip_mc )
 456             x264_mb_mc( h );
 457         x264_macroblock_encode_skip( h );
 458         return;
 459     }
 460
 461     if( h->mb.i_type == I_16x16 )
 462     {
 463         const int i_mode = h->mb.i_intra16x16_pred_mode;
 464         h->mb.b_transform_8x8 = 0;
 465
 466         if( h->mb.b_lossless )
 467             x264_predict_lossless_16x16( h, i_mode );
 468         else
 469             h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 470
 471         /* encode the 16x16 macroblock */
 472         x264_mb_encode_i16x16( h, i_qp );
 473     }
 474     else if( h->mb.i_type == I_8x8 )
 475     {
 476         DECLARE_ALIGNED_16( uint8_t edge[33] );
 477         h->mb.b_transform_8x8 = 1;
 478         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 479         if( h->mb.i_skip_intra )
 480         {
 481             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 482             /* In RD mode, restore the now-overwritten DCT data. */
 483             if( h->mb.i_skip_intra == 2 )
 484                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 485         }
 486         for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 487         {
 488             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 489             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 490             x264_predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 491
 492             if( h->mb.b_lossless )
 493                 x264_predict_lossless_8x8( h, p_dst, i, i_mode, edge );
 494             else
 495                 h->predict_8x8[i_mode]( p_dst, edge );
 496
 497             x264_mb_encode_i8x8( h, i, i_qp );
 498         }
 499         for( i = 0; i < 4; i++ )
 500             nnz8x8[i] = array_non_zero( h->dct.luma8x8[i] );
 501     }
 502     else if( h->mb.i_type == I_4x4 )
 503     {
 504         h->mb.b_transform_8x8 = 0;
 505         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 506         if( h->mb.i_skip_intra )
 507         {
 508             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 509             /* In RD mode, restore the now-overwritten DCT data. */
 510             if( h->mb.i_skip_intra == 2 )
 511                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 512         }
 513         for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 514         {
 515             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
 516             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 517
 518             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 519                 /* emulate missing topright samples */
 520                 *(uint32_t*) &p_dst[4-FDEC_STRIDE] = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 521
 522             if( h->mb.b_lossless )
 523                 x264_predict_lossless_4x4( h, p_dst, i, i_mode );
 524             else
 525                 h->predict_4x4[i_mode]( p_dst );
 526             x264_mb_encode_i4x4( h, i, i_qp );
 527         }
 528     }
 529     else    /* Inter MB */
 530     {
 531         int i8x8, i4x4;
 532         int i_decimate_mb = 0;
 533
 534         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
 535         if( !h->mb.b_skip_mc )
 536             x264_mb_mc( h );
 537
 538         if( h->mb.b_lossless )
 539         {
 540             if( h->mb.b_transform_8x8 )
 541                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 542                 {
 543                     int x = 8*(i8x8&1);
 544                     int y = 8*(i8x8>>1);
 545                     h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8],
 546                                         h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
 547                                         h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
 548                     nnz8x8[i8x8] = array_non_zero( h->dct.luma8x8[i8x8] );
 549                 }
 550             else
 551                 for( i4x4 = 0; i4x4 < 16; i4x4++ )
 552                 {
 553                     h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 554                                         h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
 555                                         h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4x4] );
 556                 }
 557         }
 558         else if( h->mb.b_transform_8x8 )
 559         {
 560             DECLARE_ALIGNED_16( int16_t dct8x8[4][8][8] );
 561             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 562             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 563             h->nr_count[1] += h->mb.b_noise_reduction * 4;
 564
 565             for( idx = 0; idx < 4; idx++ )
 566             {
 567                 if( h->mb.b_noise_reduction )
 568                     h->quantf.denoise_dct( *dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
 569                 x264_quant_8x8( h, dct8x8[idx], i_qp, 0, idx );
 570
 571                 h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 572
 573                 if( b_decimate )
 574                 {
 575                     int i_decimate_8x8 = h->quantf.decimate_score64( h->dct.luma8x8[idx] );
 576                     i_decimate_mb += i_decimate_8x8;
 577                     if( i_decimate_8x8 < 4 )
 578                         nnz8x8[idx] = 0;
 579                 }
 580                 else
 581                     nnz8x8[idx] = array_non_zero( dct8x8[idx] );
 582             }
 583
 584             if( i_decimate_mb < 6 && b_decimate )
 585                 *(uint32_t*)nnz8x8 = 0;
 586             else
 587             {
 588                 for( idx = 0; idx < 4; idx++ )
 589                     if( nnz8x8[idx] )
 590                     {
 591                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 592                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 593                     }
 594             }
 595         }
 596         else
 597         {
 598             DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 599             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 600             h->nr_count[0] += h->mb.b_noise_reduction * 16;
 601
 602             for( i8x8 = 0; i8x8 < 4; i8x8++ )
 603             {
 604                 int i_decimate_8x8;
 605
 606                 /* encode one 4x4 block */
 607                 i_decimate_8x8 = 0;
 608                 for( i4x4 = 0; i4x4 < 4; i4x4++ )
 609                 {
 610                     idx = i8x8 * 4 + i4x4;
 611
 612                     if( h->mb.b_noise_reduction )
 613                         h->quantf.denoise_dct( *dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
 614                     x264_quant_4x4( h, dct4x4[idx], i_qp, DCT_LUMA_4x4, 0, idx );
 615
 616                     h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 617
 618                     if( b_decimate && i_decimate_8x8 < 6 )
 619                         i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[idx] );
 620                 }
 621
 622                 /* decimate this 8x8 block */
 623                 i_decimate_mb += i_decimate_8x8;
 624                 if( i_decimate_8x8 < 4 && b_decimate )
 625                     nnz8x8[i8x8] = 0;
 626             }
 627
 628             if( i_decimate_mb < 6 && b_decimate )
 629                 *(uint32_t*)nnz8x8 = 0;
 630             else
 631             {
 632                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 633                     if( nnz8x8[i8x8] )
 634                     {
 635                         for( i = 0; i < 4; i++ )
 636                             h->quantf.dequant_4x4( dct4x4[i8x8*4+i], h->dequant4_mf[CQM_4PY], i_qp );
 637                         h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 638                     }
 639             }
 640         }
 641     }
 642
 643     /* encode chroma */
 644     if( IS_INTRA( h->mb.i_type ) )
 645     {
 646         const int i_mode = h->mb.i_chroma_pred_mode;
 647         if( h->mb.b_lossless )
 648             x264_predict_lossless_8x8_chroma( h, i_mode );
 649         else
 650         {
 651             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 652             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 653         }
 654     }
 655
 656     /* encode the 8x8 blocks */
 657     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 658
 659     /* coded block pattern and non_zero_count */
 660     h->mb.i_cbp_luma = 0x00;
 661     if( h->mb.i_type == I_16x16 )
 662     {
 663         for( i = 0; i < 16; i++ )
 664         {
 665             int nz = array_non_zero( h->dct.luma4x4[i] );
 666             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 667             h->mb.i_cbp_luma |= nz;
 668         }
 669         h->mb.i_cbp_luma *= 0xf;
 670         h->mb.cache.non_zero_count[x264_scan8[24]] = array_non_zero( h->dct.luma16x16_dc );
 671     }
 672     else
 673     {
 674         for( i = 0; i < 4; i++)
 675         {
 676             if(!nnz8x8[i])
 677             {
 678                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[0+i*4]] = 0;
 679                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[2+i*4]] = 0;
 680             }
 681             else if( h->mb.b_transform_8x8 )
 682             {
 683                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[0+4*i]] = nnz8x8[i] * 0x0101;
 684                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[2+4*i]] = nnz8x8[i] * 0x0101;
 685                 h->mb.i_cbp_luma |= nnz8x8[i] << i;
 686             }
 687             else
 688             {
 689                 int nz, cbp = 0;
 690                 for( j = 0; j < 4; j++ )
 691                 {
 692                     nz = array_non_zero( h->dct.luma4x4[j+4*i] );
 693                     h->mb.cache.non_zero_count[x264_scan8[j+4*i]] = nz;
 694                     cbp |= nz;
 695                 }
 696                 h->mb.i_cbp_luma |= cbp << i;
 697             }
 698         }
 699         h->mb.cache.non_zero_count[x264_scan8[24]] = 0;
 700     }
 701
 702     if( h->param.b_cabac )
 703     {
 704         i_cbp_dc = h->mb.cache.non_zero_count[x264_scan8[24]]
 705                  | h->mb.cache.non_zero_count[x264_scan8[25]] << 1
 706                  | h->mb.cache.non_zero_count[x264_scan8[26]] << 2;
 707     }
 708
 709     /* store cbp */
 710     h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
 711
 712     /* Check for P_SKIP
 713      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 714      *      (if multiple mv give same result)*/
 715     if( !b_force_no_skip )
 716     {
 717         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 718             !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
 719             *(uint32_t*)h->mb.cache.mv[0][x264_scan8[0]] == *(uint32_t*)h->mb.cache.pskip_mv
 720             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 721         {
 722             h->mb.i_type = P_SKIP;
 723         }
 724
 725         /* Check for B_SKIP */
 726         if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
 727         {
 728             h->mb.i_type = B_SKIP;
 729         }
 730     }
 731 }
 732
 733 /*****************************************************************************
 734  * x264_macroblock_probe_skip:
 735  *  Check if the current MB could be encoded as a [PB]_SKIP (it supposes you use
 736  *  the previous QP
 737  *****************************************************************************/
 738 int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
 739 {
 740     DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 741     DECLARE_ALIGNED_16( int16_t dct2x2[2][2] );
 742     DECLARE_ALIGNED_16( int16_t dctscan[16] );
 743
 744     int i_qp = h->mb.i_qp;
 745     int mvp[2];
 746     int ch, thresh;
 747
 748     int i8x8, i4x4;
 749     int i_decimate_mb;
 750
 751     if( !b_bidir )
 752     {
 753         /* Get the MV */
 754         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 755         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 756
 757         /* Motion compensation */
 758         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 759                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 760                        mvp[0], mvp[1], 16, 16 );
 761     }
 762
 763     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 764     {
 765         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
 766         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
 767         /* get luma diff */
 768         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
 769                                     h->mb.pic.p_fdec[0] + fdec_offset );
 770         /* encode one 4x4 block */
 771         for( i4x4 = 0; i4x4 < 4; i4x4++ )
 772         {
 773             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 774             if( !array_non_zero(dct4x4[i4x4]) )
 775                 continue;
 776             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 777             i_decimate_mb += h->quantf.decimate_score16( dctscan );
 778             if( i_decimate_mb >= 6 )
 779                 return 0;
 780         }
 781     }
 782
 783     /* encode chroma */
 784     i_qp = h->mb.i_chroma_qp;
 785     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 786
 787     for( ch = 0; ch < 2; ch++ )
 788     {
 789         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 790         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 791
 792         if( !b_bidir )
 793         {
 794             h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 795                              h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 796                              mvp[0], mvp[1], 8, 8 );
 797         }
 798
 799         /* there is almost never a termination during chroma, but we can't avoid the check entirely */
 800         /* so instead we check SSD and skip the actual check if the score is low enough. */
 801         if( h->pixf.ssd[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE ) < thresh )
 802             continue;
 803
 804         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 805
 806         /* calculate dct DC */
 807         dct2x2dc( dct2x2, dct4x4 );
 808         h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4PC][i_qp][0]>>1, h->quant4_bias[CQM_4PC][i_qp][0]<<1 );
 809         if( array_non_zero(dct2x2) )
 810             return 0;
 811
 812         /* calculate dct coeffs */
 813         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 814         {
 815             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 816             if( !array_non_zero(dct4x4[i4x4]) )
 817                 continue;
 818             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 819             i_decimate_mb += h->quantf.decimate_score15( dctscan );
 820             if( i_decimate_mb >= 7 )
 821                 return 0;
 822         }
 823     }
 824
 825     h->mb.b_skip_mc = 1;
 826     return 1;
 827 }
 828
 829 /****************************************************************************
 830  * DCT-domain noise reduction / adaptive deadzone
 831  * from libavcodec
 832  ****************************************************************************/
 833
 834 void x264_noise_reduction_update( x264_t *h )
 835 {
 836     int cat, i;
 837     for( cat = 0; cat < 2; cat++ )
 838     {
 839         int size = cat ? 64 : 16;
 840         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
 841
 842         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
 843         {
 844             for( i = 0; i < size; i++ )
 845                 h->nr_residual_sum[cat][i] >>= 1;
 846             h->nr_count[cat] >>= 1;
 847         }
 848
 849         for( i = 0; i < size; i++ )
 850             h->nr_offset[cat][i] =
 851                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
 852                  + h->nr_residual_sum[cat][i]/2)
 853               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
 854     }
 855 }
 856
 857 /*****************************************************************************
 858  * RD only; 4 calls to this do not make up for one macroblock_encode.
 859  * doesn't transform chroma dc.
 860  *****************************************************************************/
 861 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
 862 {
 863     int i_qp = h->mb.i_qp;
 864     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
 865     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
 866     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 867     int nnz8x8 = 0;
 868     int ch;
 869
 870     x264_mb_mc_8x8( h, i8 );
 871
 872     if( h->mb.b_lossless )
 873     {
 874         int i4;
 875         if( h->mb.b_transform_8x8 )
 876         {
 877             h->zigzagf.sub_8x8( h->dct.luma8x8[i8], p_fenc, p_fdec );
 878             nnz8x8 = array_non_zero( h->dct.luma8x8[i8] );
 879             *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+0]] = 0x0101 * nnz8x8;
 880             *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+2]] = 0x0101 * nnz8x8;
 881         }
 882         else
 883         {
 884             for( i4 = i8*4; i4 < i8*4+4; i4++ )
 885             {
 886                 int nz;
 887                 h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
 888                                     h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
 889                                     h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
 890                 nz = array_non_zero( h->dct.luma4x4[i4] );
 891                 h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
 892                 nnz8x8 |= nz;
 893             }
 894         }
 895         for( ch = 0; ch < 2; ch++ )
 896         {
 897             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
 898             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
 899             h->zigzagf.sub_4x4( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec );
 900             h->dct.luma4x4[16+i8+ch*4][0] = 0;
 901         }
 902         h->mb.cache.non_zero_count[x264_scan8[16+i8]] = array_non_zero( h->dct.luma4x4[16+i8] );
 903         h->mb.cache.non_zero_count[x264_scan8[20+i8]] = array_non_zero( h->dct.luma4x4[20+i8] );
 904     }
 905     else
 906     {
 907         if( h->mb.b_transform_8x8 )
 908         {
 909             DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 910             h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
 911             x264_quant_8x8( h, dct8x8, i_qp, 0, i8 );
 912             h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
 913
 914             if( b_decimate && !h->mb.b_trellis )
 915                 nnz8x8 = 4 <= h->quantf.decimate_score64( h->dct.luma8x8[i8] );
 916             else
 917                 nnz8x8 = array_non_zero( dct8x8 );
 918
 919             if( nnz8x8 )
 920             {
 921                 h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
 922                 h->dctf.add8x8_idct8( p_fdec, dct8x8 );
 923                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+0]] = 0x0101;
 924                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+2]] = 0x0101;
 925             }
 926             else
 927             {
 928                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+0]] = 0;
 929                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+2]] = 0;
 930             }
 931         }
 932         else
 933         {
 934             int i4;
 935             DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 936             h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
 937             for( i4 = 0; i4 < 4; i4++ )
 938                 x264_quant_4x4( h, dct4x4[i4], i_qp, DCT_LUMA_4x4, 0, i8*4+i4 );
 939
 940             for( i4 = 0; i4 < 4; i4++ )
 941                 h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
 942
 943             if( b_decimate )
 944             {
 945                 int i_decimate_8x8 = 0;
 946                 for( i4 = 0; i4 < 4 && i_decimate_8x8 < 4; i4++ )
 947                     i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[i8*4+i4] );
 948                 nnz8x8 = 4 <= i_decimate_8x8;
 949             }
 950             else
 951                 nnz8x8 = array_non_zero( dct4x4 );
 952
 953             if( nnz8x8 )
 954             {
 955                 for( i4 = 0; i4 < 4; i4++ )
 956                 {
 957                     if( array_non_zero( dct4x4[i4] ) )
 958                     {
 959                         h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
 960                         h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = 1;
 961                     }
 962                     else
 963                         h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = 0;
 964                 }
 965                 h->dctf.add8x8_idct( p_fdec, dct4x4 );
 966             }
 967             else
 968             {
 969                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+0]] = 0;
 970                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[i8*4+2]] = 0;
 971             }
 972         }
 973
 974         i_qp = h->mb.i_chroma_qp;
 975
 976         for( ch = 0; ch < 2; ch++ )
 977         {
 978             DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
 979             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
 980             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
 981
 982             h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
 983             dct4x4[0][0] = 0;
 984
 985             if( h->mb.b_trellis )
 986                 x264_quant_4x4_trellis( h, dct4x4, CQM_4PC, i_qp, DCT_CHROMA_AC, 0, 0 );
 987             else
 988                 h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 989
 990             if( array_non_zero( dct4x4 ) )
 991             {
 992                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
 993                 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
 994                 h->dctf.add4x4_idct( p_fdec, dct4x4 );
 995                 h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = 1;
 996             }
 997             else
 998                 h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = 0;
 999         }
1000     }
1001     h->mb.i_cbp_luma &= ~(1 << i8);
1002     h->mb.i_cbp_luma |= nnz8x8 << i8;
1003     h->mb.i_cbp_chroma = 0x02;
1004 }
1005
1006 /*****************************************************************************
1007  * RD only, luma only
1008  *****************************************************************************/
1009 void x264_macroblock_encode_p4x4( x264_t *h, int i4 )
1010 {
1011     int i_qp = h->mb.i_qp;
1012     uint8_t *p_fenc = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[i4]];
1013     uint8_t *p_fdec = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
1014     const int i_ref = h->mb.cache.ref[0][x264_scan8[i4]];
1015     const int mvx   = x264_clip3( h->mb.cache.mv[0][x264_scan8[i4]][0], h->mb.mv_min[0], h->mb.mv_max[0] );
1016     const int mvy   = x264_clip3( h->mb.cache.mv[0][x264_scan8[i4]][1], h->mb.mv_min[1], h->mb.mv_max[1] );
1017
1018     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 );
1019
1020     if( h->mb.b_lossless )
1021     {
1022         h->zigzagf.sub_4x4( h->dct.luma4x4[i4], p_fenc, p_fdec );
1023         h->mb.cache.non_zero_count[x264_scan8[i4]] = array_non_zero( h->dct.luma4x4[i4] );
1024     }
1025     else
1026     {
1027         DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
1028         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1029         x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 0, i4 );
1030         if( array_non_zero( dct4x4 ) )
1031         {
1032             h->zigzagf.scan_4x4( h->dct.luma4x4[i4], dct4x4 );
1033             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PY], i_qp );
1034             h->dctf.add4x4_idct( p_fdec, dct4x4 );
1035             h->mb.cache.non_zero_count[x264_scan8[i4]] = 1;
1036         }
1037         else
1038             h->mb.cache.non_zero_count[x264_scan8[i4]] = 0;
1039     }
1040 }