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 #define ZIG(i,y,x) level[i] = dct[x][y];
  29 static inline void zigzag_scan_2x2_dc( int16_t level[4], int16_t dct[2][2] )
  30 {
  31     ZIG(0,0,0)
  32     ZIG(1,0,1)
  33     ZIG(2,1,0)
  34     ZIG(3,1,1)
  35 }
  36 #undef ZIG
  37
  38 /* (ref: JVT-B118)
  39  * x264_mb_decimate_score: given dct coeffs it returns a score to see if we could empty this dct coeffs
  40  * to 0 (low score means set it to null)
  41  * Used in inter macroblock (luma and chroma)
  42  *  luma: for a 8x8 block: if score < 4 -> null
  43  *        for the complete mb: if score < 6 -> null
  44  *  chroma: for the complete mb: if score < 7 -> null
  45  */
  46 static int x264_mb_decimate_score( int16_t *dct, int i_max )
  47 {
  48     static const int i_ds_table4[16] = {
  49         3,2,2,1,1,1,0,0,0,0,0,0,0,0,0,0 };
  50     static const int i_ds_table8[64] = {
  51         3,3,3,3,2,2,2,2,2,2,2,2,1,1,1,1,
  52         1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,
  53         0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
  54         0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 };
  55
  56     const int *ds_table = (i_max == 64) ? i_ds_table8 : i_ds_table4;
  57     int i_score = 0;
  58     int idx = i_max - 1;
  59
  60     while( idx >= 0 && dct[idx] == 0 )
  61         idx--;
  62
  63     while( idx >= 0 )
  64     {
  65         int i_run;
  66
  67         if( (unsigned)(dct[idx--] + 1) > 2 )
  68             return 9;
  69
  70         i_run = 0;
  71         while( idx >= 0 && dct[idx] == 0 )
  72         {
  73             idx--;
  74             i_run++;
  75         }
  76         i_score += ds_table[i_run];
  77     }
  78
  79     return i_score;
  80 }
  81
  82 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qscale )
  83 {
  84     uint8_t *p_src = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[idx]];
  85     uint8_t *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[idx]];
  86     DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
  87
  88     if( h->mb.b_lossless )
  89     {
  90         h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
  91         return;
  92     }
  93
  94     h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
  95
  96     if( h->mb.b_trellis )
  97         x264_quant_4x4_trellis( h, dct4x4, CQM_4IY, i_qscale, DCT_LUMA_4x4, 1 );
  98     else
  99         h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4IY][i_qscale], h->quant4_bias[CQM_4IY][i_qscale] );
 100
 101     if( array_non_zero( dct4x4 ) )
 102     {
 103         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4 );
 104         h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qscale );
 105
 106         /* output samples to fdec */
 107         h->dctf.add4x4_idct( p_dst, dct4x4 );
 108     }
 109     else
 110         memset( h->dct.luma4x4[idx], 0, sizeof(h->dct.luma4x4[idx]));
 111 }
 112
 113 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qscale )
 114 {
 115     int x = 8 * (idx&1);
 116     int y = 8 * (idx>>1);
 117     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
 118     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
 119     DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 120
 121     h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
 122
 123     if( h->mb.b_trellis )
 124         x264_quant_8x8_trellis( h, dct8x8, CQM_8IY, i_qscale, 1 );
 125     else
 126         h->quantf.quant_8x8( dct8x8, h->quant8_mf[CQM_8IY][i_qscale], h->quant8_bias[CQM_8IY][i_qscale] );
 127
 128     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
 129     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qscale );
 130     h->dctf.add8x8_idct8( p_dst, dct8x8 );
 131 }
 132
 133 static void x264_mb_encode_i16x16( x264_t *h, int i_qscale )
 134 {
 135     uint8_t  *p_src = h->mb.pic.p_fenc[0];
 136     uint8_t  *p_dst = h->mb.pic.p_fdec[0];
 137
 138     DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 139     DECLARE_ALIGNED_16( int16_t dct_dc4x4[4][4] );
 140
 141     int i;
 142
 143     if( h->mb.b_lossless )
 144     {
 145         for( i = 0; i < 16; i++ )
 146         {
 147             int oe = block_idx_xy_fenc[i];
 148             int od = block_idx_xy_fdec[i];
 149             h->zigzagf.sub_4x4( h->dct.luma4x4[i], p_src+oe, p_dst+od );
 150             dct_dc4x4[0][block_idx_yx_1d[i]] = h->dct.luma4x4[i][0];
 151             h->dct.luma4x4[i][0] = 0;
 152         }
 153         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 154         return;
 155     }
 156
 157     h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
 158     for( i = 0; i < 16; i++ )
 159     {
 160         /* copy dc coeff */
 161         dct_dc4x4[0][block_idx_xy_1d[i]] = dct4x4[i][0][0];
 162         dct4x4[i][0][0] = 0;
 163
 164         /* quant/scan/dequant */
 165         if( h->mb.b_trellis )
 166             x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IY, i_qscale, DCT_LUMA_AC, 1 );
 167         else
 168             h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[CQM_4IY][i_qscale], h->quant4_bias[CQM_4IY][i_qscale] );
 169
 170         h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[i] );
 171         h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IY], i_qscale );
 172     }
 173
 174     h->dctf.dct4x4dc( dct_dc4x4 );
 175     h->quantf.quant_4x4_dc( dct_dc4x4, h->quant4_mf[CQM_4IY][i_qscale][0]>>1, h->quant4_bias[CQM_4IY][i_qscale][0]<<1 );
 176     h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 177
 178     /* output samples to fdec */
 179     h->dctf.idct4x4dc( dct_dc4x4 );
 180     x264_mb_dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[CQM_4IY], i_qscale );  /* XXX not inversed */
 181
 182     /* calculate dct coeffs */
 183     for( i = 0; i < 16; i++ )
 184     {
 185         /* copy dc coeff */
 186         dct4x4[i][0][0] = dct_dc4x4[0][block_idx_xy_1d[i]];
 187     }
 188     /* put pixels to fdec */
 189     h->dctf.add16x16_idct( p_dst, dct4x4 );
 190 }
 191
 192 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qscale )
 193 {
 194     int i, ch;
 195     int b_decimate = b_inter && (h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate);
 196
 197     for( ch = 0; ch < 2; ch++ )
 198     {
 199         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 200         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 201         int i_decimate_score = 0;
 202
 203         DECLARE_ALIGNED_16( int16_t dct2x2[2][2]  );
 204         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 205
 206         if( h->mb.b_lossless )
 207         {
 208             for( i = 0; i < 4; i++ )
 209             {
 210                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 211                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 212                 h->zigzagf.sub_4x4( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od );
 213                 h->dct.chroma_dc[ch][i] = h->dct.luma4x4[16+i+ch*4][0];
 214                 h->dct.luma4x4[16+i+ch*4][0] = 0;
 215             }
 216             continue;
 217         }
 218
 219         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 220         /* calculate dct coeffs */
 221         for( i = 0; i < 4; i++ )
 222         {
 223             /* copy dc coeff */
 224             dct2x2[i>>1][i&1] = dct4x4[i][0][0];
 225             dct4x4[i][0][0] = 0;
 226
 227             /* no trellis; it doesn't seem to help chroma noticeably */
 228             h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[CQM_4IC+b_inter][i_qscale], h->quant4_bias[CQM_4IC+b_inter][i_qscale] );
 229             h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 230
 231             if( b_decimate )
 232             {
 233                 i_decimate_score += x264_mb_decimate_score( h->dct.luma4x4[16+i+ch*4]+1, 15 );
 234             }
 235         }
 236
 237         h->dctf.dct2x2dc( dct2x2 );
 238         h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4IC+b_inter][i_qscale][0]>>1, h->quant4_bias[CQM_4IC+b_inter][i_qscale][0]<<1 );
 239         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 240
 241         /* output samples to fdec */
 242         h->dctf.idct2x2dc( dct2x2 );
 243         x264_mb_dequant_2x2_dc( dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qscale );  /* XXX not inversed */
 244
 245         if( b_decimate && i_decimate_score < 7 )
 246         {
 247             /* Near null chroma 8x8 block so make it null (bits saving) */
 248             memset( &h->dct.luma4x4[16+ch*4], 0, 4 * sizeof( *h->dct.luma4x4 ) );
 249             if( !array_non_zero( dct2x2 ) )
 250                 continue;
 251             memset( dct4x4, 0, sizeof( dct4x4 ) );
 252         }
 253         else
 254         {
 255             for( i = 0; i < 4; i++ )
 256                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qscale );
 257         }
 258         dct4x4[0][0][0] = dct2x2[0][0];
 259         dct4x4[1][0][0] = dct2x2[0][1];
 260         dct4x4[2][0][0] = dct2x2[1][0];
 261         dct4x4[3][0][0] = dct2x2[1][1];
 262         h->dctf.add8x8_idct( p_dst, dct4x4 );
 263     }
 264
 265     /* coded block pattern */
 266     h->mb.i_cbp_chroma = 0;
 267     for( i = 0; i < 8; i++ )
 268     {
 269         int nz = array_non_zero( h->dct.luma4x4[16+i] );
 270         h->mb.cache.non_zero_count[x264_scan8[16+i]] = nz;
 271         h->mb.i_cbp_chroma |= nz;
 272     }
 273     if( h->mb.i_cbp_chroma )
 274         h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
 275     else if( array_non_zero( h->dct.chroma_dc ) )
 276         h->mb.i_cbp_chroma = 1;    /* dc only */
 277 }
 278
 279 static void x264_macroblock_encode_skip( x264_t *h )
 280 {
 281     h->mb.i_cbp_luma = 0x00;
 282     h->mb.i_cbp_chroma = 0x00;
 283     memset( h->mb.cache.non_zero_count, 0, X264_SCAN8_SIZE );
 284     /* store cbp */
 285     h->mb.cbp[h->mb.i_mb_xy] = 0;
 286 }
 287
 288 /*****************************************************************************
 289  * x264_macroblock_encode_pskip:
 290  *  Encode an already marked skip block
 291  *****************************************************************************/
 292 static void x264_macroblock_encode_pskip( x264_t *h )
 293 {
 294     const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 295                                 h->mb.mv_min[0], h->mb.mv_max[0] );
 296     const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 297                                 h->mb.mv_min[1], h->mb.mv_max[1] );
 298
 299     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 300     if( !h->mb.b_skip_mc )
 301     {
 302         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 303                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 304                        mvx, mvy, 16, 16 );
 305
 306         h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 307                          h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 308                          mvx, mvy, 8, 8 );
 309
 310         h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 311                          h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 312                          mvx, mvy, 8, 8 );
 313     }
 314
 315     x264_macroblock_encode_skip( h );
 316 }
 317
 318 /*****************************************************************************
 319  * x264_macroblock_encode:
 320  *****************************************************************************/
 321 void x264_macroblock_encode( x264_t *h )
 322 {
 323     int i_cbp_dc = 0;
 324     int i_qp = h->mb.i_qp;
 325     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 326     int b_force_no_skip = 0;
 327     int i,j,idx;
 328     uint8_t nnz8x8[4] = {1,1,1,1};
 329
 330     if( h->sh.b_mbaff
 331         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 332         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 333     {
 334         /* The first skip is predicted to be a frame mb pair.
 335          * We don't yet support the aff part of mbaff, so force it to non-skip
 336          * so that we can pick the aff flag. */
 337         b_force_no_skip = 1;
 338         if( IS_SKIP(h->mb.i_type) )
 339         {
 340             if( h->mb.i_type == P_SKIP )
 341                 h->mb.i_type = P_L0;
 342             else if( h->mb.i_type == B_SKIP )
 343                 h->mb.i_type = B_DIRECT;
 344         }
 345     }
 346
 347     if( h->mb.i_type == P_SKIP )
 348     {
 349         /* A bit special */
 350         x264_macroblock_encode_pskip( h );
 351         return;
 352     }
 353     if( h->mb.i_type == B_SKIP )
 354     {
 355         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 356         if( !h->mb.b_skip_mc )
 357             x264_mb_mc( h );
 358         x264_macroblock_encode_skip( h );
 359         return;
 360     }
 361
 362     if( h->mb.i_type == I_16x16 )
 363     {
 364         const int i_mode = h->mb.i_intra16x16_pred_mode;
 365         h->mb.b_transform_8x8 = 0;
 366         /* do the right prediction */
 367         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 368
 369         /* encode the 16x16 macroblock */
 370         x264_mb_encode_i16x16( h, i_qp );
 371     }
 372     else if( h->mb.i_type == I_8x8 )
 373     {
 374         DECLARE_ALIGNED_16( uint8_t edge[33] );
 375         h->mb.b_transform_8x8 = 1;
 376         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 377         if( h->mb.i_skip_intra )
 378         {
 379             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 380             /* In RD mode, restore the now-overwritten DCT data. */
 381             if( h->mb.i_skip_intra == 2 )
 382                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 383         }
 384         for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 385         {
 386             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 387             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 388
 389             x264_predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 390             h->predict_8x8[i_mode]( p_dst, edge );
 391             x264_mb_encode_i8x8( h, i, i_qp );
 392         }
 393         for( i = 0; i < 4; i++ )
 394             nnz8x8[i] = array_non_zero( h->dct.luma8x8[i] );
 395     }
 396     else if( h->mb.i_type == I_4x4 )
 397     {
 398         h->mb.b_transform_8x8 = 0;
 399         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 400         if( h->mb.i_skip_intra )
 401         {
 402             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 403             /* In RD mode, restore the now-overwritten DCT data. */
 404             if( h->mb.i_skip_intra == 2 )
 405                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 406         }
 407         for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 408         {
 409             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
 410             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 411
 412             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 413                 /* emulate missing topright samples */
 414                 *(uint32_t*) &p_dst[4-FDEC_STRIDE] = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 415
 416             h->predict_4x4[i_mode]( p_dst );
 417             x264_mb_encode_i4x4( h, i, i_qp );
 418         }
 419     }
 420     else    /* Inter MB */
 421     {
 422         int i8x8, i4x4;
 423         int i_decimate_mb = 0;
 424
 425         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
 426         if( !h->mb.b_skip_mc )
 427             x264_mb_mc( h );
 428
 429         if( h->mb.b_lossless )
 430         {
 431             for( i4x4 = 0; i4x4 < 16; i4x4++ )
 432             {
 433                 h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 434                                     h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
 435                                     h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4x4] );
 436             }
 437         }
 438         else if( h->mb.b_transform_8x8 )
 439         {
 440             DECLARE_ALIGNED_16( int16_t dct8x8[4][8][8] );
 441             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 442             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 443             h->nr_count[1] += h->mb.b_noise_reduction * 4;
 444
 445             for( idx = 0; idx < 4; idx++ )
 446             {
 447                 if( h->mb.b_noise_reduction )
 448                     h->quantf.denoise_dct( *dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
 449                 if( h->mb.b_trellis )
 450                     x264_quant_8x8_trellis( h, dct8x8[idx], CQM_8PY, i_qp, 0 );
 451                 else
 452                     h->quantf.quant_8x8( dct8x8[idx], h->quant8_mf[CQM_8PY][i_qp], h->quant8_bias[CQM_8PY][i_qp] );
 453
 454                 h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 455
 456                 if( b_decimate )
 457                 {
 458                     int i_decimate_8x8 = x264_mb_decimate_score( h->dct.luma8x8[idx], 64 );
 459                     i_decimate_mb += i_decimate_8x8;
 460                     if( i_decimate_8x8 < 4 )
 461                         nnz8x8[idx] = 0;
 462                 }
 463                 else
 464                     nnz8x8[idx] = array_non_zero( dct8x8[idx] );
 465             }
 466
 467             if( i_decimate_mb < 6 && b_decimate )
 468                 *(uint32_t*)nnz8x8 = 0;
 469             else
 470             {
 471                 for( idx = 0; idx < 4; idx++ )
 472                     if( nnz8x8[idx] )
 473                     {
 474                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 475                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 476                     }
 477             }
 478         }
 479         else
 480         {
 481             DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 482             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 483             h->nr_count[0] += h->mb.b_noise_reduction * 16;
 484
 485             for( i8x8 = 0; i8x8 < 4; i8x8++ )
 486             {
 487                 int i_decimate_8x8;
 488
 489                 /* encode one 4x4 block */
 490                 i_decimate_8x8 = 0;
 491                 for( i4x4 = 0; i4x4 < 4; i4x4++ )
 492                 {
 493                     idx = i8x8 * 4 + i4x4;
 494
 495                     if( h->mb.b_noise_reduction )
 496                         h->quantf.denoise_dct( *dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
 497                     if( h->mb.b_trellis )
 498                         x264_quant_4x4_trellis( h, dct4x4[idx], CQM_4PY, i_qp, DCT_LUMA_4x4, 0 );
 499                     else
 500                         h->quantf.quant_4x4( dct4x4[idx], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 501
 502                     h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 503
 504                     if( b_decimate && i_decimate_8x8 <= 6 )
 505                         i_decimate_8x8 += x264_mb_decimate_score( h->dct.luma4x4[idx], 16 );
 506                 }
 507
 508                 /* decimate this 8x8 block */
 509                 i_decimate_mb += i_decimate_8x8;
 510                 if( i_decimate_8x8 < 4 && b_decimate )
 511                     nnz8x8[i8x8] = 0;
 512             }
 513
 514             if( i_decimate_mb < 6 && b_decimate )
 515                 *(uint32_t*)nnz8x8 = 0;
 516             else
 517             {
 518                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 519                     if( nnz8x8[i8x8] )
 520                     {
 521                         for( i = 0; i < 4; i++ )
 522                             h->quantf.dequant_4x4( dct4x4[i8x8*4+i], h->dequant4_mf[CQM_4PY], i_qp );
 523                         h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 524                     }
 525             }
 526         }
 527     }
 528
 529     /* encode chroma */
 530     if( IS_INTRA( h->mb.i_type ) )
 531     {
 532         const int i_mode = h->mb.i_chroma_pred_mode;
 533         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 534         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 535     }
 536
 537     /* encode the 8x8 blocks */
 538     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 539
 540     /* coded block pattern and non_zero_count */
 541     h->mb.i_cbp_luma = 0x00;
 542     if( h->mb.i_type == I_16x16 )
 543     {
 544         for( i = 0; i < 16; i++ )
 545         {
 546             int nz = array_non_zero( h->dct.luma4x4[i] );
 547             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 548             h->mb.i_cbp_luma |= nz;
 549         }
 550         h->mb.i_cbp_luma *= 0xf;
 551     }
 552     else
 553     {
 554         for( i = 0; i < 4; i++)
 555         {
 556             if(!nnz8x8[i])
 557             {
 558                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[0+i*4]] = 0;
 559                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[2+i*4]] = 0;
 560             }
 561             else if( h->mb.b_transform_8x8 )
 562             {
 563                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[0+4*i]] = nnz8x8[i] * 0x0101;
 564                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[2+4*i]] = nnz8x8[i] * 0x0101;
 565                 h->mb.i_cbp_luma |= nnz8x8[i] << i;
 566             }
 567             else
 568             {
 569                 int nz, cbp = 0;
 570                 for( j = 0; j < 4; j++ )
 571                 {
 572                     nz = array_non_zero( h->dct.luma4x4[j+4*i] );
 573                     h->mb.cache.non_zero_count[x264_scan8[j+4*i]] = nz;
 574                     cbp |= nz;
 575                 }
 576                 h->mb.i_cbp_luma |= cbp << i;
 577             }
 578         }
 579     }
 580
 581     if( h->param.b_cabac )
 582     {
 583         i_cbp_dc = ( h->mb.i_type == I_16x16 && array_non_zero( h->dct.luma16x16_dc ) )
 584                  | array_non_zero( h->dct.chroma_dc[0] ) << 1
 585                  | array_non_zero( h->dct.chroma_dc[1] ) << 2;
 586     }
 587
 588     /* store cbp */
 589     h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
 590
 591     /* Check for P_SKIP
 592      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 593      *      (if multiple mv give same result)*/
 594     if( !b_force_no_skip )
 595     {
 596         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 597             !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
 598             *(uint32_t*)h->mb.cache.mv[0][x264_scan8[0]] == *(uint32_t*)h->mb.cache.pskip_mv
 599             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 600         {
 601             h->mb.i_type = P_SKIP;
 602         }
 603
 604         /* Check for B_SKIP */
 605         if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
 606         {
 607             h->mb.i_type = B_SKIP;
 608         }
 609     }
 610 }
 611
 612 /*****************************************************************************
 613  * x264_macroblock_probe_skip:
 614  *  Check if the current MB could be encoded as a [PB]_SKIP (it supposes you use
 615  *  the previous QP
 616  *****************************************************************************/
 617 int x264_macroblock_probe_skip( x264_t *h, const int b_bidir )
 618 {
 619     DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 620     DECLARE_ALIGNED_16( int16_t dct2x2[2][2] );
 621     DECLARE_ALIGNED_16( int16_t dctscan[16] );
 622
 623     int i_qp = h->mb.i_qp;
 624     int mvp[2];
 625     int ch, thresh;
 626
 627     int i8x8, i4x4;
 628     int i_decimate_mb;
 629
 630     if( !b_bidir )
 631     {
 632         /* Get the MV */
 633         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 634         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 635
 636         /* Motion compensation */
 637         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 638                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 639                        mvp[0], mvp[1], 16, 16 );
 640     }
 641
 642     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 643     {
 644         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
 645         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
 646         /* get luma diff */
 647         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
 648                                     h->mb.pic.p_fdec[0] + fdec_offset );
 649         /* encode one 4x4 block */
 650         for( i4x4 = 0; i4x4 < 4; i4x4++ )
 651         {
 652             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 653             if( !array_non_zero(dct4x4[i4x4]) )
 654                 continue;
 655             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 656             i_decimate_mb += x264_mb_decimate_score( dctscan, 16 );
 657             if( i_decimate_mb >= 6 )
 658                 return 0;
 659         }
 660     }
 661
 662     /* encode chroma */
 663     i_qp = h->mb.i_chroma_qp;
 664     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 665
 666     for( ch = 0; ch < 2; ch++ )
 667     {
 668         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 669         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 670
 671         if( !b_bidir )
 672         {
 673             h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 674                              h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 675                              mvp[0], mvp[1], 8, 8 );
 676         }
 677
 678         /* there is almost never a termination during chroma, but we can't avoid the check entirely */
 679         /* so instead we check SSD and skip the actual check if the score is low enough. */
 680         if( h->pixf.ssd[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE ) < thresh )
 681             continue;
 682
 683         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 684
 685         /* calculate dct DC */
 686         dct2x2[0][0] = dct4x4[0][0][0];
 687         dct2x2[0][1] = dct4x4[1][0][0];
 688         dct2x2[1][0] = dct4x4[2][0][0];
 689         dct2x2[1][1] = dct4x4[3][0][0];
 690         h->dctf.dct2x2dc( dct2x2 );
 691         h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4PC][i_qp][0]>>1, h->quant4_bias[CQM_4PC][i_qp][0]<<1 );
 692         if( array_non_zero(dct2x2) )
 693             return 0;
 694
 695         /* calculate dct coeffs */
 696         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 697         {
 698             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 699             if( !array_non_zero(dct4x4[i4x4]) )
 700                 continue;
 701             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 702             i_decimate_mb += x264_mb_decimate_score( dctscan+1, 15 );
 703             if( i_decimate_mb >= 7 )
 704                 return 0;
 705         }
 706     }
 707
 708     h->mb.b_skip_mc = 1;
 709     return 1;
 710 }
 711
 712 /****************************************************************************
 713  * DCT-domain noise reduction / adaptive deadzone
 714  * from libavcodec
 715  ****************************************************************************/
 716
 717 void x264_noise_reduction_update( x264_t *h )
 718 {
 719     int cat, i;
 720     for( cat = 0; cat < 2; cat++ )
 721     {
 722         int size = cat ? 64 : 16;
 723         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
 724
 725         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
 726         {
 727             for( i = 0; i < size; i++ )
 728                 h->nr_residual_sum[cat][i] >>= 1;
 729             h->nr_count[cat] >>= 1;
 730         }
 731
 732         for( i = 0; i < size; i++ )
 733             h->nr_offset[cat][i] =
 734                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
 735                  + h->nr_residual_sum[cat][i]/2)
 736               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
 737     }
 738 }
 739
 740 /*****************************************************************************
 741  * RD only; 4 calls to this do not make up for one macroblock_encode.
 742  * doesn't transform chroma dc.
 743  *****************************************************************************/
 744 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
 745 {
 746     int i_qp = h->mb.i_qp;
 747     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
 748     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
 749     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 750     int nnz8x8 = 0;
 751     int ch;
 752
 753     x264_mb_mc_8x8( h, i8 );
 754
 755     if( h->mb.b_lossless )
 756     {
 757         int i4;
 758         for( i4 = i8*4; i4 < i8*4+4; i4++ )
 759         {
 760             h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
 761                                 h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
 762                                 h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
 763             nnz8x8 |= array_non_zero( h->dct.luma4x4[i4] );
 764         }
 765         for( ch = 0; ch < 2; ch++ )
 766         {
 767             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
 768             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
 769             h->zigzagf.sub_4x4( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec );
 770             h->dct.luma4x4[16+i8+ch*4][0] = 0;
 771         }
 772     }
 773     else
 774     {
 775         if( h->mb.b_transform_8x8 )
 776         {
 777             DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 778             h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
 779             h->quantf.quant_8x8( dct8x8, h->quant8_mf[CQM_8PY][i_qp], h->quant8_bias[CQM_8PY][i_qp] );
 780             h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
 781
 782             if( b_decimate )
 783                 nnz8x8 = 4 <= x264_mb_decimate_score( h->dct.luma8x8[i8], 64 );
 784             else
 785                 nnz8x8 = array_non_zero( dct8x8 );
 786
 787             if( nnz8x8 )
 788             {
 789                 h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
 790                 h->dctf.add8x8_idct8( p_fdec, dct8x8 );
 791             }
 792         }
 793         else
 794         {
 795             int i4;
 796             DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 797             h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
 798             for( i4 = 0; i4 < 4; i4++ )
 799                 h->quantf.quant_4x4( dct4x4[i4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 800             for( i4 = 0; i4 < 4; i4++ )
 801                 h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
 802
 803             if( b_decimate )
 804             {
 805                 int i_decimate_8x8 = 0;
 806                 for( i4 = 0; i4 < 4 && i_decimate_8x8 < 4; i4++ )
 807                     i_decimate_8x8 += x264_mb_decimate_score( h->dct.luma4x4[i8*4+i4], 16 );
 808                 nnz8x8 = 4 <= i_decimate_8x8;
 809             }
 810             else
 811                 nnz8x8 = array_non_zero( dct4x4 );
 812
 813             if( nnz8x8 )
 814             {
 815                 for( i4 = 0; i4 < 4; i4++ )
 816                     h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
 817                 h->dctf.add8x8_idct( p_fdec, dct4x4 );
 818             }
 819         }
 820
 821         i_qp = h->mb.i_chroma_qp;
 822
 823         for( ch = 0; ch < 2; ch++ )
 824         {
 825             DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
 826             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
 827             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
 828
 829             h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
 830             h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 831             h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
 832             h->dct.luma4x4[16+i8+ch*4][0] = 0;
 833             if( array_non_zero( dct4x4 ) )
 834             {
 835                 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
 836                 h->dctf.add4x4_idct( p_fdec, dct4x4 );
 837             }
 838         }
 839     }
 840     h->mb.i_cbp_luma &= ~(1 << i8);
 841     h->mb.i_cbp_luma |= nnz8x8 << i8;
 842     h->mb.i_cbp_chroma = 0x02;
 843 }