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

   1 /*****************************************************************************
   2  * macroblock.c: h264 encoder library
   3  *****************************************************************************
   4  * Copyright (C) 2003 Laurent Aimar
   5  * $Id: macroblock.c,v 1.1 2004/06/03 19:27:08 fenrir Exp $
   6  *
   7  * Authors: Laurent Aimar <fenrir@via.ecp.fr>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, write to the Free Software
  21  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
  22  *****************************************************************************/
  23
  24 #include "common/common.h"
  25 #include "macroblock.h"
  26
  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( abs( dct[idx--] ) > 1 )
  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     int x = 4 * block_idx_x[idx];
  85     int y = 4 * block_idx_y[idx];
  86     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
  87     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
  88     DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
  89
  90     if( h->mb.b_lossless )
  91     {
  92         h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
  93         return;
  94     }
  95
  96     h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
  97
  98     if( h->mb.b_trellis )
  99         x264_quant_4x4_trellis( h, dct4x4, CQM_4IY, i_qscale, DCT_LUMA_4x4, 1 );
 100     else
 101         h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4IY][i_qscale], h->quant4_bias[CQM_4IY][i_qscale] );
 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
 110 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qscale )
 111 {
 112     int x = 8 * (idx&1);
 113     int y = 8 * (idx>>1);
 114     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
 115     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
 116     DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 117
 118     h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
 119
 120     if( h->mb.b_trellis )
 121         x264_quant_8x8_trellis( h, dct8x8, CQM_8IY, i_qscale, 1 );
 122     else
 123         h->quantf.quant_8x8( dct8x8, h->quant8_mf[CQM_8IY][i_qscale], h->quant8_bias[CQM_8IY][i_qscale] );
 124
 125     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
 126     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qscale );
 127     h->dctf.add8x8_idct8( p_dst, dct8x8 );
 128 }
 129
 130 static void x264_mb_encode_i16x16( x264_t *h, int i_qscale )
 131 {
 132     uint8_t  *p_src = h->mb.pic.p_fenc[0];
 133     uint8_t  *p_dst = h->mb.pic.p_fdec[0];
 134
 135     DECLARE_ALIGNED_16( int16_t dct4x4[16+1][4][4] );
 136
 137     int i;
 138
 139     if( h->mb.b_lossless )
 140     {
 141         for( i = 0; i < 16; i++ )
 142         {
 143             int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 144             int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 145             h->zigzagf.sub_4x4( h->dct.luma4x4[i], p_src+oe, p_dst+od );
 146             dct4x4[0][block_idx_x[i]][block_idx_y[i]] = p_src[oe] - p_dst[od];
 147             p_dst[od] = p_src[oe];
 148         }
 149         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct4x4[0] );
 150         return;
 151     }
 152
 153     h->dctf.sub16x16_dct( &dct4x4[1], p_src, p_dst );
 154     for( i = 0; i < 16; i++ )
 155     {
 156         /* copy dc coeff */
 157         dct4x4[0][block_idx_y[i]][block_idx_x[i]] = dct4x4[1+i][0][0];
 158
 159         /* quant/scan/dequant */
 160         if( h->mb.b_trellis )
 161             x264_quant_4x4_trellis( h, dct4x4[1+i], CQM_4IY, i_qscale, DCT_LUMA_AC, 1 );
 162         else
 163             h->quantf.quant_4x4( dct4x4[1+i], h->quant4_mf[CQM_4IY][i_qscale], h->quant4_bias[CQM_4IY][i_qscale] );
 164
 165         h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[1+i] );
 166         h->quantf.dequant_4x4( dct4x4[1+i], h->dequant4_mf[CQM_4IY], i_qscale );
 167     }
 168
 169     h->dctf.dct4x4dc( dct4x4[0] );
 170     h->quantf.quant_4x4_dc( dct4x4[0], h->quant4_mf[CQM_4IY][i_qscale][0]>>1, h->quant4_bias[CQM_4IY][i_qscale][0]<<1 );
 171     h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct4x4[0] );
 172
 173     /* output samples to fdec */
 174     h->dctf.idct4x4dc( dct4x4[0] );
 175     x264_mb_dequant_4x4_dc( dct4x4[0], h->dequant4_mf[CQM_4IY], i_qscale );  /* XXX not inversed */
 176
 177     /* calculate dct coeffs */
 178     for( i = 0; i < 16; i++ )
 179     {
 180         /* copy dc coeff */
 181         dct4x4[1+i][0][0] = dct4x4[0][block_idx_y[i]][block_idx_x[i]];
 182     }
 183     /* put pixels to fdec */
 184     h->dctf.add16x16_idct( p_dst, &dct4x4[1] );
 185 }
 186
 187 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qscale )
 188 {
 189     int i, ch;
 190     int b_decimate = b_inter && (h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate);
 191
 192     for( ch = 0; ch < 2; ch++ )
 193     {
 194         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 195         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 196         int i_decimate_score = 0;
 197
 198         DECLARE_ALIGNED_16( int16_t dct2x2[2][2]  );
 199         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 200
 201         if( h->mb.b_lossless )
 202         {
 203             for( i = 0; i < 4; i++ )
 204             {
 205                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 206                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 207                 h->zigzagf.sub_4x4( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od );
 208                 h->dct.chroma_dc[ch][i] = p_src[oe] - p_dst[od];
 209                 p_dst[od] = p_src[oe];
 210             }
 211             continue;
 212         }
 213
 214         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 215         /* calculate dct coeffs */
 216         for( i = 0; i < 4; i++ )
 217         {
 218             /* copy dc coeff */
 219             dct2x2[block_idx_y[i]][block_idx_x[i]] = dct4x4[i][0][0];
 220
 221             /* no trellis; it doesn't seem to help chroma noticeably */
 222             h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[CQM_4IC+b_inter][i_qscale], h->quant4_bias[CQM_4IC+b_inter][i_qscale] );
 223             h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 224
 225             if( b_decimate )
 226             {
 227                 i_decimate_score += x264_mb_decimate_score( h->dct.luma4x4[16+i+ch*4]+1, 15 );
 228             }
 229         }
 230
 231         h->dctf.dct2x2dc( dct2x2 );
 232         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 );
 233         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 234
 235         /* output samples to fdec */
 236         h->dctf.idct2x2dc( dct2x2 );
 237         x264_mb_dequant_2x2_dc( dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qscale );  /* XXX not inversed */
 238
 239         if( b_decimate && i_decimate_score < 7 )
 240         {
 241             /* Near null chroma 8x8 block so make it null (bits saving) */
 242             memset( &h->dct.luma4x4[16+ch*4], 0, 4 * sizeof( *h->dct.luma4x4 ) );
 243             if( !array_non_zero( dct2x2 ) )
 244                 continue;
 245             memset( dct4x4, 0, sizeof( dct4x4 ) );
 246         }
 247         else
 248         {
 249             for( i = 0; i < 4; i++ )
 250                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qscale );
 251         }
 252
 253         for( i = 0; i < 4; i++ )
 254             dct4x4[i][0][0] = dct2x2[0][i];
 255         h->dctf.add8x8_idct( p_dst, dct4x4 );
 256     }
 257
 258     /* coded block pattern */
 259     h->mb.i_cbp_chroma = 0;
 260     for( i = 0; i < 8; i++ )
 261     {
 262         int nz = array_non_zero_count( h->dct.luma4x4[16+i]+1, 15 );
 263         h->mb.cache.non_zero_count[x264_scan8[16+i]] = nz;
 264         h->mb.i_cbp_chroma |= nz;
 265     }
 266     if( h->mb.i_cbp_chroma )
 267         h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
 268     else if( array_non_zero( h->dct.chroma_dc ) )
 269         h->mb.i_cbp_chroma = 1;    /* dc only */
 270 }
 271
 272 static void x264_macroblock_encode_skip( x264_t *h )
 273 {
 274     int i;
 275     h->mb.i_cbp_luma = 0x00;
 276     h->mb.i_cbp_chroma = 0x00;
 277
 278     for( i = 0; i < 16+8; i++ )
 279     {
 280         h->mb.cache.non_zero_count[x264_scan8[i]] = 0;
 281     }
 282
 283     /* store cbp */
 284     h->mb.cbp[h->mb.i_mb_xy] = 0;
 285 }
 286
 287 /*****************************************************************************
 288  * x264_macroblock_encode_pskip:
 289  *  Encode an already marked skip block
 290  *****************************************************************************/
 291 void x264_macroblock_encode_pskip( x264_t *h )
 292 {
 293     const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 294                                 h->mb.mv_min[0], h->mb.mv_max[0] );
 295     const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 296                                 h->mb.mv_min[1], h->mb.mv_max[1] );
 297
 298     /* Motion compensation XXX probably unneeded */
 299     h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 300                    h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 301                    mvx, mvy, 16, 16 );
 302
 303     /* Chroma MC */
 304     h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 305                      h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 306                      mvx, mvy, 8, 8 );
 307
 308     h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 309                      h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 310                      mvx, mvy, 8, 8 );
 311
 312     x264_macroblock_encode_skip( h );
 313 }
 314
 315 /*****************************************************************************
 316  * x264_macroblock_encode:
 317  *****************************************************************************/
 318 void x264_macroblock_encode( x264_t *h )
 319 {
 320     int i_cbp_dc = 0;
 321     int i_qp = h->mb.i_qp;
 322     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 323     int b_force_no_skip = 0;
 324     int i;
 325
 326     if( h->sh.b_mbaff
 327         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 328         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 329     {
 330         /* The first skip is predicted to be a frame mb pair.
 331          * We don't yet support the aff part of mbaff, so force it to non-skip
 332          * so that we can pick the aff flag. */
 333         b_force_no_skip = 1;
 334         if( IS_SKIP(h->mb.i_type) )
 335         {
 336             if( h->mb.i_type == P_SKIP )
 337                 h->mb.i_type = P_L0;
 338             else if( h->mb.i_type == B_SKIP )
 339                 h->mb.i_type = B_DIRECT;
 340         }
 341     }
 342
 343     if( h->mb.i_type == P_SKIP )
 344     {
 345         /* A bit special */
 346         x264_macroblock_encode_pskip( h );
 347         return;
 348     }
 349     if( h->mb.i_type == B_SKIP )
 350     {
 351         /* XXX motion compensation is probably unneeded */
 352         x264_mb_mc( h );
 353         x264_macroblock_encode_skip( h );
 354         return;
 355     }
 356
 357     if( h->mb.i_type == I_16x16 )
 358     {
 359         const int i_mode = h->mb.i_intra16x16_pred_mode;
 360         h->mb.b_transform_8x8 = 0;
 361         /* do the right prediction */
 362         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 363
 364         /* encode the 16x16 macroblock */
 365         x264_mb_encode_i16x16( h, i_qp );
 366     }
 367     else if( h->mb.i_type == I_8x8 )
 368     {
 369         DECLARE_ALIGNED_16( uint8_t edge[33] );
 370         h->mb.b_transform_8x8 = 1;
 371         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 372         if( h->mb.i_skip_intra )
 373         {
 374             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 375             /* In RD mode, restore the now-overwritten DCT data. */
 376             if( h->mb.i_skip_intra == 2 )
 377                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 378         }
 379         for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 380         {
 381             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 382             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 383
 384             x264_predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 385             h->predict_8x8[i_mode]( p_dst, edge );
 386             x264_mb_encode_i8x8( h, i, i_qp );
 387         }
 388     }
 389     else if( h->mb.i_type == I_4x4 )
 390     {
 391         h->mb.b_transform_8x8 = 0;
 392         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 393         if( h->mb.i_skip_intra )
 394         {
 395             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 396             /* In RD mode, restore the now-overwritten DCT data. */
 397             if( h->mb.i_skip_intra == 2 )
 398                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 399         }
 400         for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 401         {
 402             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][4 * block_idx_x[i] + 4 * block_idx_y[i] * FDEC_STRIDE];
 403             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 404
 405             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 406                 /* emulate missing topright samples */
 407                 *(uint32_t*) &p_dst[4-FDEC_STRIDE] = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 408
 409             h->predict_4x4[i_mode]( p_dst );
 410             x264_mb_encode_i4x4( h, i, i_qp );
 411         }
 412     }
 413     else    /* Inter MB */
 414     {
 415         int i8x8, i4x4, idx;
 416         int i_decimate_mb = 0;
 417
 418         /* Motion compensation */
 419         x264_mb_mc( h );
 420
 421         if( h->mb.b_lossless )
 422         {
 423             for( i4x4 = 0; i4x4 < 16; i4x4++ )
 424             {
 425                 int x = 4*block_idx_x[i4x4];
 426                 int y = 4*block_idx_y[i4x4];
 427                 h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 428                                     h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
 429                                     h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
 430             }
 431         }
 432         else if( h->mb.b_transform_8x8 )
 433         {
 434             DECLARE_ALIGNED_16( int16_t dct8x8[4][8][8] );
 435             int nnz8x8[4] = {1,1,1,1};
 436             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 437             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 438
 439             for( idx = 0; idx < 4; idx++ )
 440             {
 441                 if( h->mb.b_noise_reduction )
 442                     x264_denoise_dct( h, (int16_t*)dct8x8[idx] );
 443                 if( h->mb.b_trellis )
 444                     x264_quant_8x8_trellis( h, dct8x8[idx], CQM_8PY, i_qp, 0 );
 445                 else
 446                     h->quantf.quant_8x8( dct8x8[idx], h->quant8_mf[CQM_8PY][i_qp], h->quant8_bias[CQM_8PY][i_qp] );
 447
 448                 h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 449
 450                 if( b_decimate )
 451                 {
 452                     int i_decimate_8x8 = x264_mb_decimate_score( h->dct.luma8x8[idx], 64 );
 453                     i_decimate_mb += i_decimate_8x8;
 454                     if( i_decimate_8x8 < 4 )
 455                     {
 456                         memset( h->dct.luma8x8[idx], 0, sizeof( h->dct.luma8x8[idx] ) );
 457                         memset( dct8x8[idx], 0, sizeof( dct8x8[idx] ) );
 458                         nnz8x8[idx] = 0;
 459                     }
 460                 }
 461                 else
 462                     nnz8x8[idx] = array_non_zero( dct8x8[idx] );
 463             }
 464
 465             if( i_decimate_mb < 6 && b_decimate )
 466                 memset( h->dct.luma8x8, 0, sizeof( h->dct.luma8x8 ) );
 467             else
 468             {
 469                 for( idx = 0; idx < 4; idx++ )
 470                     if( nnz8x8[idx] )
 471                     {
 472                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 473                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 474                     }
 475             }
 476         }
 477         else
 478         {
 479             DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 480             int nnz8x8[4] = {1,1,1,1};
 481             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 482
 483             for( i8x8 = 0; i8x8 < 4; i8x8++ )
 484             {
 485                 int i_decimate_8x8;
 486
 487                 /* encode one 4x4 block */
 488                 i_decimate_8x8 = 0;
 489                 for( i4x4 = 0; i4x4 < 4; i4x4++ )
 490                 {
 491                     idx = i8x8 * 4 + i4x4;
 492
 493                     if( h->mb.b_noise_reduction )
 494                         x264_denoise_dct( h, (int16_t*)dct4x4[idx] );
 495                     if( h->mb.b_trellis )
 496                         x264_quant_4x4_trellis( h, dct4x4[idx], CQM_4PY, i_qp, DCT_LUMA_4x4, 0 );
 497                     else
 498                         h->quantf.quant_4x4( dct4x4[idx], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 499
 500                     h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 501
 502                     if( b_decimate )
 503                         i_decimate_8x8 += x264_mb_decimate_score( h->dct.luma4x4[idx], 16 );
 504                 }
 505
 506                 /* decimate this 8x8 block */
 507                 i_decimate_mb += i_decimate_8x8;
 508                 if( i_decimate_8x8 < 4 && b_decimate )
 509                 {
 510                     memset( &dct4x4[i8x8*4], 0, 4 * sizeof( *dct4x4 ) );
 511                     memset( &h->dct.luma4x4[i8x8*4], 0, 4 * sizeof( *h->dct.luma4x4 ) );
 512                     nnz8x8[i8x8] = 0;
 513                 }
 514             }
 515
 516             if( i_decimate_mb < 6 && b_decimate )
 517                 memset( h->dct.luma4x4, 0, 16 * sizeof( *h->dct.luma4x4 ) );
 518             else
 519             {
 520                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 521                     if( nnz8x8[i8x8] )
 522                     {
 523                         for( i = 0; i < 4; i++ )
 524                             h->quantf.dequant_4x4( dct4x4[i8x8*4+i], h->dequant4_mf[CQM_4PY], i_qp );
 525                         h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 526                     }
 527             }
 528         }
 529     }
 530
 531     /* encode chroma */
 532     if( IS_INTRA( h->mb.i_type ) )
 533     {
 534         const int i_mode = h->mb.i_chroma_pred_mode;
 535         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 536         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 537     }
 538
 539     /* encode the 8x8 blocks */
 540     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 541
 542     /* coded block pattern and non_zero_count */
 543     h->mb.i_cbp_luma = 0x00;
 544     if( h->mb.i_type == I_16x16 )
 545     {
 546         for( i = 0; i < 16; i++ )
 547         {
 548             const int nz = array_non_zero_count( h->dct.luma4x4[i]+1, 15 );
 549             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 550             if( nz > 0 )
 551                 h->mb.i_cbp_luma = 0x0f;
 552         }
 553     }
 554     else if( h->mb.b_transform_8x8 )
 555     {
 556         /* coded_block_flag is enough for CABAC.
 557          * the full non_zero_count is done only in CAVLC. */
 558         for( i = 0; i < 4; i++ )
 559         {
 560             const int nz = array_non_zero( h->dct.luma8x8[i] );
 561             int j;
 562             for( j = 0; j < 4; j++ )
 563                 h->mb.cache.non_zero_count[x264_scan8[4*i+j]] = nz;
 564             if( nz > 0 )
 565                 h->mb.i_cbp_luma |= 1 << i;
 566         }
 567     }
 568     else
 569     {
 570         for( i = 0; i < 16; i++ )
 571         {
 572             const int nz = array_non_zero_count( h->dct.luma4x4[i], 16 );
 573             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 574             if( nz > 0 )
 575                 h->mb.i_cbp_luma |= 1 << (i/4);
 576         }
 577     }
 578
 579     if( h->param.b_cabac )
 580     {
 581         i_cbp_dc = ( h->mb.i_type == I_16x16 && array_non_zero( h->dct.luma16x16_dc ) )
 582                  | array_non_zero( h->dct.chroma_dc[0] ) << 1
 583                  | array_non_zero( h->dct.chroma_dc[1] ) << 2;
 584     }
 585
 586     /* store cbp */
 587     h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
 588
 589     /* Check for P_SKIP
 590      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 591      *      (if multiple mv give same result)*/
 592     if( !b_force_no_skip )
 593     {
 594         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 595             h->mb.i_cbp_luma == 0x00 && h->mb.i_cbp_chroma == 0x00 &&
 596             h->mb.cache.mv[0][x264_scan8[0]][0] == h->mb.cache.pskip_mv[0] &&
 597             h->mb.cache.mv[0][x264_scan8[0]][1] == h->mb.cache.pskip_mv[1] &&
 598             h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 599         {
 600             h->mb.i_type = P_SKIP;
 601         }
 602
 603         /* Check for B_SKIP */
 604         if( h->mb.i_type == B_DIRECT &&
 605             h->mb.i_cbp_luma == 0x00 && h->mb.i_cbp_chroma== 0x00 )
 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[16][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;
 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     /* get luma diff */
 643     h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0],
 644                                   h->mb.pic.p_fdec[0] );
 645
 646     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 647     {
 648         /* encode one 4x4 block */
 649         for( i4x4 = 0; i4x4 < 4; i4x4++ )
 650         {
 651             const int idx = i8x8 * 4 + i4x4;
 652
 653             h->quantf.quant_4x4( dct4x4[idx], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 654             h->zigzagf.scan_4x4( dctscan, dct4x4[idx] );
 655
 656             i_decimate_mb += x264_mb_decimate_score( dctscan, 16 );
 657
 658             if( i_decimate_mb >= 6 )
 659             {
 660                 /* not as P_SKIP */
 661                 return 0;
 662             }
 663         }
 664     }
 665
 666     /* encode chroma */
 667     i_qp = h->mb.i_chroma_qp;
 668
 669     for( ch = 0; ch < 2; ch++ )
 670     {
 671         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 672         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 673
 674         if( !b_bidir )
 675         {
 676             h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 677                              h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 678                              mvp[0], mvp[1], 8, 8 );
 679         }
 680
 681         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 682
 683         /* calculate dct DC */
 684         dct2x2[0][0] = dct4x4[0][0][0];
 685         dct2x2[0][1] = dct4x4[1][0][0];
 686         dct2x2[1][0] = dct4x4[2][0][0];
 687         dct2x2[1][1] = dct4x4[3][0][0];
 688         h->dctf.dct2x2dc( dct2x2 );
 689         h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4PC][i_qp][0]>>1, h->quant4_bias[CQM_4PC][i_qp][0]<<1 );
 690         if( dct2x2[0][0] || dct2x2[0][1] || dct2x2[1][0] || dct2x2[1][1]  )
 691         {
 692             /* can't be */
 693             return 0;
 694         }
 695
 696         /* calculate dct coeffs */
 697         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 698         {
 699             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 700             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 701
 702             i_decimate_mb += x264_mb_decimate_score( dctscan+1, 15 );
 703             if( i_decimate_mb >= 7 )
 704             {
 705                 return 0;
 706             }
 707         }
 708     }
 709
 710     return 1;
 711 }
 712
 713 /****************************************************************************
 714  * DCT-domain noise reduction / adaptive deadzone
 715  * from libavcodec
 716  ****************************************************************************/
 717
 718 void x264_noise_reduction_update( x264_t *h )
 719 {
 720     int cat, i;
 721     for( cat = 0; cat < 2; cat++ )
 722     {
 723         int size = cat ? 64 : 16;
 724         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
 725
 726         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
 727         {
 728             for( i = 0; i < size; i++ )
 729                 h->nr_residual_sum[cat][i] >>= 1;
 730             h->nr_count[cat] >>= 1;
 731         }
 732
 733         for( i = 0; i < size; i++ )
 734             h->nr_offset[cat][i] =
 735                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
 736                  + h->nr_residual_sum[cat][i]/2)
 737               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
 738     }
 739 }
 740
 741 void x264_denoise_dct( x264_t *h, int16_t *dct )
 742 {
 743     const int cat = h->mb.b_transform_8x8;
 744     int i;
 745
 746     h->nr_count[cat]++;
 747
 748     for( i = (cat ? 63 : 15); i >= 1; i-- )
 749     {
 750         int level = dct[i];
 751         if( level )
 752         {
 753             if( level > 0 )
 754             {
 755                 h->nr_residual_sum[cat][i] += level;
 756                 level -= h->nr_offset[cat][i];
 757                 if( level < 0 )
 758                     level = 0;
 759             }
 760             else
 761             {
 762                 h->nr_residual_sum[cat][i] -= level;
 763                 level += h->nr_offset[cat][i];
 764                 if( level > 0 )
 765                     level = 0;
 766             }
 767             dct[i] = level;
 768         }
 769     }
 770 }
 771
 772 /*****************************************************************************
 773  * RD only; 4 calls to this do not make up for one macroblock_encode.
 774  * doesn't transform chroma dc.
 775  *****************************************************************************/
 776 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
 777 {
 778     int i_qp = h->mb.i_qp;
 779     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
 780     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
 781     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 782     int nnz8x8;
 783     int ch;
 784
 785     x264_mb_mc_8x8( h, i8 );
 786
 787     if( h->mb.b_transform_8x8 )
 788     {
 789         DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 790         h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
 791         h->quantf.quant_8x8( dct8x8, h->quant8_mf[CQM_8PY][i_qp], h->quant8_bias[CQM_8PY][i_qp] );
 792         h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
 793
 794         if( b_decimate )
 795             nnz8x8 = 4 <= x264_mb_decimate_score( h->dct.luma8x8[i8], 64 );
 796         else
 797             nnz8x8 = array_non_zero( dct8x8 );
 798
 799         if( nnz8x8 )
 800         {
 801             h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
 802             h->dctf.add8x8_idct8( p_fdec, dct8x8 );
 803         }
 804     }
 805     else
 806     {
 807         int i4;
 808         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 809         h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
 810         h->quantf.quant_4x4( dct4x4[0], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 811         h->quantf.quant_4x4( dct4x4[1], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 812         h->quantf.quant_4x4( dct4x4[2], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 813         h->quantf.quant_4x4( dct4x4[3], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 814         for( i4 = 0; i4 < 4; i4++ )
 815             h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
 816
 817         if( b_decimate )
 818         {
 819             int i_decimate_8x8 = 0;
 820             for( i4 = 0; i4 < 4 && i_decimate_8x8 < 4; i4++ )
 821                 i_decimate_8x8 += x264_mb_decimate_score( h->dct.luma4x4[i8*4+i4], 16 );
 822             nnz8x8 = 4 <= i_decimate_8x8;
 823         }
 824         else
 825             nnz8x8 = array_non_zero( dct4x4 );
 826
 827         if( nnz8x8 )
 828         {
 829             for( i4 = 0; i4 < 4; i4++ )
 830                 h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
 831             h->dctf.add8x8_idct( p_fdec, dct4x4 );
 832         }
 833     }
 834
 835     i_qp = h->mb.i_chroma_qp;
 836
 837     for( ch = 0; ch < 2; ch++ )
 838     {
 839         DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
 840         p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
 841         p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
 842
 843         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
 844         h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 845         h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
 846         if( array_non_zero( dct4x4 ) )
 847         {
 848             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
 849             h->dctf.add4x4_idct( p_fdec, dct4x4 );
 850         }
 851     }
 852
 853     if( nnz8x8 )
 854         h->mb.i_cbp_luma |= (1 << i8);
 855     else
 856         h->mb.i_cbp_luma &= ~(1 << i8);
 857     h->mb.i_cbp_chroma = 0x02;
 858 }