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]] = h->dct.luma4x4[i][0];
 147         }
 148         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct4x4[0] );
 149         return;
 150     }
 151
 152     h->dctf.sub16x16_dct( &dct4x4[1], p_src, p_dst );
 153     for( i = 0; i < 16; i++ )
 154     {
 155         /* copy dc coeff */
 156         dct4x4[0][block_idx_y[i]][block_idx_x[i]] = dct4x4[1+i][0][0];
 157         dct4x4[1+i][0][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] = h->dct.luma4x4[16+i+ch*4][0];
 209             }
 210             continue;
 211         }
 212
 213         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 214         /* calculate dct coeffs */
 215         for( i = 0; i < 4; i++ )
 216         {
 217             /* copy dc coeff */
 218             dct2x2[block_idx_y[i]][block_idx_x[i]] = dct4x4[i][0][0];
 219             dct4x4[i][0][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( h->dct.luma4x4[16+i] );
 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     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 299     if( !h->mb.b_skip_pbskip_mc )
 300     {
 301         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 302                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 303                        mvx, mvy, 16, 16 );
 304
 305         h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 306                          h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 307                          mvx, mvy, 8, 8 );
 308
 309         h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 310                          h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 311                          mvx, mvy, 8, 8 );
 312     }
 313
 314     x264_macroblock_encode_skip( h );
 315 }
 316
 317 /*****************************************************************************
 318  * x264_macroblock_encode:
 319  *****************************************************************************/
 320 void x264_macroblock_encode( x264_t *h )
 321 {
 322     int i_cbp_dc = 0;
 323     int i_qp = h->mb.i_qp;
 324     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 325     int b_force_no_skip = 0;
 326     int i,j,idx;
 327     uint8_t nnz8x8[4] = {1,1,1,1};
 328
 329     if( h->sh.b_mbaff
 330         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 331         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 332     {
 333         /* The first skip is predicted to be a frame mb pair.
 334          * We don't yet support the aff part of mbaff, so force it to non-skip
 335          * so that we can pick the aff flag. */
 336         b_force_no_skip = 1;
 337         if( IS_SKIP(h->mb.i_type) )
 338         {
 339             if( h->mb.i_type == P_SKIP )
 340                 h->mb.i_type = P_L0;
 341             else if( h->mb.i_type == B_SKIP )
 342                 h->mb.i_type = B_DIRECT;
 343         }
 344     }
 345
 346     if( h->mb.i_type == P_SKIP )
 347     {
 348         /* A bit special */
 349         x264_macroblock_encode_pskip( h );
 350         return;
 351     }
 352     if( h->mb.i_type == B_SKIP )
 353     {
 354         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 355         if( !h->mb.b_skip_pbskip_mc )
 356             x264_mb_mc( h );
 357         x264_macroblock_encode_skip( h );
 358         return;
 359     }
 360
 361     if( h->mb.i_type == I_16x16 )
 362     {
 363         const int i_mode = h->mb.i_intra16x16_pred_mode;
 364         h->mb.b_transform_8x8 = 0;
 365         /* do the right prediction */
 366         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 367
 368         /* encode the 16x16 macroblock */
 369         x264_mb_encode_i16x16( h, i_qp );
 370     }
 371     else if( h->mb.i_type == I_8x8 )
 372     {
 373         DECLARE_ALIGNED_16( uint8_t edge[33] );
 374         h->mb.b_transform_8x8 = 1;
 375         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 376         if( h->mb.i_skip_intra )
 377         {
 378             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 379             /* In RD mode, restore the now-overwritten DCT data. */
 380             if( h->mb.i_skip_intra == 2 )
 381                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 382         }
 383         for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 384         {
 385             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 386             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 387
 388             x264_predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 389             h->predict_8x8[i_mode]( p_dst, edge );
 390             x264_mb_encode_i8x8( h, i, i_qp );
 391         }
 392         for( i = 0; i < 4; i++ )
 393             nnz8x8[i] = array_non_zero( h->dct.luma8x8[i] );
 394     }
 395     else if( h->mb.i_type == I_4x4 )
 396     {
 397         h->mb.b_transform_8x8 = 0;
 398         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 399         if( h->mb.i_skip_intra )
 400         {
 401             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 402             /* In RD mode, restore the now-overwritten DCT data. */
 403             if( h->mb.i_skip_intra == 2 )
 404                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 405         }
 406         for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 407         {
 408             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][4 * block_idx_x[i] + 4 * block_idx_y[i] * FDEC_STRIDE];
 409             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 410
 411             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 412                 /* emulate missing topright samples */
 413                 *(uint32_t*) &p_dst[4-FDEC_STRIDE] = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 414
 415             h->predict_4x4[i_mode]( p_dst );
 416             x264_mb_encode_i4x4( h, i, i_qp );
 417         }
 418     }
 419     else    /* Inter MB */
 420     {
 421         int i8x8, i4x4;
 422         int i_decimate_mb = 0;
 423
 424         /* Motion compensation */
 425         x264_mb_mc( h );
 426
 427         if( h->mb.b_lossless )
 428         {
 429             for( i4x4 = 0; i4x4 < 16; i4x4++ )
 430             {
 431                 int x = 4*block_idx_x[i4x4];
 432                 int y = 4*block_idx_y[i4x4];
 433                 h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 434                                     h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
 435                                     h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
 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
 444             for( idx = 0; idx < 4; idx++ )
 445             {
 446                 if( h->mb.b_noise_reduction )
 447                     x264_denoise_dct( h, (int16_t*)dct8x8[idx] );
 448                 if( h->mb.b_trellis )
 449                     x264_quant_8x8_trellis( h, dct8x8[idx], CQM_8PY, i_qp, 0 );
 450                 else
 451                     h->quantf.quant_8x8( dct8x8[idx], h->quant8_mf[CQM_8PY][i_qp], h->quant8_bias[CQM_8PY][i_qp] );
 452
 453                 h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 454
 455                 if( b_decimate )
 456                 {
 457                     int i_decimate_8x8 = x264_mb_decimate_score( h->dct.luma8x8[idx], 64 );
 458                     i_decimate_mb += i_decimate_8x8;
 459                     if( i_decimate_8x8 < 4 )
 460                         nnz8x8[idx] = 0;
 461                 }
 462                 else
 463                     nnz8x8[idx] = array_non_zero( dct8x8[idx] );
 464             }
 465
 466             if( i_decimate_mb < 6 && b_decimate )
 467                 *(uint32_t*)nnz8x8 = 0;
 468             else
 469             {
 470                 for( idx = 0; idx < 4; idx++ )
 471                     if( nnz8x8[idx] )
 472                     {
 473                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 474                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 475                     }
 476             }
 477         }
 478         else
 479         {
 480             DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 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                     nnz8x8[i8x8] = 0;
 510             }
 511
 512             if( i_decimate_mb < 6 && b_decimate )
 513                 *(uint32_t*)nnz8x8 = 0;
 514             else
 515             {
 516                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 517                     if( nnz8x8[i8x8] )
 518                     {
 519                         for( i = 0; i < 4; i++ )
 520                             h->quantf.dequant_4x4( dct4x4[i8x8*4+i], h->dequant4_mf[CQM_4PY], i_qp );
 521                         h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 522                     }
 523             }
 524         }
 525     }
 526
 527     /* encode chroma */
 528     if( IS_INTRA( h->mb.i_type ) )
 529     {
 530         const int i_mode = h->mb.i_chroma_pred_mode;
 531         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 532         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 533     }
 534
 535     /* encode the 8x8 blocks */
 536     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 537
 538     /* coded block pattern and non_zero_count */
 539     h->mb.i_cbp_luma = 0x00;
 540     if( h->mb.i_type == I_16x16 )
 541     {
 542         for( i = 0; i < 16; i++ )
 543         {
 544             int nz = array_non_zero( h->dct.luma4x4[i] );
 545             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 546             h->mb.i_cbp_luma |= nz;
 547         }
 548         h->mb.i_cbp_luma *= 0xf;
 549     }
 550     else
 551     {
 552         for( i = 0; i < 4; i++)
 553         {
 554             if(!nnz8x8[i])
 555                 for( j = 0; j < 4; j++ )
 556                     h->mb.cache.non_zero_count[x264_scan8[j+i*4]] = 0;
 557             else if( h->mb.b_transform_8x8 )
 558             {
 559                 int nz = nnz8x8[i];
 560                 for( j = 0; j < 4; j++ )
 561                     h->mb.cache.non_zero_count[x264_scan8[j+4*i]] = nz;
 562                 h->mb.i_cbp_luma |= nz << i;
 563             }
 564             else
 565             {
 566                 for( j = 0; j < 4; j++ )
 567                 {
 568                     int nz = array_non_zero( h->dct.luma4x4[j+i*4] );
 569                     h->mb.cache.non_zero_count[x264_scan8[j+i*4]] = nz;
 570                     h->mb.i_cbp_luma |= nz << i;
 571                 }
 572             }
 573         }
 574     }
 575
 576     if( h->param.b_cabac )
 577     {
 578         i_cbp_dc = ( h->mb.i_type == I_16x16 && array_non_zero( h->dct.luma16x16_dc ) )
 579                  | array_non_zero( h->dct.chroma_dc[0] ) << 1
 580                  | array_non_zero( h->dct.chroma_dc[1] ) << 2;
 581     }
 582
 583     /* store cbp */
 584     h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
 585
 586     /* Check for P_SKIP
 587      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 588      *      (if multiple mv give same result)*/
 589     if( !b_force_no_skip )
 590     {
 591         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 592             h->mb.i_cbp_luma == 0x00 && h->mb.i_cbp_chroma == 0x00 &&
 593             h->mb.cache.mv[0][x264_scan8[0]][0] == h->mb.cache.pskip_mv[0] &&
 594             h->mb.cache.mv[0][x264_scan8[0]][1] == h->mb.cache.pskip_mv[1] &&
 595             h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 596         {
 597             h->mb.i_type = P_SKIP;
 598         }
 599
 600         /* Check for B_SKIP */
 601         if( h->mb.i_type == B_DIRECT &&
 602             h->mb.i_cbp_luma == 0x00 && h->mb.i_cbp_chroma== 0x00 )
 603         {
 604             h->mb.i_type = B_SKIP;
 605         }
 606     }
 607 }
 608
 609 /*****************************************************************************
 610  * x264_macroblock_probe_skip:
 611  *  Check if the current MB could be encoded as a [PB]_SKIP (it supposes you use
 612  *  the previous QP
 613  *****************************************************************************/
 614 int x264_macroblock_probe_skip( x264_t *h, const int b_bidir )
 615 {
 616     DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 617     DECLARE_ALIGNED_16( int16_t dct2x2[2][2] );
 618     DECLARE_ALIGNED_16( int16_t dctscan[16] );
 619
 620     int i_qp = h->mb.i_qp;
 621     int mvp[2];
 622     int ch;
 623
 624     int i8x8, i4x4;
 625     int i_decimate_mb;
 626
 627     if( !b_bidir )
 628     {
 629         /* Get the MV */
 630         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 631         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 632
 633         /* Motion compensation */
 634         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 635                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 636                        mvp[0], mvp[1], 16, 16 );
 637     }
 638
 639     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 640     {
 641         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
 642         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
 643         /* get luma diff */
 644         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
 645                                     h->mb.pic.p_fdec[0] + fdec_offset );
 646         /* encode one 4x4 block */
 647         for( i4x4 = 0; i4x4 < 4; i4x4++ )
 648         {
 649             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 650             if( !array_non_zero(dct4x4[i4x4]) )
 651                 continue;
 652             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 653             i_decimate_mb += x264_mb_decimate_score( dctscan, 16 );
 654             if( i_decimate_mb >= 6 )
 655                 return 0;
 656         }
 657     }
 658
 659     /* encode chroma */
 660     i_qp = h->mb.i_chroma_qp;
 661
 662     for( ch = 0; ch < 2; ch++ )
 663     {
 664         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 665         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 666
 667         if( !b_bidir )
 668         {
 669             h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 670                              h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 671                              mvp[0], mvp[1], 8, 8 );
 672         }
 673
 674         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 675
 676         /* calculate dct DC */
 677         dct2x2[0][0] = dct4x4[0][0][0];
 678         dct2x2[0][1] = dct4x4[1][0][0];
 679         dct2x2[1][0] = dct4x4[2][0][0];
 680         dct2x2[1][1] = dct4x4[3][0][0];
 681         h->dctf.dct2x2dc( dct2x2 );
 682         h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4PC][i_qp][0]>>1, h->quant4_bias[CQM_4PC][i_qp][0]<<1 );
 683         if( array_non_zero(dct2x2) )
 684             return 0;
 685
 686         /* calculate dct coeffs */
 687         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 688         {
 689             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 690             if( !array_non_zero(dct4x4[i4x4]) )
 691                 continue;
 692             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 693             i_decimate_mb += x264_mb_decimate_score( dctscan+1, 15 );
 694             if( i_decimate_mb >= 7 )
 695                 return 0;
 696         }
 697     }
 698
 699     h->mb.b_skip_pbskip_mc = 1;
 700     return 1;
 701 }
 702
 703 /****************************************************************************
 704  * DCT-domain noise reduction / adaptive deadzone
 705  * from libavcodec
 706  ****************************************************************************/
 707
 708 void x264_noise_reduction_update( x264_t *h )
 709 {
 710     int cat, i;
 711     for( cat = 0; cat < 2; cat++ )
 712     {
 713         int size = cat ? 64 : 16;
 714         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
 715
 716         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
 717         {
 718             for( i = 0; i < size; i++ )
 719                 h->nr_residual_sum[cat][i] >>= 1;
 720             h->nr_count[cat] >>= 1;
 721         }
 722
 723         for( i = 0; i < size; i++ )
 724             h->nr_offset[cat][i] =
 725                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
 726                  + h->nr_residual_sum[cat][i]/2)
 727               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
 728     }
 729 }
 730
 731 void x264_denoise_dct( x264_t *h, int16_t *dct )
 732 {
 733     const int cat = h->mb.b_transform_8x8;
 734     int i;
 735
 736     h->nr_count[cat]++;
 737
 738     for( i = (cat ? 63 : 15); i >= 1; i-- )
 739     {
 740         int level = dct[i];
 741         if( level )
 742         {
 743             if( level > 0 )
 744             {
 745                 h->nr_residual_sum[cat][i] += level;
 746                 level -= h->nr_offset[cat][i];
 747                 if( level < 0 )
 748                     level = 0;
 749             }
 750             else
 751             {
 752                 h->nr_residual_sum[cat][i] -= level;
 753                 level += h->nr_offset[cat][i];
 754                 if( level > 0 )
 755                     level = 0;
 756             }
 757             dct[i] = level;
 758         }
 759     }
 760 }
 761
 762 /*****************************************************************************
 763  * RD only; 4 calls to this do not make up for one macroblock_encode.
 764  * doesn't transform chroma dc.
 765  *****************************************************************************/
 766 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
 767 {
 768     int i_qp = h->mb.i_qp;
 769     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
 770     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
 771     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 772     int nnz8x8;
 773     int ch;
 774
 775     x264_mb_mc_8x8( h, i8 );
 776
 777     if( h->mb.b_transform_8x8 )
 778     {
 779         DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 780         h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
 781         h->quantf.quant_8x8( dct8x8, h->quant8_mf[CQM_8PY][i_qp], h->quant8_bias[CQM_8PY][i_qp] );
 782         h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
 783
 784         if( b_decimate )
 785             nnz8x8 = 4 <= x264_mb_decimate_score( h->dct.luma8x8[i8], 64 );
 786         else
 787             nnz8x8 = array_non_zero( dct8x8 );
 788
 789         if( nnz8x8 )
 790         {
 791             h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
 792             h->dctf.add8x8_idct8( p_fdec, dct8x8 );
 793         }
 794     }
 795     else
 796     {
 797         int i4;
 798         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 799         h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
 800         h->quantf.quant_4x4( dct4x4[0], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 801         h->quantf.quant_4x4( dct4x4[1], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 802         h->quantf.quant_4x4( dct4x4[2], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 803         h->quantf.quant_4x4( dct4x4[3], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 804         for( i4 = 0; i4 < 4; i4++ )
 805             h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
 806
 807         if( b_decimate )
 808         {
 809             int i_decimate_8x8 = 0;
 810             for( i4 = 0; i4 < 4 && i_decimate_8x8 < 4; i4++ )
 811                 i_decimate_8x8 += x264_mb_decimate_score( h->dct.luma4x4[i8*4+i4], 16 );
 812             nnz8x8 = 4 <= i_decimate_8x8;
 813         }
 814         else
 815             nnz8x8 = array_non_zero( dct4x4 );
 816
 817         if( nnz8x8 )
 818         {
 819             for( i4 = 0; i4 < 4; i4++ )
 820                 h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
 821             h->dctf.add8x8_idct( p_fdec, dct4x4 );
 822         }
 823     }
 824
 825     i_qp = h->mb.i_chroma_qp;
 826
 827     for( ch = 0; ch < 2; ch++ )
 828     {
 829         DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
 830         p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
 831         p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
 832
 833         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
 834         h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 835         h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
 836         if( array_non_zero( dct4x4 ) )
 837         {
 838             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
 839             h->dctf.add4x4_idct( p_fdec, dct4x4 );
 840         }
 841     }
 842
 843     h->mb.i_cbp_luma &= ~(1 << i8);
 844     h->mb.i_cbp_luma |= nnz8x8 << i8;
 845     h->mb.i_cbp_chroma = 0x02;
 846 }