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., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
  23  *****************************************************************************/
  24
  25 #include "common/common.h"
  26 #include "macroblock.h"
  27
  28
  29 #define ZIG(i,y,x) level[i] = dct[x][y];
  30 static inline void zigzag_scan_2x2_dc( int16_t level[4], int16_t dct[2][2] )
  31 {
  32     ZIG(0,0,0)
  33     ZIG(1,0,1)
  34     ZIG(2,1,0)
  35     ZIG(3,1,1)
  36 }
  37 #undef ZIG
  38
  39 /* (ref: JVT-B118)
  40  * x264_mb_decimate_score: given dct coeffs it returns a score to see if we could empty this dct coeffs
  41  * to 0 (low score means set it to null)
  42  * Used in inter macroblock (luma and chroma)
  43  *  luma: for a 8x8 block: if score < 4 -> null
  44  *        for the complete mb: if score < 6 -> null
  45  *  chroma: for the complete mb: if score < 7 -> null
  46  */
  47 static int x264_mb_decimate_score( int16_t *dct, int i_max )
  48 {
  49     static const int i_ds_table4[16] = {
  50         3,2,2,1,1,1,0,0,0,0,0,0,0,0,0,0 };
  51     static const int i_ds_table8[64] = {
  52         3,3,3,3,2,2,2,2,2,2,2,2,1,1,1,1,
  53         1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,
  54         0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
  55         0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 };
  56
  57     const int *ds_table = (i_max == 64) ? i_ds_table8 : i_ds_table4;
  58     int i_score = 0;
  59     int idx = i_max - 1;
  60
  61     while( idx >= 0 && dct[idx] == 0 )
  62         idx--;
  63
  64     while( idx >= 0 )
  65     {
  66         int i_run;
  67
  68         if( (unsigned)(dct[idx--] + 1) > 2 )
  69             return 9;
  70
  71         i_run = 0;
  72         while( idx >= 0 && dct[idx] == 0 )
  73         {
  74             idx--;
  75             i_run++;
  76         }
  77         i_score += ds_table[i_run];
  78     }
  79
  80     return i_score;
  81 }
  82
  83 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qscale )
  84 {
  85     int x = 4 * block_idx_x[idx];
  86     int y = 4 * block_idx_y[idx];
  87     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
  88     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
  89     DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
  90
  91     if( h->mb.b_lossless )
  92     {
  93         h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
  94         return;
  95     }
  96
  97     h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
  98
  99     if( h->mb.b_trellis )
 100         x264_quant_4x4_trellis( h, dct4x4, CQM_4IY, i_qscale, DCT_LUMA_4x4, 1 );
 101     else
 102         h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4IY][i_qscale], h->quant4_bias[CQM_4IY][i_qscale] );
 103
 104     if( array_non_zero( dct4x4 ) )
 105     {
 106         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4 );
 107         h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qscale );
 108
 109         /* output samples to fdec */
 110         h->dctf.add4x4_idct( p_dst, dct4x4 );
 111     }
 112     else
 113         memset( h->dct.luma4x4[idx], 0, sizeof(h->dct.luma4x4[idx]));
 114 }
 115
 116 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qscale )
 117 {
 118     int x = 8 * (idx&1);
 119     int y = 8 * (idx>>1);
 120     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
 121     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
 122     DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 123
 124     h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
 125
 126     if( h->mb.b_trellis )
 127         x264_quant_8x8_trellis( h, dct8x8, CQM_8IY, i_qscale, 1 );
 128     else
 129         h->quantf.quant_8x8( dct8x8, h->quant8_mf[CQM_8IY][i_qscale], h->quant8_bias[CQM_8IY][i_qscale] );
 130
 131     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
 132     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qscale );
 133     h->dctf.add8x8_idct8( p_dst, dct8x8 );
 134 }
 135
 136 static void x264_mb_encode_i16x16( x264_t *h, int i_qscale )
 137 {
 138     uint8_t  *p_src = h->mb.pic.p_fenc[0];
 139     uint8_t  *p_dst = h->mb.pic.p_fdec[0];
 140
 141     DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 142     DECLARE_ALIGNED_16( int16_t dct_dc4x4[4][4] );
 143
 144     int i;
 145
 146     if( h->mb.b_lossless )
 147     {
 148         for( i = 0; i < 16; i++ )
 149         {
 150             int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 151             int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 152             h->zigzagf.sub_4x4( h->dct.luma4x4[i], p_src+oe, p_dst+od );
 153             dct_dc4x4[block_idx_x[i]][block_idx_y[i]] = h->dct.luma4x4[i][0];
 154             h->dct.luma4x4[i][0] = 0;
 155         }
 156         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 157         return;
 158     }
 159
 160     h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
 161     for( i = 0; i < 16; i++ )
 162     {
 163         /* copy dc coeff */
 164         dct_dc4x4[block_idx_y[i]][block_idx_x[i]] = dct4x4[i][0][0];
 165         dct4x4[i][0][0] = 0;
 166
 167         /* quant/scan/dequant */
 168         if( h->mb.b_trellis )
 169             x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IY, i_qscale, DCT_LUMA_AC, 1 );
 170         else
 171             h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[CQM_4IY][i_qscale], h->quant4_bias[CQM_4IY][i_qscale] );
 172
 173         h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[i] );
 174         h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IY], i_qscale );
 175     }
 176
 177     h->dctf.dct4x4dc( dct_dc4x4 );
 178     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 );
 179     h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 180
 181     /* output samples to fdec */
 182     h->dctf.idct4x4dc( dct_dc4x4 );
 183     x264_mb_dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[CQM_4IY], i_qscale );  /* XXX not inversed */
 184
 185     /* calculate dct coeffs */
 186     for( i = 0; i < 16; i++ )
 187     {
 188         /* copy dc coeff */
 189         dct4x4[i][0][0] = dct_dc4x4[block_idx_y[i]][block_idx_x[i]];
 190     }
 191     /* put pixels to fdec */
 192     h->dctf.add16x16_idct( p_dst, dct4x4 );
 193 }
 194
 195 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qscale )
 196 {
 197     int i, ch;
 198     int b_decimate = b_inter && (h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate);
 199
 200     for( ch = 0; ch < 2; ch++ )
 201     {
 202         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 203         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 204         int i_decimate_score = 0;
 205
 206         DECLARE_ALIGNED_16( int16_t dct2x2[2][2]  );
 207         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 208
 209         if( h->mb.b_lossless )
 210         {
 211             for( i = 0; i < 4; i++ )
 212             {
 213                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 214                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 215                 h->zigzagf.sub_4x4( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od );
 216                 h->dct.chroma_dc[ch][i] = h->dct.luma4x4[16+i+ch*4][0];
 217                 h->dct.luma4x4[16+i+ch*4][0] = 0;
 218             }
 219             continue;
 220         }
 221
 222         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 223         /* calculate dct coeffs */
 224         for( i = 0; i < 4; i++ )
 225         {
 226             /* copy dc coeff */
 227             dct2x2[block_idx_y[i]][block_idx_x[i]] = dct4x4[i][0][0];
 228             dct4x4[i][0][0] = 0;
 229
 230             /* no trellis; it doesn't seem to help chroma noticeably */
 231             h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[CQM_4IC+b_inter][i_qscale], h->quant4_bias[CQM_4IC+b_inter][i_qscale] );
 232             h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 233
 234             if( b_decimate )
 235             {
 236                 i_decimate_score += x264_mb_decimate_score( h->dct.luma4x4[16+i+ch*4]+1, 15 );
 237             }
 238         }
 239
 240         h->dctf.dct2x2dc( dct2x2 );
 241         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 );
 242         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 243
 244         /* output samples to fdec */
 245         h->dctf.idct2x2dc( dct2x2 );
 246         x264_mb_dequant_2x2_dc( dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qscale );  /* XXX not inversed */
 247
 248         if( b_decimate && i_decimate_score < 7 )
 249         {
 250             /* Near null chroma 8x8 block so make it null (bits saving) */
 251             memset( &h->dct.luma4x4[16+ch*4], 0, 4 * sizeof( *h->dct.luma4x4 ) );
 252             if( !array_non_zero( dct2x2 ) )
 253                 continue;
 254             memset( dct4x4, 0, sizeof( dct4x4 ) );
 255         }
 256         else
 257         {
 258             for( i = 0; i < 4; i++ )
 259                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qscale );
 260         }
 261
 262         for( i = 0; i < 4; i++ )
 263             dct4x4[i][0][0] = dct2x2[0][i];
 264         h->dctf.add8x8_idct( p_dst, dct4x4 );
 265     }
 266
 267     /* coded block pattern */
 268     h->mb.i_cbp_chroma = 0;
 269     for( i = 0; i < 8; i++ )
 270     {
 271         int nz = array_non_zero( h->dct.luma4x4[16+i] );
 272         h->mb.cache.non_zero_count[x264_scan8[16+i]] = nz;
 273         h->mb.i_cbp_chroma |= nz;
 274     }
 275     if( h->mb.i_cbp_chroma )
 276         h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
 277     else if( array_non_zero( h->dct.chroma_dc ) )
 278         h->mb.i_cbp_chroma = 1;    /* dc only */
 279 }
 280
 281 static void x264_macroblock_encode_skip( x264_t *h )
 282 {
 283     h->mb.i_cbp_luma = 0x00;
 284     h->mb.i_cbp_chroma = 0x00;
 285     memset( h->mb.cache.non_zero_count, 0, X264_SCAN8_SIZE );
 286     /* store cbp */
 287     h->mb.cbp[h->mb.i_mb_xy] = 0;
 288 }
 289
 290 /*****************************************************************************
 291  * x264_macroblock_encode_pskip:
 292  *  Encode an already marked skip block
 293  *****************************************************************************/
 294 void x264_macroblock_encode_pskip( x264_t *h )
 295 {
 296     const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 297                                 h->mb.mv_min[0], h->mb.mv_max[0] );
 298     const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 299                                 h->mb.mv_min[1], h->mb.mv_max[1] );
 300
 301     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 302     if( !h->mb.b_skip_mc )
 303     {
 304         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 305                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 306                        mvx, mvy, 16, 16 );
 307
 308         h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 309                          h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 310                          mvx, mvy, 8, 8 );
 311
 312         h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 313                          h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 314                          mvx, mvy, 8, 8 );
 315     }
 316
 317     x264_macroblock_encode_skip( h );
 318 }
 319
 320 /*****************************************************************************
 321  * x264_macroblock_encode:
 322  *****************************************************************************/
 323 void x264_macroblock_encode( x264_t *h )
 324 {
 325     int i_cbp_dc = 0;
 326     int i_qp = h->mb.i_qp;
 327     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 328     int b_force_no_skip = 0;
 329     int i,j,idx;
 330     uint8_t nnz8x8[4] = {1,1,1,1};
 331
 332     if( h->sh.b_mbaff
 333         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 334         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 335     {
 336         /* The first skip is predicted to be a frame mb pair.
 337          * We don't yet support the aff part of mbaff, so force it to non-skip
 338          * so that we can pick the aff flag. */
 339         b_force_no_skip = 1;
 340         if( IS_SKIP(h->mb.i_type) )
 341         {
 342             if( h->mb.i_type == P_SKIP )
 343                 h->mb.i_type = P_L0;
 344             else if( h->mb.i_type == B_SKIP )
 345                 h->mb.i_type = B_DIRECT;
 346         }
 347     }
 348
 349     if( h->mb.i_type == P_SKIP )
 350     {
 351         /* A bit special */
 352         x264_macroblock_encode_pskip( h );
 353         return;
 354     }
 355     if( h->mb.i_type == B_SKIP )
 356     {
 357         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 358         if( !h->mb.b_skip_mc )
 359             x264_mb_mc( h );
 360         x264_macroblock_encode_skip( h );
 361         return;
 362     }
 363
 364     if( h->mb.i_type == I_16x16 )
 365     {
 366         const int i_mode = h->mb.i_intra16x16_pred_mode;
 367         h->mb.b_transform_8x8 = 0;
 368         /* do the right prediction */
 369         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 370
 371         /* encode the 16x16 macroblock */
 372         x264_mb_encode_i16x16( h, i_qp );
 373     }
 374     else if( h->mb.i_type == I_8x8 )
 375     {
 376         DECLARE_ALIGNED_16( uint8_t edge[33] );
 377         h->mb.b_transform_8x8 = 1;
 378         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 379         if( h->mb.i_skip_intra )
 380         {
 381             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 382             /* In RD mode, restore the now-overwritten DCT data. */
 383             if( h->mb.i_skip_intra == 2 )
 384                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 385         }
 386         for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 387         {
 388             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 389             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 390
 391             x264_predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 392             h->predict_8x8[i_mode]( p_dst, edge );
 393             x264_mb_encode_i8x8( h, i, i_qp );
 394         }
 395         for( i = 0; i < 4; i++ )
 396             nnz8x8[i] = array_non_zero( h->dct.luma8x8[i] );
 397     }
 398     else if( h->mb.i_type == I_4x4 )
 399     {
 400         h->mb.b_transform_8x8 = 0;
 401         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 402         if( h->mb.i_skip_intra )
 403         {
 404             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 405             /* In RD mode, restore the now-overwritten DCT data. */
 406             if( h->mb.i_skip_intra == 2 )
 407                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 408         }
 409         for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 410         {
 411             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][4 * block_idx_x[i] + 4 * block_idx_y[i] * FDEC_STRIDE];
 412             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 413
 414             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 415                 /* emulate missing topright samples */
 416                 *(uint32_t*) &p_dst[4-FDEC_STRIDE] = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 417
 418             h->predict_4x4[i_mode]( p_dst );
 419             x264_mb_encode_i4x4( h, i, i_qp );
 420         }
 421     }
 422     else    /* Inter MB */
 423     {
 424         int i8x8, i4x4;
 425         int i_decimate_mb = 0;
 426
 427         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
 428         if( !h->mb.b_skip_mc )
 429             x264_mb_mc( h );
 430
 431         if( h->mb.b_lossless )
 432         {
 433             for( i4x4 = 0; i4x4 < 16; i4x4++ )
 434             {
 435                 int x = 4*block_idx_x[i4x4];
 436                 int y = 4*block_idx_y[i4x4];
 437                 h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 438                                     h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
 439                                     h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
 440             }
 441         }
 442         else if( h->mb.b_transform_8x8 )
 443         {
 444             DECLARE_ALIGNED_16( int16_t dct8x8[4][8][8] );
 445             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 446             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 447             h->nr_count[1] += h->mb.b_noise_reduction * 4;
 448
 449             for( idx = 0; idx < 4; idx++ )
 450             {
 451                 if( h->mb.b_noise_reduction )
 452                     h->quantf.denoise_dct_core( *dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
 453                 if( h->mb.b_trellis )
 454                     x264_quant_8x8_trellis( h, dct8x8[idx], CQM_8PY, i_qp, 0 );
 455                 else
 456                     h->quantf.quant_8x8( dct8x8[idx], h->quant8_mf[CQM_8PY][i_qp], h->quant8_bias[CQM_8PY][i_qp] );
 457
 458                 h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 459
 460                 if( b_decimate )
 461                 {
 462                     int i_decimate_8x8 = x264_mb_decimate_score( h->dct.luma8x8[idx], 64 );
 463                     i_decimate_mb += i_decimate_8x8;
 464                     if( i_decimate_8x8 < 4 )
 465                         nnz8x8[idx] = 0;
 466                 }
 467                 else
 468                     nnz8x8[idx] = array_non_zero( dct8x8[idx] );
 469             }
 470
 471             if( i_decimate_mb < 6 && b_decimate )
 472                 *(uint32_t*)nnz8x8 = 0;
 473             else
 474             {
 475                 for( idx = 0; idx < 4; idx++ )
 476                     if( nnz8x8[idx] )
 477                     {
 478                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 479                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 480                     }
 481             }
 482         }
 483         else
 484         {
 485             DECLARE_ALIGNED_16( int16_t dct4x4[16][4][4] );
 486             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 487             h->nr_count[0] += h->mb.b_noise_reduction * 16;
 488
 489             for( i8x8 = 0; i8x8 < 4; i8x8++ )
 490             {
 491                 int i_decimate_8x8;
 492
 493                 /* encode one 4x4 block */
 494                 i_decimate_8x8 = 0;
 495                 for( i4x4 = 0; i4x4 < 4; i4x4++ )
 496                 {
 497                     idx = i8x8 * 4 + i4x4;
 498
 499                     if( h->mb.b_noise_reduction )
 500                         h->quantf.denoise_dct_core( *dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
 501                     if( h->mb.b_trellis )
 502                         x264_quant_4x4_trellis( h, dct4x4[idx], CQM_4PY, i_qp, DCT_LUMA_4x4, 0 );
 503                     else
 504                         h->quantf.quant_4x4( dct4x4[idx], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 505
 506                     h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 507
 508                     if( b_decimate && i_decimate_8x8 <= 6 )
 509                         i_decimate_8x8 += x264_mb_decimate_score( h->dct.luma4x4[idx], 16 );
 510                 }
 511
 512                 /* decimate this 8x8 block */
 513                 i_decimate_mb += i_decimate_8x8;
 514                 if( i_decimate_8x8 < 4 && b_decimate )
 515                     nnz8x8[i8x8] = 0;
 516             }
 517
 518             if( i_decimate_mb < 6 && b_decimate )
 519                 *(uint32_t*)nnz8x8 = 0;
 520             else
 521             {
 522                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 523                     if( nnz8x8[i8x8] )
 524                     {
 525                         for( i = 0; i < 4; i++ )
 526                             h->quantf.dequant_4x4( dct4x4[i8x8*4+i], h->dequant4_mf[CQM_4PY], i_qp );
 527                         h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 528                     }
 529             }
 530         }
 531     }
 532
 533     /* encode chroma */
 534     if( IS_INTRA( h->mb.i_type ) )
 535     {
 536         const int i_mode = h->mb.i_chroma_pred_mode;
 537         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 538         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 539     }
 540
 541     /* encode the 8x8 blocks */
 542     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 543
 544     /* coded block pattern and non_zero_count */
 545     h->mb.i_cbp_luma = 0x00;
 546     if( h->mb.i_type == I_16x16 )
 547     {
 548         for( i = 0; i < 16; i++ )
 549         {
 550             int nz = array_non_zero( h->dct.luma4x4[i] );
 551             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 552             h->mb.i_cbp_luma |= nz;
 553         }
 554         h->mb.i_cbp_luma *= 0xf;
 555     }
 556     else
 557     {
 558         for( i = 0; i < 4; i++)
 559         {
 560             if(!nnz8x8[i])
 561             {
 562                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[0+i*4]] = 0;
 563                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[2+i*4]] = 0;
 564             }
 565             else if( h->mb.b_transform_8x8 )
 566             {
 567                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[0+4*i]] = nnz8x8[i] * 0x0101;
 568                 *(uint16_t*)&h->mb.cache.non_zero_count[x264_scan8[2+4*i]] = nnz8x8[i] * 0x0101;
 569                 h->mb.i_cbp_luma |= nnz8x8[i] << i;
 570             }
 571             else
 572             {
 573                 int nz, cbp = 0;
 574                 for( j = 0; j < 4; j++ )
 575                 {
 576                     nz = array_non_zero( h->dct.luma4x4[j+4*i] );
 577                     h->mb.cache.non_zero_count[x264_scan8[j+4*i]] = nz;
 578                     cbp |= nz;
 579                 }
 580                 h->mb.i_cbp_luma |= cbp << i;
 581             }
 582         }
 583     }
 584
 585     if( h->param.b_cabac )
 586     {
 587         i_cbp_dc = ( h->mb.i_type == I_16x16 && array_non_zero( h->dct.luma16x16_dc ) )
 588                  | array_non_zero( h->dct.chroma_dc[0] ) << 1
 589                  | array_non_zero( h->dct.chroma_dc[1] ) << 2;
 590     }
 591
 592     /* store cbp */
 593     h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
 594
 595     /* Check for P_SKIP
 596      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 597      *      (if multiple mv give same result)*/
 598     if( !b_force_no_skip )
 599     {
 600         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 601             !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
 602             *(uint32_t*)h->mb.cache.mv[0][x264_scan8[0]] == *(uint32_t*)h->mb.cache.pskip_mv
 603             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 604         {
 605             h->mb.i_type = P_SKIP;
 606         }
 607
 608         /* Check for B_SKIP */
 609         if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
 610         {
 611             h->mb.i_type = B_SKIP;
 612         }
 613     }
 614 }
 615
 616 /*****************************************************************************
 617  * x264_macroblock_probe_skip:
 618  *  Check if the current MB could be encoded as a [PB]_SKIP (it supposes you use
 619  *  the previous QP
 620  *****************************************************************************/
 621 int x264_macroblock_probe_skip( x264_t *h, const int b_bidir )
 622 {
 623     DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 624     DECLARE_ALIGNED_16( int16_t dct2x2[2][2] );
 625     DECLARE_ALIGNED_16( int16_t dctscan[16] );
 626
 627     int i_qp = h->mb.i_qp;
 628     int mvp[2];
 629     int ch, thresh;
 630
 631     int i8x8, i4x4;
 632     int i_decimate_mb;
 633
 634     if( !b_bidir )
 635     {
 636         /* Get the MV */
 637         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 638         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 639
 640         /* Motion compensation */
 641         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 642                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 643                        mvp[0], mvp[1], 16, 16 );
 644     }
 645
 646     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 647     {
 648         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
 649         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
 650         /* get luma diff */
 651         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
 652                                     h->mb.pic.p_fdec[0] + fdec_offset );
 653         /* encode one 4x4 block */
 654         for( i4x4 = 0; i4x4 < 4; i4x4++ )
 655         {
 656             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 657             if( !array_non_zero(dct4x4[i4x4]) )
 658                 continue;
 659             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 660             i_decimate_mb += x264_mb_decimate_score( dctscan, 16 );
 661             if( i_decimate_mb >= 6 )
 662                 return 0;
 663         }
 664     }
 665
 666     /* encode chroma */
 667     i_qp = h->mb.i_chroma_qp;
 668     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 669
 670     for( ch = 0; ch < 2; ch++ )
 671     {
 672         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 673         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 674
 675         if( !b_bidir )
 676         {
 677             h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 678                              h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 679                              mvp[0], mvp[1], 8, 8 );
 680         }
 681
 682         /* there is almost never a termination during chroma, but we can't avoid the check entirely */
 683         /* so instead we check SSD and skip the actual check if the score is low enough. */
 684         if( h->pixf.ssd[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE ) < thresh )
 685             continue;
 686
 687         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 688
 689         /* calculate dct DC */
 690         dct2x2[0][0] = dct4x4[0][0][0];
 691         dct2x2[0][1] = dct4x4[1][0][0];
 692         dct2x2[1][0] = dct4x4[2][0][0];
 693         dct2x2[1][1] = dct4x4[3][0][0];
 694         h->dctf.dct2x2dc( dct2x2 );
 695         h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4PC][i_qp][0]>>1, h->quant4_bias[CQM_4PC][i_qp][0]<<1 );
 696         if( array_non_zero(dct2x2) )
 697             return 0;
 698
 699         /* calculate dct coeffs */
 700         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 701         {
 702             h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 703             if( !array_non_zero(dct4x4[i4x4]) )
 704                 continue;
 705             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 706             i_decimate_mb += x264_mb_decimate_score( dctscan+1, 15 );
 707             if( i_decimate_mb >= 7 )
 708                 return 0;
 709         }
 710     }
 711
 712     h->mb.b_skip_mc = 1;
 713     return 1;
 714 }
 715
 716 /****************************************************************************
 717  * DCT-domain noise reduction / adaptive deadzone
 718  * from libavcodec
 719  ****************************************************************************/
 720
 721 void x264_noise_reduction_update( x264_t *h )
 722 {
 723     int cat, i;
 724     for( cat = 0; cat < 2; cat++ )
 725     {
 726         int size = cat ? 64 : 16;
 727         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
 728
 729         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
 730         {
 731             for( i = 0; i < size; i++ )
 732                 h->nr_residual_sum[cat][i] >>= 1;
 733             h->nr_count[cat] >>= 1;
 734         }
 735
 736         for( i = 0; i < size; i++ )
 737             h->nr_offset[cat][i] =
 738                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
 739                  + h->nr_residual_sum[cat][i]/2)
 740               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
 741     }
 742 }
 743
 744 /*****************************************************************************
 745  * RD only; 4 calls to this do not make up for one macroblock_encode.
 746  * doesn't transform chroma dc.
 747  *****************************************************************************/
 748 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
 749 {
 750     int i_qp = h->mb.i_qp;
 751     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
 752     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
 753     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 754     int nnz8x8;
 755     int ch;
 756
 757     x264_mb_mc_8x8( h, i8 );
 758
 759     if( h->mb.b_transform_8x8 )
 760     {
 761         DECLARE_ALIGNED_16( int16_t dct8x8[8][8] );
 762         h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
 763         h->quantf.quant_8x8( dct8x8, h->quant8_mf[CQM_8PY][i_qp], h->quant8_bias[CQM_8PY][i_qp] );
 764         h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
 765
 766         if( b_decimate )
 767             nnz8x8 = 4 <= x264_mb_decimate_score( h->dct.luma8x8[i8], 64 );
 768         else
 769             nnz8x8 = array_non_zero( dct8x8 );
 770
 771         if( nnz8x8 )
 772         {
 773             h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
 774             h->dctf.add8x8_idct8( p_fdec, dct8x8 );
 775         }
 776     }
 777     else
 778     {
 779         int i4;
 780         DECLARE_ALIGNED_16( int16_t dct4x4[4][4][4] );
 781         h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
 782         for( i4 = 0; i4 < 4; i4++ )
 783             h->quantf.quant_4x4( dct4x4[i4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] );
 784         for( i4 = 0; i4 < 4; i4++ )
 785             h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
 786
 787         if( b_decimate )
 788         {
 789             int i_decimate_8x8 = 0;
 790             for( i4 = 0; i4 < 4 && i_decimate_8x8 < 4; i4++ )
 791                 i_decimate_8x8 += x264_mb_decimate_score( h->dct.luma4x4[i8*4+i4], 16 );
 792             nnz8x8 = 4 <= i_decimate_8x8;
 793         }
 794         else
 795             nnz8x8 = array_non_zero( dct4x4 );
 796
 797         if( nnz8x8 )
 798         {
 799             for( i4 = 0; i4 < 4; i4++ )
 800                 h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
 801             h->dctf.add8x8_idct( p_fdec, dct4x4 );
 802         }
 803     }
 804
 805     i_qp = h->mb.i_chroma_qp;
 806
 807     for( ch = 0; ch < 2; ch++ )
 808     {
 809         DECLARE_ALIGNED_16( int16_t dct4x4[4][4] );
 810         p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
 811         p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
 812
 813         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
 814         h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
 815         h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
 816         if( array_non_zero( dct4x4 ) )
 817         {
 818             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
 819             h->dctf.add4x4_idct( p_fdec, dct4x4 );
 820         }
 821     }
 822
 823     h->mb.i_cbp_luma &= ~(1 << i8);
 824     h->mb.i_cbp_luma |= nnz8x8 << i8;
 825     h->mb.i_cbp_chroma = 0x02;
 826 }