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

   1 /*****************************************************************************
   2  * macroblock.c: h264 encoder library
   3  *****************************************************************************
   4  * Copyright (C) 2003-2008 x264 project
   5  *
   6  * Authors: Laurent Aimar <fenrir@via.ecp.fr>
   7  *          Loren Merritt <lorenm@u.washington.edu>
   8  *          Fiona Glaser <fiona@x264.com>
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published by
  12  * the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  * GNU General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * along with this program; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
  23  *****************************************************************************/
  24
  25 #include "common/common.h"
  26 #include "macroblock.h"
  27
  28 /* These chroma DC functions don't have assembly versions and are only used here. */
  29
  30 #define ZIG(i,y,x) level[i] = dct[x*2+y];
  31 static inline void zigzag_scan_2x2_dc( int16_t level[4], int16_t dct[4] )
  32 {
  33     ZIG(0,0,0)
  34     ZIG(1,0,1)
  35     ZIG(2,1,0)
  36     ZIG(3,1,1)
  37 }
  38 #undef ZIG
  39
  40 #define IDCT_DEQUANT_START \
  41     int d0 = dct[0] + dct[1]; \
  42     int d1 = dct[2] + dct[3]; \
  43     int d2 = dct[0] - dct[1]; \
  44     int d3 = dct[2] - dct[3]; \
  45     int dmf = dequant_mf[i_qp%6][0] << i_qp/6;
  46
  47 static inline void idct_dequant_2x2_dc( int16_t dct[4], int16_t dct4x4[4][16], int dequant_mf[6][16], int i_qp )
  48 {
  49     IDCT_DEQUANT_START
  50     dct4x4[0][0] = (d0 + d1) * dmf >> 5;
  51     dct4x4[1][0] = (d0 - d1) * dmf >> 5;
  52     dct4x4[2][0] = (d2 + d3) * dmf >> 5;
  53     dct4x4[3][0] = (d2 - d3) * dmf >> 5;
  54 }
  55
  56 static inline void idct_dequant_2x2_dconly( int16_t out[4], int16_t dct[4], int dequant_mf[6][16], int i_qp )
  57 {
  58     IDCT_DEQUANT_START
  59     out[0] = (d0 + d1) * dmf >> 5;
  60     out[1] = (d0 - d1) * dmf >> 5;
  61     out[2] = (d2 + d3) * dmf >> 5;
  62     out[3] = (d2 - d3) * dmf >> 5;
  63 }
  64
  65 static inline void dct2x2dc( int16_t d[4], int16_t dct4x4[4][16] )
  66 {
  67     int d0 = dct4x4[0][0] + dct4x4[1][0];
  68     int d1 = dct4x4[2][0] + dct4x4[3][0];
  69     int d2 = dct4x4[0][0] - dct4x4[1][0];
  70     int d3 = dct4x4[2][0] - dct4x4[3][0];
  71     d[0] = d0 + d1;
  72     d[2] = d2 + d3;
  73     d[1] = d0 - d1;
  74     d[3] = d2 - d3;
  75     dct4x4[0][0] = 0;
  76     dct4x4[1][0] = 0;
  77     dct4x4[2][0] = 0;
  78     dct4x4[3][0] = 0;
  79 }
  80
  81 static ALWAYS_INLINE int x264_quant_4x4( x264_t *h, int16_t dct[16], int i_qp, int i_ctxBlockCat, int b_intra, int idx )
  82 {
  83     int i_quant_cat = b_intra ? CQM_4IY : CQM_4PY;
  84     if( h->mb.b_trellis )
  85         return x264_quant_4x4_trellis( h, dct, i_quant_cat, i_qp, i_ctxBlockCat, b_intra, 0, idx );
  86     else
  87         return h->quantf.quant_4x4( dct, h->quant4_mf[i_quant_cat][i_qp], h->quant4_bias[i_quant_cat][i_qp] );
  88 }
  89
  90 static ALWAYS_INLINE int x264_quant_8x8( x264_t *h, int16_t dct[64], int i_qp, int b_intra, int idx )
  91 {
  92     int i_quant_cat = b_intra ? CQM_8IY : CQM_8PY;
  93     if( h->mb.b_trellis )
  94         return x264_quant_8x8_trellis( h, dct, i_quant_cat, i_qp, b_intra, idx );
  95     else
  96         return h->quantf.quant_8x8( dct, h->quant8_mf[i_quant_cat][i_qp], h->quant8_bias[i_quant_cat][i_qp] );
  97 }
  98
  99 /* All encoding functions must output the correct CBP and NNZ values.
 100  * The entropy coding functions will check CBP first, then NNZ, before
 101  * actually reading the DCT coefficients.  NNZ still must be correct even
 102  * if CBP is zero because of the use of NNZ values for context selection.
 103  * "NNZ" need only be 0 or 1 rather than the exact coefficient count because
 104  * that is only needed in CAVLC, and will be calculated by CAVLC's residual
 105  * coding and stored as necessary. */
 106
 107 /* This means that decimation can be done merely by adjusting the CBP and NNZ
 108  * rather than memsetting the coefficients. */
 109
 110 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qp )
 111 {
 112     int nz;
 113     uint8_t *p_src = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[idx]];
 114     uint8_t *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[idx]];
 115     ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
 116
 117     if( h->mb.b_lossless )
 118     {
 119         nz = h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
 120         h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 121         h->mb.i_cbp_luma |= nz<<(idx>>2);
 122         return;
 123     }
 124
 125     h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
 126
 127     nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 1, idx );
 128     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 129     if( nz )
 130     {
 131         h->mb.i_cbp_luma |= 1<<(idx>>2);
 132         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4 );
 133         h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qp );
 134         h->dctf.add4x4_idct( p_dst, dct4x4 );
 135     }
 136 }
 137
 138 #define STORE_8x8_NNZ(idx,nz)\
 139 {\
 140     M16( &h->mb.cache.non_zero_count[x264_scan8[idx*4+0]] ) = nz * 0x0101;\
 141     M16( &h->mb.cache.non_zero_count[x264_scan8[idx*4+2]] ) = nz * 0x0101;\
 142 }
 143
 144 #define CLEAR_16x16_NNZ \
 145 {\
 146     M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = 0;\
 147     M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = 0;\
 148     M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = 0;\
 149     M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = 0;\
 150 }
 151
 152 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qp )
 153 {
 154     int x = 8 * (idx&1);
 155     int y = 8 * (idx>>1);
 156     int nz;
 157     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
 158     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
 159     ALIGNED_ARRAY_16( int16_t, dct8x8,[64] );
 160
 161     if( h->mb.b_lossless )
 162     {
 163         nz = h->zigzagf.sub_8x8( h->dct.luma8x8[idx], p_src, p_dst );
 164         STORE_8x8_NNZ(idx,nz);
 165         h->mb.i_cbp_luma |= nz<<idx;
 166         return;
 167     }
 168
 169     h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
 170
 171     nz = x264_quant_8x8( h, dct8x8, i_qp, 1, idx );
 172     if( nz )
 173     {
 174         h->mb.i_cbp_luma |= 1<<idx;
 175         h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
 176         h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qp );
 177         h->dctf.add8x8_idct8( p_dst, dct8x8 );
 178         STORE_8x8_NNZ(idx,1);
 179     }
 180     else
 181         STORE_8x8_NNZ(idx,0);
 182 }
 183
 184 static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
 185 {
 186     uint8_t  *p_src = h->mb.pic.p_fenc[0];
 187     uint8_t  *p_dst = h->mb.pic.p_fdec[0];
 188
 189     ALIGNED_ARRAY_16( int16_t, dct4x4,[16],[16] );
 190     ALIGNED_ARRAY_16( int16_t, dct_dc4x4,[16] );
 191
 192     int nz;
 193     int decimate_score = h->mb.b_dct_decimate ? 0 : 9;
 194
 195     if( h->mb.b_lossless )
 196     {
 197         for( int i = 0; i < 16; i++ )
 198         {
 199             int oe = block_idx_xy_fenc[i];
 200             int od = block_idx_xy_fdec[i];
 201             nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[i], p_src+oe, p_dst+od, &dct_dc4x4[block_idx_yx_1d[i]] );
 202             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 203             h->mb.i_cbp_luma |= nz;
 204         }
 205         h->mb.i_cbp_luma *= 0xf;
 206         h->mb.cache.non_zero_count[x264_scan8[24]] = array_non_zero( dct_dc4x4 );
 207         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 208         return;
 209     }
 210
 211     h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
 212
 213     for( int i = 0; i < 16; i++ )
 214     {
 215         /* copy dc coeff */
 216         dct_dc4x4[block_idx_xy_1d[i]] = dct4x4[i][0];
 217         dct4x4[i][0] = 0;
 218
 219         /* quant/scan/dequant */
 220         nz = x264_quant_4x4( h, dct4x4[i], i_qp, DCT_LUMA_AC, 1, i );
 221         h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 222         if( nz )
 223         {
 224             h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[i] );
 225             h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IY], i_qp );
 226             if( decimate_score < 6 ) decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[i] );
 227             h->mb.i_cbp_luma = 0xf;
 228         }
 229     }
 230
 231     /* Writing the 16 CBFs in an i16x16 block is quite costly, so decimation can save many bits. */
 232     /* More useful with CAVLC, but still useful with CABAC. */
 233     if( decimate_score < 6 )
 234     {
 235         h->mb.i_cbp_luma = 0;
 236         CLEAR_16x16_NNZ
 237     }
 238
 239     h->dctf.dct4x4dc( dct_dc4x4 );
 240     if( h->mb.b_trellis )
 241         nz = x264_quant_dc_trellis( h, dct_dc4x4, CQM_4IY, i_qp, DCT_LUMA_DC, 1, 0 );
 242     else
 243         nz = h->quantf.quant_4x4_dc( dct_dc4x4, h->quant4_mf[CQM_4IY][i_qp][0]>>1, h->quant4_bias[CQM_4IY][i_qp][0]<<1 );
 244
 245     h->mb.cache.non_zero_count[x264_scan8[24]] = nz;
 246     if( nz )
 247     {
 248         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 249
 250         /* output samples to fdec */
 251         h->dctf.idct4x4dc( dct_dc4x4 );
 252         h->quantf.dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[CQM_4IY], i_qp );  /* XXX not inversed */
 253         if( h->mb.i_cbp_luma )
 254             for( int i = 0; i < 16; i++ )
 255                 dct4x4[i][0] = dct_dc4x4[block_idx_xy_1d[i]];
 256     }
 257
 258     /* put pixels to fdec */
 259     if( h->mb.i_cbp_luma )
 260         h->dctf.add16x16_idct( p_dst, dct4x4 );
 261     else if( nz )
 262         h->dctf.add16x16_idct_dc( p_dst, dct_dc4x4 );
 263 }
 264
 265 static inline int idct_dequant_round_2x2_dc( int16_t ref[4], int16_t dct[4], int dequant_mf[6][16], int i_qp )
 266 {
 267     int16_t out[4];
 268     idct_dequant_2x2_dconly( out, dct, dequant_mf, i_qp );
 269     return ((ref[0] ^ (out[0]+32))
 270           | (ref[1] ^ (out[1]+32))
 271           | (ref[2] ^ (out[2]+32))
 272           | (ref[3] ^ (out[3]+32))) >> 6;
 273 }
 274
 275 /* Round down coefficients losslessly in DC-only chroma blocks.
 276  * Unlike luma blocks, this can't be done with a lookup table or
 277  * other shortcut technique because of the interdependencies
 278  * between the coefficients due to the chroma DC transform. */
 279 static inline int x264_mb_optimize_chroma_dc( x264_t *h, int b_inter, int i_qp, int16_t dct2x2[4] )
 280 {
 281     int16_t dct2x2_orig[4];
 282     int coeff, nz;
 283
 284     /* If the QP is too high, there's no benefit to rounding optimization. */
 285     if( h->dequant4_mf[CQM_4IC + b_inter][i_qp%6][0] << (i_qp/6) > 32*64 )
 286         return 1;
 287
 288     idct_dequant_2x2_dconly( dct2x2_orig, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 289     dct2x2_orig[0] += 32;
 290     dct2x2_orig[1] += 32;
 291     dct2x2_orig[2] += 32;
 292     dct2x2_orig[3] += 32;
 293
 294     /* If the DC coefficients already round to zero, terminate early. */
 295     if( !((dct2x2_orig[0]|dct2x2_orig[1]|dct2x2_orig[2]|dct2x2_orig[3])>>6) )
 296         return 0;
 297
 298     /* Start with the highest frequency coefficient... is this the best option? */
 299     for( nz = 0, coeff = h->quantf.coeff_last[DCT_CHROMA_DC]( dct2x2 ); coeff >= 0; coeff-- )
 300     {
 301         int level = dct2x2[coeff];
 302         int sign = level>>31 | 1; /* dct2x2[coeff] < 0 ? -1 : 1 */
 303
 304         while( level )
 305         {
 306             dct2x2[coeff] = level - sign;
 307             if( idct_dequant_round_2x2_dc( dct2x2_orig, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp ) )
 308             {
 309                 nz = 1;
 310                 dct2x2[coeff] = level;
 311                 break;
 312             }
 313             level -= sign;
 314         }
 315     }
 316
 317     return nz;
 318 }
 319
 320 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
 321 {
 322     int nz, nz_dc;
 323     int b_decimate = b_inter && h->mb.b_dct_decimate;
 324     ALIGNED_ARRAY_16( int16_t, dct2x2,[4] );
 325     h->mb.i_cbp_chroma = 0;
 326
 327     /* Early termination: check variance of chroma residual before encoding.
 328      * Don't bother trying early termination at low QPs.
 329      * Values are experimentally derived. */
 330     if( b_decimate && i_qp >= (h->mb.b_trellis ? 12 : 18) )
 331     {
 332         int thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 333         int ssd[2];
 334         int score  = h->pixf.var2_8x8( h->mb.pic.p_fenc[1], FENC_STRIDE, h->mb.pic.p_fdec[1], FDEC_STRIDE, &ssd[0] );
 335             score += h->pixf.var2_8x8( h->mb.pic.p_fenc[2], FENC_STRIDE, h->mb.pic.p_fdec[2], FDEC_STRIDE, &ssd[1] );
 336         if( score < thresh*4 )
 337         {
 338             h->mb.cache.non_zero_count[x264_scan8[16]] = 0;
 339             h->mb.cache.non_zero_count[x264_scan8[17]] = 0;
 340             h->mb.cache.non_zero_count[x264_scan8[18]] = 0;
 341             h->mb.cache.non_zero_count[x264_scan8[19]] = 0;
 342             h->mb.cache.non_zero_count[x264_scan8[20]] = 0;
 343             h->mb.cache.non_zero_count[x264_scan8[21]] = 0;
 344             h->mb.cache.non_zero_count[x264_scan8[22]] = 0;
 345             h->mb.cache.non_zero_count[x264_scan8[23]] = 0;
 346             M16( &h->mb.cache.non_zero_count[x264_scan8[25]] ) = 0;
 347
 348             for( int ch = 0; ch < 2; ch++ )
 349             {
 350                 if( ssd[ch] > thresh )
 351                 {
 352                     h->dctf.sub8x8_dct_dc( dct2x2, h->mb.pic.p_fenc[1+ch], h->mb.pic.p_fdec[1+ch] );
 353                     if( h->mb.b_trellis )
 354                         nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 355                     else
 356                         nz_dc = h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4IC+b_inter][i_qp][0]>>1, h->quant4_bias[CQM_4IC+b_inter][i_qp][0]<<1 );
 357
 358                     if( nz_dc )
 359                     {
 360                         if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
 361                             continue;
 362                         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 1;
 363                         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 364                         idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 365                         h->dctf.add8x8_idct_dc( h->mb.pic.p_fdec[1+ch], dct2x2 );
 366                         h->mb.i_cbp_chroma = 1;
 367                     }
 368                 }
 369             }
 370             return;
 371         }
 372     }
 373
 374     for( int ch = 0; ch < 2; ch++ )
 375     {
 376         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 377         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 378         int i_decimate_score = 0;
 379         int nz_ac = 0;
 380
 381         ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
 382
 383         if( h->mb.b_lossless )
 384         {
 385             for( int i = 0; i < 4; i++ )
 386             {
 387                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 388                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 389                 nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od, &h->dct.chroma_dc[ch][i] );
 390                 h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 391                 h->mb.i_cbp_chroma |= nz;
 392             }
 393             h->mb.cache.non_zero_count[x264_scan8[25]+ch] = array_non_zero( h->dct.chroma_dc[ch] );
 394             continue;
 395         }
 396
 397         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 398         dct2x2dc( dct2x2, dct4x4 );
 399         /* calculate dct coeffs */
 400         for( int i = 0; i < 4; i++ )
 401         {
 402             if( h->mb.b_trellis )
 403                 nz = x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 1, 0 );
 404             else
 405                 nz = h->quantf.quant_4x4( dct4x4[i], h->quant4_mf[CQM_4IC+b_inter][i_qp], h->quant4_bias[CQM_4IC+b_inter][i_qp] );
 406             h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 407             if( nz )
 408             {
 409                 nz_ac = 1;
 410                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 411                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 412                 if( b_decimate )
 413                     i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[16+i+ch*4] );
 414             }
 415         }
 416
 417         if( h->mb.b_trellis )
 418             nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 419         else
 420             nz_dc = h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4IC+b_inter][i_qp][0]>>1, h->quant4_bias[CQM_4IC+b_inter][i_qp][0]<<1 );
 421
 422         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = nz_dc;
 423
 424         if( (b_decimate && i_decimate_score < 7) || !nz_ac )
 425         {
 426             /* Decimate the block */
 427             h->mb.cache.non_zero_count[x264_scan8[16+0]+24*ch] = 0;
 428             h->mb.cache.non_zero_count[x264_scan8[16+1]+24*ch] = 0;
 429             h->mb.cache.non_zero_count[x264_scan8[16+2]+24*ch] = 0;
 430             h->mb.cache.non_zero_count[x264_scan8[16+3]+24*ch] = 0;
 431             if( !nz_dc ) /* Whole block is empty */
 432                 continue;
 433             if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
 434             {
 435                 h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 0;
 436                 continue;
 437             }
 438             /* DC-only */
 439             zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 440             idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 441             h->dctf.add8x8_idct_dc( p_dst, dct2x2 );
 442         }
 443         else
 444         {
 445             h->mb.i_cbp_chroma = 1;
 446             if( nz_dc )
 447             {
 448                 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 449                 idct_dequant_2x2_dc( dct2x2, dct4x4, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 450             }
 451             h->dctf.add8x8_idct( p_dst, dct4x4 );
 452         }
 453     }
 454
 455     /* 0 = none, 1 = DC only, 2 = DC+AC */
 456     h->mb.i_cbp_chroma = ((!!M16( &h->mb.cache.non_zero_count[x264_scan8[25]] )) | h->mb.i_cbp_chroma) + h->mb.i_cbp_chroma;
 457 }
 458
 459 static void x264_macroblock_encode_skip( x264_t *h )
 460 {
 461     for( int i = 0; i < sizeof( h->mb.cache.non_zero_count ); i += 16 )
 462         M128( &h->mb.cache.non_zero_count[i] ) = M128_ZERO;
 463     h->mb.i_cbp_luma = 0;
 464     h->mb.i_cbp_chroma = 0;
 465     h->mb.cbp[h->mb.i_mb_xy] = 0;
 466 }
 467
 468 /*****************************************************************************
 469  * x264_macroblock_encode_pskip:
 470  *  Encode an already marked skip block
 471  *****************************************************************************/
 472 static void x264_macroblock_encode_pskip( x264_t *h )
 473 {
 474     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 475     if( !h->mb.b_skip_mc )
 476     {
 477         int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 478                               h->mb.mv_min[0], h->mb.mv_max[0] );
 479         int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 480                               h->mb.mv_min[1], h->mb.mv_max[1] );
 481
 482         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 483                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 484                        mvx, mvy, 16, 16, &h->sh.weight[0][0] );
 485
 486         /* Special case for mv0, which is (of course) very common in P-skip mode. */
 487         if( mvx | mvy )
 488         {
 489             h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 490                              h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 491                              mvx, mvy, 8, 8 );
 492             h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 493                              h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 494                              mvx, mvy, 8, 8 );
 495         }
 496         else
 497         {
 498             h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1], 8 );
 499             h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2], 8 );
 500         }
 501
 502         if( h->sh.weight[0][1].weightfn )
 503             h->sh.weight[0][1].weightfn[8>>2]( h->mb.pic.p_fdec[1], FDEC_STRIDE,
 504                                                h->mb.pic.p_fdec[1], FDEC_STRIDE,
 505                                                &h->sh.weight[0][1], 8 );
 506
 507         if( h->sh.weight[0][2].weightfn )
 508             h->sh.weight[0][2].weightfn[8>>2]( h->mb.pic.p_fdec[2], FDEC_STRIDE,
 509                                                h->mb.pic.p_fdec[2], FDEC_STRIDE,
 510                                                &h->sh.weight[0][2], 8 );
 511     }
 512
 513     x264_macroblock_encode_skip( h );
 514 }
 515
 516 /*****************************************************************************
 517  * Intra prediction for predictive lossless mode.
 518  *****************************************************************************/
 519
 520 /* Note that these functions take a shortcut (mc.copy instead of actual pixel prediction) which assumes
 521  * that the edge pixels of the reconstructed frame are the same as that of the source frame.  This means
 522  * they will only work correctly if the neighboring blocks are losslessly coded.  In practice, this means
 523  * lossless mode cannot be mixed with lossy mode within a frame. */
 524 /* This can be resolved by explicitly copying the edge pixels after doing the mc.copy, but this doesn't
 525  * need to be done unless we decide to allow mixing lossless and lossy compression. */
 526
 527 void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
 528 {
 529     int stride = h->fenc->i_stride[1] << h->mb.b_interlaced;
 530     if( i_mode == I_PRED_CHROMA_V )
 531     {
 532         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-stride, stride, 8 );
 533         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-stride, stride, 8 );
 534     }
 535     else if( i_mode == I_PRED_CHROMA_H )
 536     {
 537         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-1, stride, 8 );
 538         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-1, stride, 8 );
 539     }
 540     else
 541     {
 542         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 543         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 544     }
 545 }
 546
 547 void x264_predict_lossless_4x4( x264_t *h, uint8_t *p_dst, int idx, int i_mode )
 548 {
 549     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 550     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
 551
 552     if( i_mode == I_PRED_4x4_V )
 553         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
 554     else if( i_mode == I_PRED_4x4_H )
 555         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-1, stride, 4 );
 556     else
 557         h->predict_4x4[i_mode]( p_dst );
 558 }
 559
 560 void x264_predict_lossless_8x8( x264_t *h, uint8_t *p_dst, int idx, int i_mode, uint8_t edge[33] )
 561 {
 562     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 563     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
 564
 565     if( i_mode == I_PRED_8x8_V )
 566         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
 567     else if( i_mode == I_PRED_8x8_H )
 568         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-1, stride, 8 );
 569     else
 570         h->predict_8x8[i_mode]( p_dst, edge );
 571 }
 572
 573 void x264_predict_lossless_16x16( x264_t *h, int i_mode )
 574 {
 575     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 576     if( i_mode == I_PRED_16x16_V )
 577         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-stride, stride, 16 );
 578     else if( i_mode == I_PRED_16x16_H )
 579         h->mc.copy_16x16_unaligned( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-1, stride, 16 );
 580     else
 581         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 582 }
 583
 584 /*****************************************************************************
 585  * x264_macroblock_encode:
 586  *****************************************************************************/
 587 void x264_macroblock_encode( x264_t *h )
 588 {
 589     int i_qp = h->mb.i_qp;
 590     int b_decimate = h->mb.b_dct_decimate;
 591     int b_force_no_skip = 0;
 592     int nz;
 593     h->mb.i_cbp_luma = 0;
 594     h->mb.cache.non_zero_count[x264_scan8[24]] = 0;
 595
 596     if( h->mb.i_type == I_PCM )
 597     {
 598         /* if PCM is chosen, we need to store reconstructed frame data */
 599         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE, 16 );
 600         h->mc.copy[PIXEL_8x8]  ( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc[1], FENC_STRIDE, 8 );
 601         h->mc.copy[PIXEL_8x8]  ( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc[2], FENC_STRIDE, 8 );
 602         return;
 603     }
 604
 605     if( h->sh.b_mbaff
 606         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 607         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 608     {
 609         /* The first skip is predicted to be a frame mb pair.
 610          * We don't yet support the aff part of mbaff, so force it to non-skip
 611          * so that we can pick the aff flag. */
 612         b_force_no_skip = 1;
 613         if( IS_SKIP(h->mb.i_type) )
 614         {
 615             if( h->mb.i_type == P_SKIP )
 616                 h->mb.i_type = P_L0;
 617             else if( h->mb.i_type == B_SKIP )
 618                 h->mb.i_type = B_DIRECT;
 619         }
 620     }
 621
 622     if( h->mb.i_type == P_SKIP )
 623     {
 624         /* A bit special */
 625         x264_macroblock_encode_pskip( h );
 626         return;
 627     }
 628     if( h->mb.i_type == B_SKIP )
 629     {
 630         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 631         if( !h->mb.b_skip_mc )
 632             x264_mb_mc( h );
 633         x264_macroblock_encode_skip( h );
 634         return;
 635     }
 636
 637     if( h->mb.i_type == I_16x16 )
 638     {
 639         const int i_mode = h->mb.i_intra16x16_pred_mode;
 640         h->mb.b_transform_8x8 = 0;
 641
 642         if( h->mb.b_lossless )
 643             x264_predict_lossless_16x16( h, i_mode );
 644         else
 645             h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 646
 647         /* encode the 16x16 macroblock */
 648         x264_mb_encode_i16x16( h, i_qp );
 649     }
 650     else if( h->mb.i_type == I_8x8 )
 651     {
 652         ALIGNED_ARRAY_16( uint8_t, edge,[33] );
 653         h->mb.b_transform_8x8 = 1;
 654         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 655         if( h->mb.i_skip_intra )
 656         {
 657             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 658             M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i8x8_nnz_buf[0];
 659             M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i8x8_nnz_buf[1];
 660             M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i8x8_nnz_buf[2];
 661             M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i8x8_nnz_buf[3];
 662             h->mb.i_cbp_luma = h->mb.pic.i8x8_cbp;
 663             /* In RD mode, restore the now-overwritten DCT data. */
 664             if( h->mb.i_skip_intra == 2 )
 665                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 666         }
 667         for( int i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 668         {
 669             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 670             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 671             h->predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 672
 673             if( h->mb.b_lossless )
 674                 x264_predict_lossless_8x8( h, p_dst, i, i_mode, edge );
 675             else
 676                 h->predict_8x8[i_mode]( p_dst, edge );
 677
 678             x264_mb_encode_i8x8( h, i, i_qp );
 679         }
 680     }
 681     else if( h->mb.i_type == I_4x4 )
 682     {
 683         h->mb.b_transform_8x8 = 0;
 684         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 685         if( h->mb.i_skip_intra )
 686         {
 687             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 688             M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i4x4_nnz_buf[0];
 689             M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i4x4_nnz_buf[1];
 690             M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i4x4_nnz_buf[2];
 691             M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i4x4_nnz_buf[3];
 692             h->mb.i_cbp_luma = h->mb.pic.i4x4_cbp;
 693             /* In RD mode, restore the now-overwritten DCT data. */
 694             if( h->mb.i_skip_intra == 2 )
 695                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 696         }
 697         for( int i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 698         {
 699             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
 700             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 701
 702             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 703                 /* emulate missing topright samples */
 704                 M32( &p_dst[4-FDEC_STRIDE] ) = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 705
 706             if( h->mb.b_lossless )
 707                 x264_predict_lossless_4x4( h, p_dst, i, i_mode );
 708             else
 709                 h->predict_4x4[i_mode]( p_dst );
 710             x264_mb_encode_i4x4( h, i, i_qp );
 711         }
 712     }
 713     else    /* Inter MB */
 714     {
 715         int i_decimate_mb = 0;
 716
 717         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
 718         if( !h->mb.b_skip_mc )
 719             x264_mb_mc( h );
 720
 721         if( h->mb.b_lossless )
 722         {
 723             if( h->mb.b_transform_8x8 )
 724                 for( int i8x8 = 0; i8x8 < 4; i8x8++ )
 725                 {
 726                     int x = 8*(i8x8&1);
 727                     int y = 8*(i8x8>>1);
 728                     nz = h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8],
 729                                         h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
 730                                         h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
 731                     STORE_8x8_NNZ(i8x8,nz);
 732                     h->mb.i_cbp_luma |= nz << i8x8;
 733                 }
 734             else
 735                 for( int i4x4 = 0; i4x4 < 16; i4x4++ )
 736                 {
 737                     nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 738                                         h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
 739                                         h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4x4] );
 740                     h->mb.cache.non_zero_count[x264_scan8[i4x4]] = nz;
 741                     h->mb.i_cbp_luma |= nz << (i4x4>>2);
 742                 }
 743         }
 744         else if( h->mb.b_transform_8x8 )
 745         {
 746             ALIGNED_ARRAY_16( int16_t, dct8x8,[4],[64] );
 747             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 748             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 749             h->nr_count[1] += h->mb.b_noise_reduction * 4;
 750
 751             for( int idx = 0; idx < 4; idx++ )
 752             {
 753                 if( h->mb.b_noise_reduction )
 754                     h->quantf.denoise_dct( dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
 755                 nz = x264_quant_8x8( h, dct8x8[idx], i_qp, 0, idx );
 756
 757                 if( nz )
 758                 {
 759                     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 760                     if( b_decimate )
 761                     {
 762                         int i_decimate_8x8 = h->quantf.decimate_score64( h->dct.luma8x8[idx] );
 763                         i_decimate_mb += i_decimate_8x8;
 764                         if( i_decimate_8x8 >= 4 )
 765                             h->mb.i_cbp_luma |= 1<<idx;
 766                     }
 767                     else
 768                         h->mb.i_cbp_luma |= 1<<idx;
 769                 }
 770             }
 771
 772             if( i_decimate_mb < 6 && b_decimate )
 773             {
 774                 h->mb.i_cbp_luma = 0;
 775                 CLEAR_16x16_NNZ
 776             }
 777             else
 778             {
 779                 for( int idx = 0; idx < 4; idx++ )
 780                 {
 781                     if( h->mb.i_cbp_luma&(1<<idx) )
 782                     {
 783                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 784                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 785                         STORE_8x8_NNZ(idx,1);
 786                     }
 787                     else
 788                         STORE_8x8_NNZ(idx,0);
 789                 }
 790             }
 791         }
 792         else
 793         {
 794             ALIGNED_ARRAY_16( int16_t, dct4x4,[16],[16] );
 795             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 796             h->nr_count[0] += h->mb.b_noise_reduction * 16;
 797
 798             for( int i8x8 = 0; i8x8 < 4; i8x8++ )
 799             {
 800                 int i_decimate_8x8 = 0;
 801                 int cbp = 0;
 802
 803                 /* encode one 4x4 block */
 804                 for( int i4x4 = 0; i4x4 < 4; i4x4++ )
 805                 {
 806                     int idx = i8x8 * 4 + i4x4;
 807
 808                     if( h->mb.b_noise_reduction )
 809                         h->quantf.denoise_dct( dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
 810                     nz = x264_quant_4x4( h, dct4x4[idx], i_qp, DCT_LUMA_4x4, 0, idx );
 811                     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 812
 813                     if( nz )
 814                     {
 815                         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 816                         h->quantf.dequant_4x4( dct4x4[idx], h->dequant4_mf[CQM_4PY], i_qp );
 817                         if( b_decimate && i_decimate_8x8 < 6 )
 818                             i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[idx] );
 819                         cbp = 1;
 820                     }
 821                 }
 822
 823                 /* decimate this 8x8 block */
 824                 i_decimate_mb += i_decimate_8x8;
 825                 if( b_decimate )
 826                 {
 827                     if( i_decimate_8x8 < 4 )
 828                         STORE_8x8_NNZ(i8x8,0)
 829                     else
 830                         h->mb.i_cbp_luma |= 1<<i8x8;
 831                 }
 832                 else if( cbp )
 833                 {
 834                     h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 835                     h->mb.i_cbp_luma |= 1<<i8x8;
 836                 }
 837             }
 838
 839             if( b_decimate )
 840             {
 841                 if( i_decimate_mb < 6 )
 842                 {
 843                     h->mb.i_cbp_luma = 0;
 844                     CLEAR_16x16_NNZ
 845                 }
 846                 else
 847                 {
 848                     for( int i8x8 = 0; i8x8 < 4; i8x8++ )
 849                         if( h->mb.i_cbp_luma&(1<<i8x8) )
 850                             h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 851                 }
 852             }
 853         }
 854     }
 855
 856     /* encode chroma */
 857     if( IS_INTRA( h->mb.i_type ) )
 858     {
 859         const int i_mode = h->mb.i_chroma_pred_mode;
 860         if( h->mb.b_lossless )
 861             x264_predict_lossless_8x8_chroma( h, i_mode );
 862         else
 863         {
 864             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 865             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 866         }
 867     }
 868
 869     /* encode the 8x8 blocks */
 870     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 871
 872     /* store cbp */
 873     int cbp = h->mb.i_cbp_chroma << 4 | h->mb.i_cbp_luma;
 874     if( h->param.b_cabac )
 875         cbp |= h->mb.cache.non_zero_count[x264_scan8[24]] << 8
 876             |  h->mb.cache.non_zero_count[x264_scan8[25]] << 9
 877             |  h->mb.cache.non_zero_count[x264_scan8[26]] << 10;
 878     h->mb.cbp[h->mb.i_mb_xy] = cbp;
 879
 880     /* Check for P_SKIP
 881      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 882      *      (if multiple mv give same result)*/
 883     if( !b_force_no_skip )
 884     {
 885         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 886             !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
 887             M32( h->mb.cache.mv[0][x264_scan8[0]] ) == M32( h->mb.cache.pskip_mv )
 888             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 889         {
 890             h->mb.i_type = P_SKIP;
 891         }
 892
 893         /* Check for B_SKIP */
 894         if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
 895         {
 896             h->mb.i_type = B_SKIP;
 897         }
 898     }
 899 }
 900
 901 /*****************************************************************************
 902  * x264_macroblock_probe_skip:
 903  *  Check if the current MB could be encoded as a [PB]_SKIP
 904  *****************************************************************************/
 905 int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
 906 {
 907     ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
 908     ALIGNED_ARRAY_16( int16_t, dct2x2,[4] );
 909     ALIGNED_ARRAY_16( int16_t, dctscan,[16] );
 910     ALIGNED_4( int16_t mvp[2] );
 911
 912     int i_qp = h->mb.i_qp;
 913     int thresh, ssd;
 914
 915     if( !b_bidir )
 916     {
 917         /* Get the MV */
 918         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 919         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 920
 921         /* Motion compensation */
 922         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 923                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 924                        mvp[0], mvp[1], 16, 16, &h->sh.weight[0][0] );
 925     }
 926
 927     for( int i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 928     {
 929         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
 930         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
 931         /* get luma diff */
 932         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
 933                                     h->mb.pic.p_fdec[0] + fdec_offset );
 934         /* encode one 4x4 block */
 935         for( int i4x4 = 0; i4x4 < 4; i4x4++ )
 936         {
 937             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] ) )
 938                 continue;
 939             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 940             i_decimate_mb += h->quantf.decimate_score16( dctscan );
 941             if( i_decimate_mb >= 6 )
 942                 return 0;
 943         }
 944     }
 945
 946     /* encode chroma */
 947     i_qp = h->mb.i_chroma_qp;
 948     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 949
 950     for( int ch = 0; ch < 2; ch++ )
 951     {
 952         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 953         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 954
 955         if( !b_bidir )
 956         {
 957             /* Special case for mv0, which is (of course) very common in P-skip mode. */
 958             if( M32( mvp ) )
 959             {
 960                 h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 961                                  h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 962                                  mvp[0], mvp[1], 8, 8 );
 963             }
 964             else
 965                 h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1+ch], FDEC_STRIDE, h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch], 8 );
 966
 967             if( h->sh.weight[0][1+ch].weightfn )
 968                 h->sh.weight[0][1+ch].weightfn[8>>2]( h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
 969                                                       h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
 970                                                       &h->sh.weight[0][1+ch], 8 );
 971         }
 972
 973         /* there is almost never a termination during chroma, but we can't avoid the check entirely */
 974         /* so instead we check SSD and skip the actual check if the score is low enough. */
 975         ssd = h->pixf.ssd[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE );
 976         if( ssd < thresh )
 977             continue;
 978
 979         /* The vast majority of chroma checks will terminate during the DC check or the higher
 980          * threshold check, so we can save time by doing a DC-only DCT. */
 981         h->dctf.sub8x8_dct_dc( dct2x2, p_src, p_dst );
 982
 983         if( h->quantf.quant_2x2_dc( dct2x2, h->quant4_mf[CQM_4PC][i_qp][0]>>1, h->quant4_bias[CQM_4PC][i_qp][0]<<1 ) )
 984             return 0;
 985
 986         /* If there wasn't a termination in DC, we can check against a much higher threshold. */
 987         if( ssd < thresh*4 )
 988             continue;
 989
 990         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 991
 992         /* calculate dct coeffs */
 993         for( int i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 994         {
 995             /* We don't need to zero the DC coefficient before quantization because we already
 996              * checked that all the DCs were zero above at twice the precision that quant4x4
 997              * uses.  This applies even though the DC here is being quantized before the 2x2
 998              * transform. */
 999             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] ) )
1000                 continue;
1001             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
1002             i_decimate_mb += h->quantf.decimate_score15( dctscan );
1003             if( i_decimate_mb >= 7 )
1004                 return 0;
1005         }
1006     }
1007
1008     h->mb.b_skip_mc = 1;
1009     return 1;
1010 }
1011
1012 /****************************************************************************
1013  * DCT-domain noise reduction / adaptive deadzone
1014  * from libavcodec
1015  ****************************************************************************/
1016
1017 void x264_noise_reduction_update( x264_t *h )
1018 {
1019     for( int cat = 0; cat < 2; cat++ )
1020     {
1021         int size = cat ? 64 : 16;
1022         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
1023
1024         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
1025         {
1026             for( int i = 0; i < size; i++ )
1027                 h->nr_residual_sum[cat][i] >>= 1;
1028             h->nr_count[cat] >>= 1;
1029         }
1030
1031         for( int i = 0; i < size; i++ )
1032             h->nr_offset[cat][i] =
1033                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
1034                  + h->nr_residual_sum[cat][i]/2)
1035               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
1036     }
1037 }
1038
1039 /*****************************************************************************
1040  * RD only; 4 calls to this do not make up for one macroblock_encode.
1041  * doesn't transform chroma dc.
1042  *****************************************************************************/
1043 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
1044 {
1045     int i_qp = h->mb.i_qp;
1046     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
1047     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
1048     int b_decimate = h->mb.b_dct_decimate;
1049     int nnz8x8 = 0;
1050     int nz;
1051
1052     if( !h->mb.b_skip_mc )
1053         x264_mb_mc_8x8( h, i8 );
1054
1055     if( h->mb.b_lossless )
1056     {
1057         if( h->mb.b_transform_8x8 )
1058         {
1059             nnz8x8 = h->zigzagf.sub_8x8( h->dct.luma8x8[i8], p_fenc, p_fdec );
1060             STORE_8x8_NNZ(i8,nnz8x8);
1061         }
1062         else
1063         {
1064             for( int i4 = i8*4; i4 < i8*4+4; i4++ )
1065             {
1066                 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
1067                                     h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
1068                                     h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
1069                 h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1070                 nnz8x8 |= nz;
1071             }
1072         }
1073         for( int ch = 0; ch < 2; ch++ )
1074         {
1075             int16_t dc;
1076             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1077             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1078             nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec, &dc );
1079             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1080         }
1081     }
1082     else
1083     {
1084         if( h->mb.b_transform_8x8 )
1085         {
1086             ALIGNED_ARRAY_16( int16_t, dct8x8,[64] );
1087             h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
1088             nnz8x8 = x264_quant_8x8( h, dct8x8, i_qp, 0, i8 );
1089             if( nnz8x8 )
1090             {
1091                 h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
1092
1093                 if( b_decimate && !h->mb.b_trellis )
1094                     nnz8x8 = 4 <= h->quantf.decimate_score64( h->dct.luma8x8[i8] );
1095
1096                 if( nnz8x8 )
1097                 {
1098                     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
1099                     h->dctf.add8x8_idct8( p_fdec, dct8x8 );
1100                     STORE_8x8_NNZ(i8,1);
1101                 }
1102                 else
1103                     STORE_8x8_NNZ(i8,0);
1104             }
1105             else
1106                 STORE_8x8_NNZ(i8,0);
1107         }
1108         else
1109         {
1110             int i_decimate_8x8 = 0;
1111             ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
1112             h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
1113             for( int i4 = 0; i4 < 4; i4++ )
1114             {
1115                 nz = x264_quant_4x4( h, dct4x4[i4], i_qp, DCT_LUMA_4x4, 0, i8*4+i4 );
1116                 h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = nz;
1117                 if( nz )
1118                 {
1119                     h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
1120                     h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
1121                     if( b_decimate )
1122                         i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[i8*4+i4] );
1123                     nnz8x8 = 1;
1124                 }
1125             }
1126
1127             if( b_decimate && i_decimate_8x8 < 4 )
1128                 nnz8x8 = 0;
1129
1130             if( nnz8x8 )
1131                 h->dctf.add8x8_idct( p_fdec, dct4x4 );
1132             else
1133                 STORE_8x8_NNZ(i8,0);
1134         }
1135
1136         i_qp = h->mb.i_chroma_qp;
1137
1138         for( int ch = 0; ch < 2; ch++ )
1139         {
1140             ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
1141             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1142             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1143
1144             h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1145             dct4x4[0] = 0;
1146
1147             if( h->mb.b_trellis )
1148                 nz = x264_quant_4x4_trellis( h, dct4x4, CQM_4PC, i_qp, DCT_CHROMA_AC, 0, 1, 0 );
1149             else
1150                 nz = h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
1151
1152             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1153             if( nz )
1154             {
1155                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
1156                 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
1157                 h->dctf.add4x4_idct( p_fdec, dct4x4 );
1158             }
1159         }
1160     }
1161     h->mb.i_cbp_luma &= ~(1 << i8);
1162     h->mb.i_cbp_luma |= nnz8x8 << i8;
1163     h->mb.i_cbp_chroma = 0x02;
1164 }
1165
1166 /*****************************************************************************
1167  * RD only, luma only
1168  *****************************************************************************/
1169 void x264_macroblock_encode_p4x4( x264_t *h, int i4 )
1170 {
1171     int i_qp = h->mb.i_qp;
1172     uint8_t *p_fenc = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[i4]];
1173     uint8_t *p_fdec = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
1174     int nz;
1175
1176     /* Don't need motion compensation as this function is only used in qpel-RD, which caches pixel data. */
1177
1178     if( h->mb.b_lossless )
1179     {
1180         nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4], p_fenc, p_fdec );
1181         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1182     }
1183     else
1184     {
1185         ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
1186         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1187         nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 0, i4 );
1188         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1189         if( nz )
1190         {
1191             h->zigzagf.scan_4x4( h->dct.luma4x4[i4], dct4x4 );
1192             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PY], i_qp );
1193             h->dctf.add4x4_idct( p_fdec, dct4x4 );
1194         }
1195     }
1196 }