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         if( score < thresh*4 )
 336             score += h->pixf.var2_8x8( h->mb.pic.p_fenc[2], FENC_STRIDE, h->mb.pic.p_fdec[2], FDEC_STRIDE, &ssd[1] );
 337         if( score < thresh*4 )
 338         {
 339             h->mb.cache.non_zero_count[x264_scan8[16]] = 0;
 340             h->mb.cache.non_zero_count[x264_scan8[17]] = 0;
 341             h->mb.cache.non_zero_count[x264_scan8[18]] = 0;
 342             h->mb.cache.non_zero_count[x264_scan8[19]] = 0;
 343             h->mb.cache.non_zero_count[x264_scan8[20]] = 0;
 344             h->mb.cache.non_zero_count[x264_scan8[21]] = 0;
 345             h->mb.cache.non_zero_count[x264_scan8[22]] = 0;
 346             h->mb.cache.non_zero_count[x264_scan8[23]] = 0;
 347             M16( &h->mb.cache.non_zero_count[x264_scan8[25]] ) = 0;
 348
 349             for( int ch = 0; ch < 2; ch++ )
 350             {
 351                 if( ssd[ch] > thresh )
 352                 {
 353                     h->dctf.sub8x8_dct_dc( dct2x2, h->mb.pic.p_fenc[1+ch], h->mb.pic.p_fdec[1+ch] );
 354                     if( h->mb.b_trellis )
 355                         nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 356                     else
 357                         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 );
 358
 359                     if( nz_dc )
 360                     {
 361                         if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
 362                             continue;
 363                         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 1;
 364                         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 365                         idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 366                         h->dctf.add8x8_idct_dc( h->mb.pic.p_fdec[1+ch], dct2x2 );
 367                         h->mb.i_cbp_chroma = 1;
 368                     }
 369                 }
 370             }
 371             return;
 372         }
 373     }
 374
 375     for( int ch = 0; ch < 2; ch++ )
 376     {
 377         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 378         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 379         int i_decimate_score = 0;
 380         int nz_ac = 0;
 381
 382         ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
 383
 384         if( h->mb.b_lossless )
 385         {
 386             for( int i = 0; i < 4; i++ )
 387             {
 388                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 389                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 390                 nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od, &h->dct.chroma_dc[ch][i] );
 391                 h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 392                 h->mb.i_cbp_chroma |= nz;
 393             }
 394             h->mb.cache.non_zero_count[x264_scan8[25]+ch] = array_non_zero( h->dct.chroma_dc[ch] );
 395             continue;
 396         }
 397
 398         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 399         dct2x2dc( dct2x2, dct4x4 );
 400         /* calculate dct coeffs */
 401         for( int i = 0; i < 4; i++ )
 402         {
 403             if( h->mb.b_trellis )
 404                 nz = x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 1, 0 );
 405             else
 406                 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] );
 407             h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 408             if( nz )
 409             {
 410                 nz_ac = 1;
 411                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 412                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 413                 if( b_decimate )
 414                     i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[16+i+ch*4] );
 415             }
 416         }
 417
 418         if( h->mb.b_trellis )
 419             nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 420         else
 421             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 );
 422
 423         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = nz_dc;
 424
 425         if( (b_decimate && i_decimate_score < 7) || !nz_ac )
 426         {
 427             /* Decimate the block */
 428             h->mb.cache.non_zero_count[x264_scan8[16+0]+24*ch] = 0;
 429             h->mb.cache.non_zero_count[x264_scan8[16+1]+24*ch] = 0;
 430             h->mb.cache.non_zero_count[x264_scan8[16+2]+24*ch] = 0;
 431             h->mb.cache.non_zero_count[x264_scan8[16+3]+24*ch] = 0;
 432             if( !nz_dc ) /* Whole block is empty */
 433                 continue;
 434             if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
 435             {
 436                 h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 0;
 437                 continue;
 438             }
 439             /* DC-only */
 440             zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 441             idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 442             h->dctf.add8x8_idct_dc( p_dst, dct2x2 );
 443         }
 444         else
 445         {
 446             h->mb.i_cbp_chroma = 1;
 447             if( nz_dc )
 448             {
 449                 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 450                 idct_dequant_2x2_dc( dct2x2, dct4x4, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 451             }
 452             h->dctf.add8x8_idct( p_dst, dct4x4 );
 453         }
 454     }
 455
 456     /* 0 = none, 1 = DC only, 2 = DC+AC */
 457     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;
 458 }
 459
 460 static void x264_macroblock_encode_skip( x264_t *h )
 461 {
 462     M32( &h->mb.cache.non_zero_count[x264_scan8[0]+0*8] ) = 0;
 463     M32( &h->mb.cache.non_zero_count[x264_scan8[0]+1*8] ) = 0;
 464     M32( &h->mb.cache.non_zero_count[x264_scan8[0]+2*8] ) = 0;
 465     M32( &h->mb.cache.non_zero_count[x264_scan8[0]+3*8] ) = 0;
 466     for( int i = 16; i < 24; i++ )
 467         h->mb.cache.non_zero_count[x264_scan8[i]] = 0;
 468     h->mb.i_cbp_luma = 0;
 469     h->mb.i_cbp_chroma = 0;
 470     h->mb.cbp[h->mb.i_mb_xy] = 0;
 471 }
 472
 473 /*****************************************************************************
 474  * x264_macroblock_encode_pskip:
 475  *  Encode an already marked skip block
 476  *****************************************************************************/
 477 static void x264_macroblock_encode_pskip( x264_t *h )
 478 {
 479     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 480     if( !h->mb.b_skip_mc )
 481     {
 482         int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 483                               h->mb.mv_min[0], h->mb.mv_max[0] );
 484         int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 485                               h->mb.mv_min[1], h->mb.mv_max[1] );
 486
 487         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 488                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 489                        mvx, mvy, 16, 16, &h->sh.weight[0][0] );
 490
 491         /* Special case for mv0, which is (of course) very common in P-skip mode. */
 492         if( mvx | mvy )
 493         {
 494             h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 495                              h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 496                              mvx, mvy, 8, 8 );
 497             h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 498                              h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 499                              mvx, mvy, 8, 8 );
 500         }
 501         else
 502         {
 503             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 );
 504             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 );
 505         }
 506
 507         if( h->sh.weight[0][1].weightfn )
 508             h->sh.weight[0][1].weightfn[8>>2]( h->mb.pic.p_fdec[1], FDEC_STRIDE,
 509                                                h->mb.pic.p_fdec[1], FDEC_STRIDE,
 510                                                &h->sh.weight[0][1], 8 );
 511
 512         if( h->sh.weight[0][2].weightfn )
 513             h->sh.weight[0][2].weightfn[8>>2]( h->mb.pic.p_fdec[2], FDEC_STRIDE,
 514                                                h->mb.pic.p_fdec[2], FDEC_STRIDE,
 515                                                &h->sh.weight[0][2], 8 );
 516     }
 517
 518     x264_macroblock_encode_skip( h );
 519 }
 520
 521 /*****************************************************************************
 522  * Intra prediction for predictive lossless mode.
 523  *****************************************************************************/
 524
 525 /* Note that these functions take a shortcut (mc.copy instead of actual pixel prediction) which assumes
 526  * that the edge pixels of the reconstructed frame are the same as that of the source frame.  This means
 527  * they will only work correctly if the neighboring blocks are losslessly coded.  In practice, this means
 528  * lossless mode cannot be mixed with lossy mode within a frame. */
 529 /* This can be resolved by explicitly copying the edge pixels after doing the mc.copy, but this doesn't
 530  * need to be done unless we decide to allow mixing lossless and lossy compression. */
 531
 532 void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
 533 {
 534     int stride = h->fenc->i_stride[1] << h->mb.b_interlaced;
 535     if( i_mode == I_PRED_CHROMA_V )
 536     {
 537         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-stride, stride, 8 );
 538         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-stride, stride, 8 );
 539     }
 540     else if( i_mode == I_PRED_CHROMA_H )
 541     {
 542         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-1, stride, 8 );
 543         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-1, stride, 8 );
 544     }
 545     else
 546     {
 547         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 548         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 549     }
 550 }
 551
 552 void x264_predict_lossless_4x4( x264_t *h, uint8_t *p_dst, int idx, int i_mode )
 553 {
 554     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 555     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
 556
 557     if( i_mode == I_PRED_4x4_V )
 558         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
 559     else if( i_mode == I_PRED_4x4_H )
 560         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-1, stride, 4 );
 561     else
 562         h->predict_4x4[i_mode]( p_dst );
 563 }
 564
 565 void x264_predict_lossless_8x8( x264_t *h, uint8_t *p_dst, int idx, int i_mode, uint8_t edge[33] )
 566 {
 567     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 568     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
 569
 570     if( i_mode == I_PRED_8x8_V )
 571         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
 572     else if( i_mode == I_PRED_8x8_H )
 573         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-1, stride, 8 );
 574     else
 575         h->predict_8x8[i_mode]( p_dst, edge );
 576 }
 577
 578 void x264_predict_lossless_16x16( x264_t *h, int i_mode )
 579 {
 580     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 581     if( i_mode == I_PRED_16x16_V )
 582         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-stride, stride, 16 );
 583     else if( i_mode == I_PRED_16x16_H )
 584         h->mc.copy_16x16_unaligned( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-1, stride, 16 );
 585     else
 586         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 587 }
 588
 589 /*****************************************************************************
 590  * x264_macroblock_encode:
 591  *****************************************************************************/
 592 void x264_macroblock_encode( x264_t *h )
 593 {
 594     int i_qp = h->mb.i_qp;
 595     int b_decimate = h->mb.b_dct_decimate;
 596     int b_force_no_skip = 0;
 597     int nz;
 598     h->mb.i_cbp_luma = 0;
 599     h->mb.cache.non_zero_count[x264_scan8[24]] = 0;
 600
 601     if( h->mb.i_type == I_PCM )
 602     {
 603         /* if PCM is chosen, we need to store reconstructed frame data */
 604         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc[0], FENC_STRIDE, 16 );
 605         h->mc.copy[PIXEL_8x8]  ( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc[1], FENC_STRIDE, 8 );
 606         h->mc.copy[PIXEL_8x8]  ( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc[2], FENC_STRIDE, 8 );
 607         return;
 608     }
 609
 610     if( h->sh.b_mbaff
 611         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 612         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 613     {
 614         /* The first skip is predicted to be a frame mb pair.
 615          * We don't yet support the aff part of mbaff, so force it to non-skip
 616          * so that we can pick the aff flag. */
 617         b_force_no_skip = 1;
 618         if( IS_SKIP(h->mb.i_type) )
 619         {
 620             if( h->mb.i_type == P_SKIP )
 621                 h->mb.i_type = P_L0;
 622             else if( h->mb.i_type == B_SKIP )
 623                 h->mb.i_type = B_DIRECT;
 624         }
 625     }
 626
 627     if( h->mb.i_type == P_SKIP )
 628     {
 629         /* A bit special */
 630         x264_macroblock_encode_pskip( h );
 631         return;
 632     }
 633     if( h->mb.i_type == B_SKIP )
 634     {
 635         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 636         if( !h->mb.b_skip_mc )
 637             x264_mb_mc( h );
 638         x264_macroblock_encode_skip( h );
 639         return;
 640     }
 641
 642     if( h->mb.i_type == I_16x16 )
 643     {
 644         const int i_mode = h->mb.i_intra16x16_pred_mode;
 645         h->mb.b_transform_8x8 = 0;
 646
 647         if( h->mb.b_lossless )
 648             x264_predict_lossless_16x16( h, i_mode );
 649         else
 650             h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 651
 652         /* encode the 16x16 macroblock */
 653         x264_mb_encode_i16x16( h, i_qp );
 654     }
 655     else if( h->mb.i_type == I_8x8 )
 656     {
 657         ALIGNED_ARRAY_16( uint8_t, edge,[33] );
 658         h->mb.b_transform_8x8 = 1;
 659         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 660         if( h->mb.i_skip_intra )
 661         {
 662             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 663             M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i8x8_nnz_buf[0];
 664             M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i8x8_nnz_buf[1];
 665             M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i8x8_nnz_buf[2];
 666             M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i8x8_nnz_buf[3];
 667             h->mb.i_cbp_luma = h->mb.pic.i8x8_cbp;
 668             /* In RD mode, restore the now-overwritten DCT data. */
 669             if( h->mb.i_skip_intra == 2 )
 670                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 671         }
 672         for( int i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 673         {
 674             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 675             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 676             h->predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 677
 678             if( h->mb.b_lossless )
 679                 x264_predict_lossless_8x8( h, p_dst, i, i_mode, edge );
 680             else
 681                 h->predict_8x8[i_mode]( p_dst, edge );
 682
 683             x264_mb_encode_i8x8( h, i, i_qp );
 684         }
 685     }
 686     else if( h->mb.i_type == I_4x4 )
 687     {
 688         h->mb.b_transform_8x8 = 0;
 689         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 690         if( h->mb.i_skip_intra )
 691         {
 692             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 693             M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i4x4_nnz_buf[0];
 694             M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i4x4_nnz_buf[1];
 695             M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i4x4_nnz_buf[2];
 696             M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i4x4_nnz_buf[3];
 697             h->mb.i_cbp_luma = h->mb.pic.i4x4_cbp;
 698             /* In RD mode, restore the now-overwritten DCT data. */
 699             if( h->mb.i_skip_intra == 2 )
 700                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 701         }
 702         for( int i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 703         {
 704             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
 705             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 706
 707             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 708                 /* emulate missing topright samples */
 709                 M32( &p_dst[4-FDEC_STRIDE] ) = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 710
 711             if( h->mb.b_lossless )
 712                 x264_predict_lossless_4x4( h, p_dst, i, i_mode );
 713             else
 714                 h->predict_4x4[i_mode]( p_dst );
 715             x264_mb_encode_i4x4( h, i, i_qp );
 716         }
 717     }
 718     else    /* Inter MB */
 719     {
 720         int i_decimate_mb = 0;
 721
 722         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
 723         if( !h->mb.b_skip_mc )
 724             x264_mb_mc( h );
 725
 726         if( h->mb.b_lossless )
 727         {
 728             if( h->mb.b_transform_8x8 )
 729                 for( int i8x8 = 0; i8x8 < 4; i8x8++ )
 730                 {
 731                     int x = 8*(i8x8&1);
 732                     int y = 8*(i8x8>>1);
 733                     nz = h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8],
 734                                         h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
 735                                         h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
 736                     STORE_8x8_NNZ(i8x8,nz);
 737                     h->mb.i_cbp_luma |= nz << i8x8;
 738                 }
 739             else
 740                 for( int i4x4 = 0; i4x4 < 16; i4x4++ )
 741                 {
 742                     nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 743                                         h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
 744                                         h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4x4] );
 745                     h->mb.cache.non_zero_count[x264_scan8[i4x4]] = nz;
 746                     h->mb.i_cbp_luma |= nz << (i4x4>>2);
 747                 }
 748         }
 749         else if( h->mb.b_transform_8x8 )
 750         {
 751             ALIGNED_ARRAY_16( int16_t, dct8x8,[4],[64] );
 752             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 753             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 754             h->nr_count[1] += h->mb.b_noise_reduction * 4;
 755
 756             for( int idx = 0; idx < 4; idx++ )
 757             {
 758                 if( h->mb.b_noise_reduction )
 759                     h->quantf.denoise_dct( dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
 760                 nz = x264_quant_8x8( h, dct8x8[idx], i_qp, 0, idx );
 761
 762                 if( nz )
 763                 {
 764                     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 765                     if( b_decimate )
 766                     {
 767                         int i_decimate_8x8 = h->quantf.decimate_score64( h->dct.luma8x8[idx] );
 768                         i_decimate_mb += i_decimate_8x8;
 769                         if( i_decimate_8x8 >= 4 )
 770                             h->mb.i_cbp_luma |= 1<<idx;
 771                     }
 772                     else
 773                         h->mb.i_cbp_luma |= 1<<idx;
 774                 }
 775             }
 776
 777             if( i_decimate_mb < 6 && b_decimate )
 778             {
 779                 h->mb.i_cbp_luma = 0;
 780                 CLEAR_16x16_NNZ
 781             }
 782             else
 783             {
 784                 for( int idx = 0; idx < 4; idx++ )
 785                 {
 786                     if( h->mb.i_cbp_luma&(1<<idx) )
 787                     {
 788                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 789                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 790                         STORE_8x8_NNZ(idx,1);
 791                     }
 792                     else
 793                         STORE_8x8_NNZ(idx,0);
 794                 }
 795             }
 796         }
 797         else
 798         {
 799             ALIGNED_ARRAY_16( int16_t, dct4x4,[16],[16] );
 800             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 801             h->nr_count[0] += h->mb.b_noise_reduction * 16;
 802
 803             for( int i8x8 = 0; i8x8 < 4; i8x8++ )
 804             {
 805                 int i_decimate_8x8 = 0;
 806                 int cbp = 0;
 807
 808                 /* encode one 4x4 block */
 809                 for( int i4x4 = 0; i4x4 < 4; i4x4++ )
 810                 {
 811                     int idx = i8x8 * 4 + i4x4;
 812
 813                     if( h->mb.b_noise_reduction )
 814                         h->quantf.denoise_dct( dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
 815                     nz = x264_quant_4x4( h, dct4x4[idx], i_qp, DCT_LUMA_4x4, 0, idx );
 816                     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 817
 818                     if( nz )
 819                     {
 820                         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 821                         h->quantf.dequant_4x4( dct4x4[idx], h->dequant4_mf[CQM_4PY], i_qp );
 822                         if( b_decimate && i_decimate_8x8 < 6 )
 823                             i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[idx] );
 824                         cbp = 1;
 825                     }
 826                 }
 827
 828                 /* decimate this 8x8 block */
 829                 i_decimate_mb += i_decimate_8x8;
 830                 if( b_decimate )
 831                 {
 832                     if( i_decimate_8x8 < 4 )
 833                         STORE_8x8_NNZ(i8x8,0)
 834                     else
 835                         h->mb.i_cbp_luma |= 1<<i8x8;
 836                 }
 837                 else if( cbp )
 838                 {
 839                     h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 840                     h->mb.i_cbp_luma |= 1<<i8x8;
 841                 }
 842             }
 843
 844             if( b_decimate )
 845             {
 846                 if( i_decimate_mb < 6 )
 847                 {
 848                     h->mb.i_cbp_luma = 0;
 849                     CLEAR_16x16_NNZ
 850                 }
 851                 else
 852                 {
 853                     for( int i8x8 = 0; i8x8 < 4; i8x8++ )
 854                         if( h->mb.i_cbp_luma&(1<<i8x8) )
 855                             h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 856                 }
 857             }
 858         }
 859     }
 860
 861     /* encode chroma */
 862     if( IS_INTRA( h->mb.i_type ) )
 863     {
 864         const int i_mode = h->mb.i_chroma_pred_mode;
 865         if( h->mb.b_lossless )
 866             x264_predict_lossless_8x8_chroma( h, i_mode );
 867         else
 868         {
 869             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 870             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 871         }
 872     }
 873
 874     /* encode the 8x8 blocks */
 875     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 876
 877     /* store cbp */
 878     int cbp = h->mb.i_cbp_chroma << 4 | h->mb.i_cbp_luma;
 879     if( h->param.b_cabac )
 880         cbp |= h->mb.cache.non_zero_count[x264_scan8[24]] << 8
 881             |  h->mb.cache.non_zero_count[x264_scan8[25]] << 9
 882             |  h->mb.cache.non_zero_count[x264_scan8[26]] << 10;
 883     h->mb.cbp[h->mb.i_mb_xy] = cbp;
 884
 885     /* Check for P_SKIP
 886      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 887      *      (if multiple mv give same result)*/
 888     if( !b_force_no_skip )
 889     {
 890         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 891             !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
 892             M32( h->mb.cache.mv[0][x264_scan8[0]] ) == M32( h->mb.cache.pskip_mv )
 893             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 894         {
 895             h->mb.i_type = P_SKIP;
 896         }
 897
 898         /* Check for B_SKIP */
 899         if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
 900         {
 901             h->mb.i_type = B_SKIP;
 902         }
 903     }
 904 }
 905
 906 /*****************************************************************************
 907  * x264_macroblock_probe_skip:
 908  *  Check if the current MB could be encoded as a [PB]_SKIP
 909  *****************************************************************************/
 910 int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
 911 {
 912     ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
 913     ALIGNED_ARRAY_16( int16_t, dct2x2,[4] );
 914     ALIGNED_ARRAY_16( int16_t, dctscan,[16] );
 915     ALIGNED_4( int16_t mvp[2] );
 916
 917     int i_qp = h->mb.i_qp;
 918     int thresh, ssd;
 919
 920     if( !b_bidir )
 921     {
 922         /* Get the MV */
 923         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 924         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 925
 926         /* Motion compensation */
 927         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 928                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 929                        mvp[0], mvp[1], 16, 16, &h->sh.weight[0][0] );
 930     }
 931
 932     for( int i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 933     {
 934         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
 935         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
 936         /* get luma diff */
 937         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
 938                                     h->mb.pic.p_fdec[0] + fdec_offset );
 939         /* encode one 4x4 block */
 940         for( int i4x4 = 0; i4x4 < 4; i4x4++ )
 941         {
 942             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] ) )
 943                 continue;
 944             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 945             i_decimate_mb += h->quantf.decimate_score16( dctscan );
 946             if( i_decimate_mb >= 6 )
 947                 return 0;
 948         }
 949     }
 950
 951     /* encode chroma */
 952     i_qp = h->mb.i_chroma_qp;
 953     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 954
 955     for( int ch = 0; ch < 2; ch++ )
 956     {
 957         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 958         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 959
 960         if( !b_bidir )
 961         {
 962             /* Special case for mv0, which is (of course) very common in P-skip mode. */
 963             if( M32( mvp ) )
 964             {
 965                 h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 966                                  h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 967                                  mvp[0], mvp[1], 8, 8 );
 968             }
 969             else
 970                 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 );
 971
 972             if( h->sh.weight[0][1+ch].weightfn )
 973                 h->sh.weight[0][1+ch].weightfn[8>>2]( h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
 974                                                       h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
 975                                                       &h->sh.weight[0][1+ch], 8 );
 976         }
 977
 978         /* there is almost never a termination during chroma, but we can't avoid the check entirely */
 979         /* so instead we check SSD and skip the actual check if the score is low enough. */
 980         ssd = h->pixf.ssd[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE );
 981         if( ssd < thresh )
 982             continue;
 983
 984         /* The vast majority of chroma checks will terminate during the DC check or the higher
 985          * threshold check, so we can save time by doing a DC-only DCT. */
 986         h->dctf.sub8x8_dct_dc( dct2x2, p_src, p_dst );
 987
 988         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 ) )
 989             return 0;
 990
 991         /* If there wasn't a termination in DC, we can check against a much higher threshold. */
 992         if( ssd < thresh*4 )
 993             continue;
 994
 995         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 996
 997         /* calculate dct coeffs */
 998         for( int i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
 999         {
1000             dct4x4[i4x4][0] = 0;
1001             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] ) )
1002                 continue;
1003             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
1004             i_decimate_mb += h->quantf.decimate_score15( dctscan );
1005             if( i_decimate_mb >= 7 )
1006                 return 0;
1007         }
1008     }
1009
1010     h->mb.b_skip_mc = 1;
1011     return 1;
1012 }
1013
1014 /****************************************************************************
1015  * DCT-domain noise reduction / adaptive deadzone
1016  * from libavcodec
1017  ****************************************************************************/
1018
1019 void x264_noise_reduction_update( x264_t *h )
1020 {
1021     for( int cat = 0; cat < 2; cat++ )
1022     {
1023         int size = cat ? 64 : 16;
1024         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
1025
1026         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
1027         {
1028             for( int i = 0; i < size; i++ )
1029                 h->nr_residual_sum[cat][i] >>= 1;
1030             h->nr_count[cat] >>= 1;
1031         }
1032
1033         for( int i = 0; i < size; i++ )
1034             h->nr_offset[cat][i] =
1035                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
1036                  + h->nr_residual_sum[cat][i]/2)
1037               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
1038     }
1039 }
1040
1041 /*****************************************************************************
1042  * RD only; 4 calls to this do not make up for one macroblock_encode.
1043  * doesn't transform chroma dc.
1044  *****************************************************************************/
1045 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
1046 {
1047     int i_qp = h->mb.i_qp;
1048     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
1049     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
1050     int b_decimate = h->mb.b_dct_decimate;
1051     int nnz8x8 = 0;
1052     int nz;
1053
1054     if( !h->mb.b_skip_mc )
1055         x264_mb_mc_8x8( h, i8 );
1056
1057     if( h->mb.b_lossless )
1058     {
1059         if( h->mb.b_transform_8x8 )
1060         {
1061             nnz8x8 = h->zigzagf.sub_8x8( h->dct.luma8x8[i8], p_fenc, p_fdec );
1062             STORE_8x8_NNZ(i8,nnz8x8);
1063         }
1064         else
1065         {
1066             for( int i4 = i8*4; i4 < i8*4+4; i4++ )
1067             {
1068                 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
1069                                     h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
1070                                     h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
1071                 h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1072                 nnz8x8 |= nz;
1073             }
1074         }
1075         for( int ch = 0; ch < 2; ch++ )
1076         {
1077             int16_t dc;
1078             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1079             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1080             nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec, &dc );
1081             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1082         }
1083     }
1084     else
1085     {
1086         if( h->mb.b_transform_8x8 )
1087         {
1088             ALIGNED_ARRAY_16( int16_t, dct8x8,[64] );
1089             h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
1090             nnz8x8 = x264_quant_8x8( h, dct8x8, i_qp, 0, i8 );
1091             if( nnz8x8 )
1092             {
1093                 h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
1094
1095                 if( b_decimate && !h->mb.b_trellis )
1096                     nnz8x8 = 4 <= h->quantf.decimate_score64( h->dct.luma8x8[i8] );
1097
1098                 if( nnz8x8 )
1099                 {
1100                     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
1101                     h->dctf.add8x8_idct8( p_fdec, dct8x8 );
1102                     STORE_8x8_NNZ(i8,1);
1103                 }
1104                 else
1105                     STORE_8x8_NNZ(i8,0);
1106             }
1107             else
1108                 STORE_8x8_NNZ(i8,0);
1109         }
1110         else
1111         {
1112             int i_decimate_8x8 = 0;
1113             ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
1114             h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
1115             for( int i4 = 0; i4 < 4; i4++ )
1116             {
1117                 nz = x264_quant_4x4( h, dct4x4[i4], i_qp, DCT_LUMA_4x4, 0, i8*4+i4 );
1118                 h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = nz;
1119                 if( nz )
1120                 {
1121                     h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
1122                     h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
1123                     if( b_decimate )
1124                         i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[i8*4+i4] );
1125                     nnz8x8 = 1;
1126                 }
1127             }
1128
1129             if( b_decimate && i_decimate_8x8 < 4 )
1130                 nnz8x8 = 0;
1131
1132             if( nnz8x8 )
1133                 h->dctf.add8x8_idct( p_fdec, dct4x4 );
1134             else
1135                 STORE_8x8_NNZ(i8,0);
1136         }
1137
1138         i_qp = h->mb.i_chroma_qp;
1139
1140         for( int ch = 0; ch < 2; ch++ )
1141         {
1142             ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
1143             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1144             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1145
1146             h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1147             dct4x4[0] = 0;
1148
1149             if( h->mb.b_trellis )
1150                 nz = x264_quant_4x4_trellis( h, dct4x4, CQM_4PC, i_qp, DCT_CHROMA_AC, 0, 1, 0 );
1151             else
1152                 nz = h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
1153
1154             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1155             if( nz )
1156             {
1157                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
1158                 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
1159                 h->dctf.add4x4_idct( p_fdec, dct4x4 );
1160             }
1161         }
1162     }
1163     h->mb.i_cbp_luma &= ~(1 << i8);
1164     h->mb.i_cbp_luma |= nnz8x8 << i8;
1165     h->mb.i_cbp_chroma = 0x02;
1166 }
1167
1168 /*****************************************************************************
1169  * RD only, luma only
1170  *****************************************************************************/
1171 void x264_macroblock_encode_p4x4( x264_t *h, int i4 )
1172 {
1173     int i_qp = h->mb.i_qp;
1174     uint8_t *p_fenc = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[i4]];
1175     uint8_t *p_fdec = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
1176     int nz;
1177
1178     /* Don't need motion compensation as this function is only used in qpel-RD, which caches pixel data. */
1179
1180     if( h->mb.b_lossless )
1181     {
1182         nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4], p_fenc, p_fdec );
1183         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1184     }
1185     else
1186     {
1187         ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
1188         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1189         nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 0, i4 );
1190         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1191         if( nz )
1192         {
1193             h->zigzagf.scan_4x4( h->dct.luma4x4[i4], dct4x4 );
1194             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PY], i_qp );
1195             h->dctf.add4x4_idct( p_fdec, dct4x4 );
1196         }
1197     }
1198 }