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]; \
  46     int qbits = i_qp/6 - 5; \
  47     if( qbits > 0 ) \
  48     { \
  49         dmf <<= qbits; \
  50         qbits = 0; \
  51     }
  52
  53 static inline void idct_dequant_2x2_dc( int16_t dct[4], int16_t dct4x4[4][16], int dequant_mf[6][16], int i_qp )
  54 {
  55     IDCT_DEQUANT_START
  56     dct4x4[0][0] = (d0 + d1) * dmf >> -qbits;
  57     dct4x4[1][0] = (d0 - d1) * dmf >> -qbits;
  58     dct4x4[2][0] = (d2 + d3) * dmf >> -qbits;
  59     dct4x4[3][0] = (d2 - d3) * dmf >> -qbits;
  60 }
  61
  62 static inline void idct_dequant_2x2_dconly( int16_t out[4], int16_t dct[4], int dequant_mf[6][16], int i_qp )
  63 {
  64     IDCT_DEQUANT_START
  65     out[0] = (d0 + d1) * dmf >> -qbits;
  66     out[1] = (d0 - d1) * dmf >> -qbits;
  67     out[2] = (d2 + d3) * dmf >> -qbits;
  68     out[3] = (d2 - d3) * dmf >> -qbits;
  69 }
  70
  71 static inline void dct2x2dc( int16_t d[4], int16_t dct4x4[4][16] )
  72 {
  73     int d0 = dct4x4[0][0] + dct4x4[1][0];
  74     int d1 = dct4x4[2][0] + dct4x4[3][0];
  75     int d2 = dct4x4[0][0] - dct4x4[1][0];
  76     int d3 = dct4x4[2][0] - dct4x4[3][0];
  77     d[0] = d0 + d1;
  78     d[2] = d2 + d3;
  79     d[1] = d0 - d1;
  80     d[3] = d2 - d3;
  81     dct4x4[0][0] = 0;
  82     dct4x4[1][0] = 0;
  83     dct4x4[2][0] = 0;
  84     dct4x4[3][0] = 0;
  85 }
  86
  87 static inline void dct2x2dc_dconly( int16_t d[4] )
  88 {
  89     int d0 = d[0] + d[1];
  90     int d1 = d[2] + d[3];
  91     int d2 = d[0] - d[1];
  92     int d3 = d[2] - d[3];
  93     d[0] = d0 + d1;
  94     d[2] = d2 + d3;
  95     d[1] = d0 - d1;
  96     d[3] = d2 - d3;
  97 }
  98
  99 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 )
 100 {
 101     int i_quant_cat = b_intra ? CQM_4IY : CQM_4PY;
 102     if( h->mb.b_trellis )
 103         return x264_quant_4x4_trellis( h, dct, i_quant_cat, i_qp, i_ctxBlockCat, b_intra, 0, idx );
 104     else
 105         return h->quantf.quant_4x4( dct, h->quant4_mf[i_quant_cat][i_qp], h->quant4_bias[i_quant_cat][i_qp] );
 106 }
 107
 108 static ALWAYS_INLINE int x264_quant_8x8( x264_t *h, int16_t dct[64], int i_qp, int b_intra, int idx )
 109 {
 110     int i_quant_cat = b_intra ? CQM_8IY : CQM_8PY;
 111     if( h->mb.b_trellis )
 112         return x264_quant_8x8_trellis( h, dct, i_quant_cat, i_qp, b_intra, idx );
 113     else
 114         return h->quantf.quant_8x8( dct, h->quant8_mf[i_quant_cat][i_qp], h->quant8_bias[i_quant_cat][i_qp] );
 115 }
 116
 117 /* All encoding functions must output the correct CBP and NNZ values.
 118  * The entropy coding functions will check CBP first, then NNZ, before
 119  * actually reading the DCT coefficients.  NNZ still must be correct even
 120  * if CBP is zero because of the use of NNZ values for context selection.
 121  * "NNZ" need only be 0 or 1 rather than the exact coefficient count because
 122  * that is only needed in CAVLC, and will be calculated by CAVLC's residual
 123  * coding and stored as necessary. */
 124
 125 /* This means that decimation can be done merely by adjusting the CBP and NNZ
 126  * rather than memsetting the coefficients. */
 127
 128 void x264_mb_encode_i4x4( x264_t *h, int idx, int i_qp )
 129 {
 130     int nz;
 131     uint8_t *p_src = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[idx]];
 132     uint8_t *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[idx]];
 133     ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
 134
 135     if( h->mb.b_lossless )
 136     {
 137         nz = h->zigzagf.sub_4x4( h->dct.luma4x4[idx], p_src, p_dst );
 138         h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 139         h->mb.i_cbp_luma |= nz<<(idx>>2);
 140         return;
 141     }
 142
 143     h->dctf.sub4x4_dct( dct4x4, p_src, p_dst );
 144
 145     nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 1, idx );
 146     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 147     if( nz )
 148     {
 149         h->mb.i_cbp_luma |= 1<<(idx>>2);
 150         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4 );
 151         h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4IY], i_qp );
 152         h->dctf.add4x4_idct( p_dst, dct4x4 );
 153     }
 154 }
 155
 156 #define STORE_8x8_NNZ(idx,nz)\
 157 {\
 158     M16( &h->mb.cache.non_zero_count[x264_scan8[idx*4+0]] ) = nz * 0x0101;\
 159     M16( &h->mb.cache.non_zero_count[x264_scan8[idx*4+2]] ) = nz * 0x0101;\
 160 }
 161
 162 #define CLEAR_16x16_NNZ \
 163 {\
 164     M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = 0;\
 165     M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = 0;\
 166     M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = 0;\
 167     M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = 0;\
 168 }
 169
 170 void x264_mb_encode_i8x8( x264_t *h, int idx, int i_qp )
 171 {
 172     int x = 8 * (idx&1);
 173     int y = 8 * (idx>>1);
 174     int nz;
 175     uint8_t *p_src = &h->mb.pic.p_fenc[0][x+y*FENC_STRIDE];
 176     uint8_t *p_dst = &h->mb.pic.p_fdec[0][x+y*FDEC_STRIDE];
 177     ALIGNED_ARRAY_16( int16_t, dct8x8,[64] );
 178
 179     if( h->mb.b_lossless )
 180     {
 181         nz = h->zigzagf.sub_8x8( h->dct.luma8x8[idx], p_src, p_dst );
 182         STORE_8x8_NNZ(idx,nz);
 183         h->mb.i_cbp_luma |= nz<<idx;
 184         return;
 185     }
 186
 187     h->dctf.sub8x8_dct8( dct8x8, p_src, p_dst );
 188
 189     nz = x264_quant_8x8( h, dct8x8, i_qp, 1, idx );
 190     if( nz )
 191     {
 192         h->mb.i_cbp_luma |= 1<<idx;
 193         h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8 );
 194         h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8IY], i_qp );
 195         h->dctf.add8x8_idct8( p_dst, dct8x8 );
 196         STORE_8x8_NNZ(idx,1);
 197     }
 198     else
 199         STORE_8x8_NNZ(idx,0);
 200 }
 201
 202 static void x264_mb_encode_i16x16( x264_t *h, int i_qp )
 203 {
 204     uint8_t  *p_src = h->mb.pic.p_fenc[0];
 205     uint8_t  *p_dst = h->mb.pic.p_fdec[0];
 206
 207     ALIGNED_ARRAY_16( int16_t, dct4x4,[16],[16] );
 208     ALIGNED_ARRAY_16( int16_t, dct_dc4x4,[16] );
 209
 210     int i, nz;
 211     int b_decimate = h->sh.i_type == SLICE_TYPE_B || (h->param.analyse.b_dct_decimate && h->sh.i_type == SLICE_TYPE_P);
 212     int decimate_score = b_decimate ? 0 : 9;
 213
 214     if( h->mb.b_lossless )
 215     {
 216         for( i = 0; i < 16; i++ )
 217         {
 218             int oe = block_idx_xy_fenc[i];
 219             int od = block_idx_xy_fdec[i];
 220             nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[i], p_src+oe, p_dst+od, &dct_dc4x4[block_idx_yx_1d[i]] );
 221             h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 222             h->mb.i_cbp_luma |= nz;
 223         }
 224         h->mb.i_cbp_luma *= 0xf;
 225         h->mb.cache.non_zero_count[x264_scan8[24]] = array_non_zero( dct_dc4x4 );
 226         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 227         return;
 228     }
 229
 230     h->dctf.sub16x16_dct( dct4x4, p_src, p_dst );
 231
 232     for( i = 0; i < 16; i++ )
 233     {
 234         /* copy dc coeff */
 235         dct_dc4x4[block_idx_xy_1d[i]] = dct4x4[i][0];
 236         dct4x4[i][0] = 0;
 237
 238         /* quant/scan/dequant */
 239         nz = x264_quant_4x4( h, dct4x4[i], i_qp, DCT_LUMA_AC, 1, i );
 240         h->mb.cache.non_zero_count[x264_scan8[i]] = nz;
 241         if( nz )
 242         {
 243             h->zigzagf.scan_4x4( h->dct.luma4x4[i], dct4x4[i] );
 244             h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IY], i_qp );
 245             if( decimate_score < 6 ) decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[i] );
 246             h->mb.i_cbp_luma = 0xf;
 247         }
 248     }
 249
 250     /* Writing the 16 CBFs in an i16x16 block is quite costly, so decimation can save many bits. */
 251     /* More useful with CAVLC, but still useful with CABAC. */
 252     if( decimate_score < 6 )
 253     {
 254         h->mb.i_cbp_luma = 0;
 255         CLEAR_16x16_NNZ
 256     }
 257
 258     h->dctf.dct4x4dc( dct_dc4x4 );
 259     if( h->mb.b_trellis )
 260         nz = x264_quant_dc_trellis( h, dct_dc4x4, CQM_4IY, i_qp, DCT_LUMA_DC, 1, 0 );
 261     else
 262         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 );
 263
 264     h->mb.cache.non_zero_count[x264_scan8[24]] = nz;
 265     if( nz )
 266     {
 267         h->zigzagf.scan_4x4( h->dct.luma16x16_dc, dct_dc4x4 );
 268
 269         /* output samples to fdec */
 270         h->dctf.idct4x4dc( dct_dc4x4 );
 271         h->quantf.dequant_4x4_dc( dct_dc4x4, h->dequant4_mf[CQM_4IY], i_qp );  /* XXX not inversed */
 272         if( h->mb.i_cbp_luma )
 273             for( i = 0; i < 16; i++ )
 274                 dct4x4[i][0] = dct_dc4x4[block_idx_xy_1d[i]];
 275     }
 276
 277     /* put pixels to fdec */
 278     if( h->mb.i_cbp_luma )
 279         h->dctf.add16x16_idct( p_dst, dct4x4 );
 280     else if( nz )
 281         h->dctf.add16x16_idct_dc( p_dst, dct_dc4x4 );
 282 }
 283
 284 static inline int idct_dequant_round_2x2_dc( int16_t ref[4], int16_t dct[4], int dequant_mf[6][16], int i_qp )
 285 {
 286     int16_t out[4];
 287     idct_dequant_2x2_dconly( out, dct, dequant_mf, i_qp );
 288     return ((ref[0] ^ (out[0]+32))
 289           | (ref[1] ^ (out[1]+32))
 290           | (ref[2] ^ (out[2]+32))
 291           | (ref[3] ^ (out[3]+32))) >> 6;
 292 }
 293
 294 /* Round down coefficients losslessly in DC-only chroma blocks.
 295  * Unlike luma blocks, this can't be done with a lookup table or
 296  * other shortcut technique because of the interdependencies
 297  * between the coefficients due to the chroma DC transform. */
 298 static inline int x264_mb_optimize_chroma_dc( x264_t *h, int b_inter, int i_qp, int16_t dct2x2[4] )
 299 {
 300     int16_t dct2x2_orig[4];
 301     int coeff;
 302     int nz = 0;
 303
 304     /* If the QP is too high, there's no benefit to rounding optimization. */
 305     if( h->dequant4_mf[CQM_4IC + b_inter][i_qp%6][0] << (i_qp/6) > 32*64 )
 306         return 1;
 307
 308     idct_dequant_2x2_dconly( dct2x2_orig, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 309     dct2x2_orig[0] += 32;
 310     dct2x2_orig[1] += 32;
 311     dct2x2_orig[2] += 32;
 312     dct2x2_orig[3] += 32;
 313
 314     /* If the DC coefficients already round to zero, terminate early. */
 315     if( !((dct2x2_orig[0]|dct2x2_orig[1]|dct2x2_orig[2]|dct2x2_orig[3])>>6) )
 316         return 0;
 317
 318     /* Start with the highest frequency coefficient... is this the best option? */
 319     for( coeff = 3; coeff >= 0; coeff-- )
 320     {
 321         int sign = dct2x2[coeff] < 0 ? -1 : 1;
 322         int level = dct2x2[coeff];
 323
 324         if( !level )
 325             continue;
 326
 327         while( level )
 328         {
 329             dct2x2[coeff] = level - sign;
 330             if( idct_dequant_round_2x2_dc( dct2x2_orig, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp ) )
 331                 break;
 332             level -= sign;
 333         }
 334
 335         nz |= level;
 336         dct2x2[coeff] = level;
 337     }
 338
 339     return !!nz;
 340 }
 341
 342 void x264_mb_encode_8x8_chroma( x264_t *h, int b_inter, int i_qp )
 343 {
 344     int i, ch, nz, nz_dc;
 345     int b_decimate = b_inter && (h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate);
 346     ALIGNED_ARRAY_16( int16_t, dct2x2,[4] );
 347     h->mb.i_cbp_chroma = 0;
 348
 349     /* Early termination: check variance of chroma residual before encoding.
 350      * Don't bother trying early termination at low QPs.
 351      * Values are experimentally derived. */
 352     if( b_decimate && i_qp >= (h->mb.b_trellis ? 12 : 18) )
 353     {
 354         int thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 355         int ssd[2];
 356         int score  = h->pixf.var2_8x8( h->mb.pic.p_fenc[1], FENC_STRIDE, h->mb.pic.p_fdec[1], FDEC_STRIDE, &ssd[0] );
 357             score += h->pixf.var2_8x8( h->mb.pic.p_fenc[2], FENC_STRIDE, h->mb.pic.p_fdec[2], FDEC_STRIDE, &ssd[1] );
 358         if( score < thresh*4 )
 359         {
 360             h->mb.cache.non_zero_count[x264_scan8[16]] = 0;
 361             h->mb.cache.non_zero_count[x264_scan8[17]] = 0;
 362             h->mb.cache.non_zero_count[x264_scan8[18]] = 0;
 363             h->mb.cache.non_zero_count[x264_scan8[19]] = 0;
 364             h->mb.cache.non_zero_count[x264_scan8[20]] = 0;
 365             h->mb.cache.non_zero_count[x264_scan8[21]] = 0;
 366             h->mb.cache.non_zero_count[x264_scan8[22]] = 0;
 367             h->mb.cache.non_zero_count[x264_scan8[23]] = 0;
 368             h->mb.cache.non_zero_count[x264_scan8[25]] = 0;
 369             h->mb.cache.non_zero_count[x264_scan8[26]] = 0;
 370             for( ch = 0; ch < 2; ch++ )
 371             {
 372                 if( ssd[ch] > thresh )
 373                 {
 374                     h->dctf.sub8x8_dct_dc( dct2x2, h->mb.pic.p_fenc[1+ch], h->mb.pic.p_fdec[1+ch] );
 375                     dct2x2dc_dconly( dct2x2 );
 376                     if( h->mb.b_trellis )
 377                         nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 378                     else
 379                         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]<<
 380     1 );
 381
 382                     if( nz_dc )
 383                     {
 384                         if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
 385                             continue;
 386                         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 1;
 387                         zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 388                         idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 389                         h->dctf.add8x8_idct_dc( h->mb.pic.p_fdec[1+ch], dct2x2 );
 390                         h->mb.i_cbp_chroma = 1;
 391                     }
 392                 }
 393             }
 394             return;
 395         }
 396     }
 397
 398     for( ch = 0; ch < 2; ch++ )
 399     {
 400         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 401         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 402         int i_decimate_score = 0;
 403         int nz_ac = 0;
 404
 405         ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
 406
 407         if( h->mb.b_lossless )
 408         {
 409             for( i = 0; i < 4; i++ )
 410             {
 411                 int oe = block_idx_x[i]*4 + block_idx_y[i]*4*FENC_STRIDE;
 412                 int od = block_idx_x[i]*4 + block_idx_y[i]*4*FDEC_STRIDE;
 413                 nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i+ch*4], p_src+oe, p_dst+od, &h->dct.chroma_dc[ch][i] );
 414                 h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 415                 h->mb.i_cbp_chroma |= nz;
 416             }
 417             h->mb.cache.non_zero_count[x264_scan8[25]+ch] = array_non_zero( h->dct.chroma_dc[ch] );
 418             continue;
 419         }
 420
 421         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 422         dct2x2dc( dct2x2, dct4x4 );
 423         /* calculate dct coeffs */
 424         for( i = 0; i < 4; i++ )
 425         {
 426             if( h->mb.b_trellis )
 427                 nz = x264_quant_4x4_trellis( h, dct4x4[i], CQM_4IC+b_inter, i_qp, DCT_CHROMA_AC, !b_inter, 1, 0 );
 428             else
 429                 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] );
 430             h->mb.cache.non_zero_count[x264_scan8[16+i+ch*4]] = nz;
 431             if( nz )
 432             {
 433                 nz_ac = 1;
 434                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i+ch*4], dct4x4[i] );
 435                 h->quantf.dequant_4x4( dct4x4[i], h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 436                 if( b_decimate )
 437                     i_decimate_score += h->quantf.decimate_score15( h->dct.luma4x4[16+i+ch*4] );
 438             }
 439         }
 440
 441         if( h->mb.b_trellis )
 442             nz_dc = x264_quant_dc_trellis( h, dct2x2, CQM_4IC+b_inter, i_qp, DCT_CHROMA_DC, !b_inter, 1 );
 443         else
 444             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 );
 445
 446         h->mb.cache.non_zero_count[x264_scan8[25]+ch] = nz_dc;
 447
 448         if( (b_decimate && i_decimate_score < 7) || !nz_ac )
 449         {
 450             /* Decimate the block */
 451             h->mb.cache.non_zero_count[x264_scan8[16+0]+24*ch] = 0;
 452             h->mb.cache.non_zero_count[x264_scan8[16+1]+24*ch] = 0;
 453             h->mb.cache.non_zero_count[x264_scan8[16+2]+24*ch] = 0;
 454             h->mb.cache.non_zero_count[x264_scan8[16+3]+24*ch] = 0;
 455             if( !nz_dc ) /* Whole block is empty */
 456                 continue;
 457             if( !x264_mb_optimize_chroma_dc( h, b_inter, i_qp, dct2x2 ) )
 458             {
 459                 h->mb.cache.non_zero_count[x264_scan8[25]+ch] = 0;
 460                 continue;
 461             }
 462             /* DC-only */
 463             zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 464             idct_dequant_2x2_dconly( dct2x2, dct2x2, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 465             h->dctf.add8x8_idct_dc( p_dst, dct2x2 );
 466         }
 467         else
 468         {
 469             h->mb.i_cbp_chroma = 1;
 470             if( nz_dc )
 471             {
 472                 zigzag_scan_2x2_dc( h->dct.chroma_dc[ch], dct2x2 );
 473                 idct_dequant_2x2_dc( dct2x2, dct4x4, h->dequant4_mf[CQM_4IC + b_inter], i_qp );
 474             }
 475             h->dctf.add8x8_idct( p_dst, dct4x4 );
 476         }
 477     }
 478
 479     if( h->mb.i_cbp_chroma )
 480         h->mb.i_cbp_chroma = 2;    /* dc+ac (we can't do only ac) */
 481     else if( h->mb.cache.non_zero_count[x264_scan8[25]] |
 482              h->mb.cache.non_zero_count[x264_scan8[26]] )
 483         h->mb.i_cbp_chroma = 1;    /* dc only */
 484 }
 485
 486 static void x264_macroblock_encode_skip( x264_t *h )
 487 {
 488     h->mb.i_cbp_luma = 0x00;
 489     h->mb.i_cbp_chroma = 0x00;
 490     memset( h->mb.cache.non_zero_count, 0, X264_SCAN8_SIZE );
 491     /* store cbp */
 492     h->mb.cbp[h->mb.i_mb_xy] = 0;
 493 }
 494
 495 /*****************************************************************************
 496  * x264_macroblock_encode_pskip:
 497  *  Encode an already marked skip block
 498  *****************************************************************************/
 499 static void x264_macroblock_encode_pskip( x264_t *h )
 500 {
 501     const int mvx = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][0],
 502                                 h->mb.mv_min[0], h->mb.mv_max[0] );
 503     const int mvy = x264_clip3( h->mb.cache.mv[0][x264_scan8[0]][1],
 504                                 h->mb.mv_min[1], h->mb.mv_max[1] );
 505
 506     /* don't do pskip motion compensation if it was already done in macroblock_analyse */
 507     if( !h->mb.b_skip_mc )
 508     {
 509         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 510                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 511                        mvx, mvy, 16, 16, &h->sh.weight[0][0] );
 512
 513         h->mc.mc_chroma( h->mb.pic.p_fdec[1],       FDEC_STRIDE,
 514                          h->mb.pic.p_fref[0][0][4], h->mb.pic.i_stride[1],
 515                          mvx, mvy, 8, 8 );
 516
 517         if( h->sh.weight[0][1].weightfn )
 518             h->sh.weight[0][1].weightfn[8>>2]( h->mb.pic.p_fdec[1], FDEC_STRIDE,
 519                                                h->mb.pic.p_fdec[1], FDEC_STRIDE,
 520                                                &h->sh.weight[0][1], 8 );
 521
 522         h->mc.mc_chroma( h->mb.pic.p_fdec[2],       FDEC_STRIDE,
 523                          h->mb.pic.p_fref[0][0][5], h->mb.pic.i_stride[2],
 524                          mvx, mvy, 8, 8 );
 525
 526         if( h->sh.weight[0][2].weightfn )
 527             h->sh.weight[0][2].weightfn[8>>2]( h->mb.pic.p_fdec[2], FDEC_STRIDE,
 528                                                h->mb.pic.p_fdec[2], FDEC_STRIDE,
 529                                                &h->sh.weight[0][2], 8 );
 530     }
 531
 532     x264_macroblock_encode_skip( h );
 533 }
 534
 535 /*****************************************************************************
 536  * Intra prediction for predictive lossless mode.
 537  *****************************************************************************/
 538
 539 /* Note that these functions take a shortcut (mc.copy instead of actual pixel prediction) which assumes
 540  * that the edge pixels of the reconstructed frame are the same as that of the source frame.  This means
 541  * they will only work correctly if the neighboring blocks are losslessly coded.  In practice, this means
 542  * lossless mode cannot be mixed with lossy mode within a frame. */
 543 /* This can be resolved by explicitly copying the edge pixels after doing the mc.copy, but this doesn't
 544  * need to be done unless we decide to allow mixing lossless and lossy compression. */
 545
 546 void x264_predict_lossless_8x8_chroma( x264_t *h, int i_mode )
 547 {
 548     int stride = h->fenc->i_stride[1] << h->mb.b_interlaced;
 549     if( i_mode == I_PRED_CHROMA_V )
 550     {
 551         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-stride, stride, 8 );
 552         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-stride, stride, 8 );
 553     }
 554     else if( i_mode == I_PRED_CHROMA_H )
 555     {
 556         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[1], FDEC_STRIDE, h->mb.pic.p_fenc_plane[1]-1, stride, 8 );
 557         h->mc.copy[PIXEL_8x8]( h->mb.pic.p_fdec[2], FDEC_STRIDE, h->mb.pic.p_fenc_plane[2]-1, stride, 8 );
 558     }
 559     else
 560     {
 561         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 562         h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 563     }
 564 }
 565
 566 void x264_predict_lossless_4x4( x264_t *h, uint8_t *p_dst, int idx, int i_mode )
 567 {
 568     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 569     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + block_idx_x[idx]*4 + block_idx_y[idx]*4 * stride;
 570
 571     if( i_mode == I_PRED_4x4_V )
 572         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-stride, stride, 4 );
 573     else if( i_mode == I_PRED_4x4_H )
 574         h->mc.copy[PIXEL_4x4]( p_dst, FDEC_STRIDE, p_src-1, stride, 4 );
 575     else
 576         h->predict_4x4[i_mode]( p_dst );
 577 }
 578
 579 void x264_predict_lossless_8x8( x264_t *h, uint8_t *p_dst, int idx, int i_mode, uint8_t edge[33] )
 580 {
 581     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 582     uint8_t *p_src = h->mb.pic.p_fenc_plane[0] + (idx&1)*8 + (idx>>1)*8*stride;
 583
 584     if( i_mode == I_PRED_8x8_V )
 585         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-stride, stride, 8 );
 586     else if( i_mode == I_PRED_8x8_H )
 587         h->mc.copy[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src-1, stride, 8 );
 588     else
 589         h->predict_8x8[i_mode]( p_dst, edge );
 590 }
 591
 592 void x264_predict_lossless_16x16( x264_t *h, int i_mode )
 593 {
 594     int stride = h->fenc->i_stride[0] << h->mb.b_interlaced;
 595     if( i_mode == I_PRED_16x16_V )
 596         h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-stride, stride, 16 );
 597     else if( i_mode == I_PRED_16x16_H )
 598         h->mc.copy_16x16_unaligned( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.p_fenc_plane[0]-1, stride, 16 );
 599     else
 600         h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 601 }
 602
 603 /*****************************************************************************
 604  * x264_macroblock_encode:
 605  *****************************************************************************/
 606 void x264_macroblock_encode( x264_t *h )
 607 {
 608     int i_cbp_dc = 0;
 609     int i_qp = h->mb.i_qp;
 610     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
 611     int b_force_no_skip = 0;
 612     int i,idx,nz;
 613     h->mb.i_cbp_luma = 0;
 614     h->mb.cache.non_zero_count[x264_scan8[24]] = 0;
 615
 616     if( h->sh.b_mbaff
 617         && h->mb.i_mb_xy == h->sh.i_first_mb + h->mb.i_mb_stride
 618         && IS_SKIP(h->mb.type[h->sh.i_first_mb]) )
 619     {
 620         /* The first skip is predicted to be a frame mb pair.
 621          * We don't yet support the aff part of mbaff, so force it to non-skip
 622          * so that we can pick the aff flag. */
 623         b_force_no_skip = 1;
 624         if( IS_SKIP(h->mb.i_type) )
 625         {
 626             if( h->mb.i_type == P_SKIP )
 627                 h->mb.i_type = P_L0;
 628             else if( h->mb.i_type == B_SKIP )
 629                 h->mb.i_type = B_DIRECT;
 630         }
 631     }
 632
 633     if( h->mb.i_type == P_SKIP )
 634     {
 635         /* A bit special */
 636         x264_macroblock_encode_pskip( h );
 637         return;
 638     }
 639     if( h->mb.i_type == B_SKIP )
 640     {
 641         /* don't do bskip motion compensation if it was already done in macroblock_analyse */
 642         if( !h->mb.b_skip_mc )
 643             x264_mb_mc( h );
 644         x264_macroblock_encode_skip( h );
 645         return;
 646     }
 647
 648     if( h->mb.i_type == I_16x16 )
 649     {
 650         const int i_mode = h->mb.i_intra16x16_pred_mode;
 651         h->mb.b_transform_8x8 = 0;
 652
 653         if( h->mb.b_lossless )
 654             x264_predict_lossless_16x16( h, i_mode );
 655         else
 656             h->predict_16x16[i_mode]( h->mb.pic.p_fdec[0] );
 657
 658         /* encode the 16x16 macroblock */
 659         x264_mb_encode_i16x16( h, i_qp );
 660     }
 661     else if( h->mb.i_type == I_8x8 )
 662     {
 663         ALIGNED_ARRAY_16( uint8_t, edge,[33] );
 664         h->mb.b_transform_8x8 = 1;
 665         /* If we already encoded 3 of the 4 i8x8 blocks, we don't have to do them again. */
 666         if( h->mb.i_skip_intra )
 667         {
 668             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i8x8_fdec_buf, 16, 16 );
 669             M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i8x8_nnz_buf[0];
 670             M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i8x8_nnz_buf[1];
 671             M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i8x8_nnz_buf[2];
 672             M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i8x8_nnz_buf[3];
 673             h->mb.i_cbp_luma = h->mb.pic.i8x8_cbp;
 674             /* In RD mode, restore the now-overwritten DCT data. */
 675             if( h->mb.i_skip_intra == 2 )
 676                 h->mc.memcpy_aligned( h->dct.luma8x8, h->mb.pic.i8x8_dct_buf, sizeof(h->mb.pic.i8x8_dct_buf) );
 677         }
 678         for( i = h->mb.i_skip_intra ? 3 : 0 ; i < 4; i++ )
 679         {
 680             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][8 * (i&1) + 8 * (i>>1) * FDEC_STRIDE];
 681             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[4*i]];
 682             h->predict_8x8_filter( p_dst, edge, h->mb.i_neighbour8[i], x264_pred_i4x4_neighbors[i_mode] );
 683
 684             if( h->mb.b_lossless )
 685                 x264_predict_lossless_8x8( h, p_dst, i, i_mode, edge );
 686             else
 687                 h->predict_8x8[i_mode]( p_dst, edge );
 688
 689             x264_mb_encode_i8x8( h, i, i_qp );
 690         }
 691     }
 692     else if( h->mb.i_type == I_4x4 )
 693     {
 694         h->mb.b_transform_8x8 = 0;
 695         /* If we already encoded 15 of the 16 i4x4 blocks, we don't have to do them again. */
 696         if( h->mb.i_skip_intra )
 697         {
 698             h->mc.copy[PIXEL_16x16]( h->mb.pic.p_fdec[0], FDEC_STRIDE, h->mb.pic.i4x4_fdec_buf, 16, 16 );
 699             M32( &h->mb.cache.non_zero_count[x264_scan8[ 0]] ) = h->mb.pic.i4x4_nnz_buf[0];
 700             M32( &h->mb.cache.non_zero_count[x264_scan8[ 2]] ) = h->mb.pic.i4x4_nnz_buf[1];
 701             M32( &h->mb.cache.non_zero_count[x264_scan8[ 8]] ) = h->mb.pic.i4x4_nnz_buf[2];
 702             M32( &h->mb.cache.non_zero_count[x264_scan8[10]] ) = h->mb.pic.i4x4_nnz_buf[3];
 703             h->mb.i_cbp_luma = h->mb.pic.i4x4_cbp;
 704             /* In RD mode, restore the now-overwritten DCT data. */
 705             if( h->mb.i_skip_intra == 2 )
 706                 h->mc.memcpy_aligned( h->dct.luma4x4, h->mb.pic.i4x4_dct_buf, sizeof(h->mb.pic.i4x4_dct_buf) );
 707         }
 708         for( i = h->mb.i_skip_intra ? 15 : 0 ; i < 16; i++ )
 709         {
 710             uint8_t  *p_dst = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i]];
 711             int      i_mode = h->mb.cache.intra4x4_pred_mode[x264_scan8[i]];
 712
 713             if( (h->mb.i_neighbour4[i] & (MB_TOPRIGHT|MB_TOP)) == MB_TOP )
 714                 /* emulate missing topright samples */
 715                 M32( &p_dst[4-FDEC_STRIDE] ) = p_dst[3-FDEC_STRIDE] * 0x01010101U;
 716
 717             if( h->mb.b_lossless )
 718                 x264_predict_lossless_4x4( h, p_dst, i, i_mode );
 719             else
 720                 h->predict_4x4[i_mode]( p_dst );
 721             x264_mb_encode_i4x4( h, i, i_qp );
 722         }
 723     }
 724     else    /* Inter MB */
 725     {
 726         int i8x8, i4x4;
 727         int i_decimate_mb = 0;
 728
 729         /* Don't repeat motion compensation if it was already done in non-RD transform analysis */
 730         if( !h->mb.b_skip_mc )
 731             x264_mb_mc( h );
 732
 733         if( h->mb.b_lossless )
 734         {
 735             if( h->mb.b_transform_8x8 )
 736                 for( i8x8 = 0; i8x8 < 4; i8x8++ )
 737                 {
 738                     int x = 8*(i8x8&1);
 739                     int y = 8*(i8x8>>1);
 740                     nz = h->zigzagf.sub_8x8( h->dct.luma8x8[i8x8],
 741                                         h->mb.pic.p_fenc[0]+x+y*FENC_STRIDE,
 742                                         h->mb.pic.p_fdec[0]+x+y*FDEC_STRIDE );
 743                     STORE_8x8_NNZ(i8x8,nz);
 744                     h->mb.i_cbp_luma |= nz << i8x8;
 745                 }
 746             else
 747                 for( i4x4 = 0; i4x4 < 16; i4x4++ )
 748                 {
 749                     nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4x4],
 750                                         h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4x4],
 751                                         h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4x4] );
 752                     h->mb.cache.non_zero_count[x264_scan8[i4x4]] = nz;
 753                     h->mb.i_cbp_luma |= nz << (i4x4>>2);
 754                 }
 755         }
 756         else if( h->mb.b_transform_8x8 )
 757         {
 758             ALIGNED_ARRAY_16( int16_t, dct8x8,[4],[64] );
 759             b_decimate &= !h->mb.b_trellis; // 8x8 trellis is inherently optimal decimation
 760             h->dctf.sub16x16_dct8( dct8x8, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 761             h->nr_count[1] += h->mb.b_noise_reduction * 4;
 762
 763             for( idx = 0; idx < 4; idx++ )
 764             {
 765                 if( h->mb.b_noise_reduction )
 766                     h->quantf.denoise_dct( dct8x8[idx], h->nr_residual_sum[1], h->nr_offset[1], 64 );
 767                 nz = x264_quant_8x8( h, dct8x8[idx], i_qp, 0, idx );
 768
 769                 if( nz )
 770                 {
 771                     h->zigzagf.scan_8x8( h->dct.luma8x8[idx], dct8x8[idx] );
 772                     if( b_decimate )
 773                     {
 774                         int i_decimate_8x8 = h->quantf.decimate_score64( h->dct.luma8x8[idx] );
 775                         i_decimate_mb += i_decimate_8x8;
 776                         if( i_decimate_8x8 >= 4 )
 777                             h->mb.i_cbp_luma |= 1<<idx;
 778                     }
 779                     else
 780                         h->mb.i_cbp_luma |= 1<<idx;
 781                 }
 782             }
 783
 784             if( i_decimate_mb < 6 && b_decimate )
 785             {
 786                 h->mb.i_cbp_luma = 0;
 787                 CLEAR_16x16_NNZ
 788             }
 789             else
 790             {
 791                 for( idx = 0; idx < 4; idx++ )
 792                 {
 793                     if( h->mb.i_cbp_luma&(1<<idx) )
 794                     {
 795                         h->quantf.dequant_8x8( dct8x8[idx], h->dequant8_mf[CQM_8PY], i_qp );
 796                         h->dctf.add8x8_idct8( &h->mb.pic.p_fdec[0][(idx&1)*8 + (idx>>1)*8*FDEC_STRIDE], dct8x8[idx] );
 797                         STORE_8x8_NNZ(idx,1);
 798                     }
 799                     else
 800                         STORE_8x8_NNZ(idx,0);
 801                 }
 802             }
 803         }
 804         else
 805         {
 806             ALIGNED_ARRAY_16( int16_t, dct4x4,[16],[16] );
 807             h->dctf.sub16x16_dct( dct4x4, h->mb.pic.p_fenc[0], h->mb.pic.p_fdec[0] );
 808             h->nr_count[0] += h->mb.b_noise_reduction * 16;
 809
 810             for( i8x8 = 0; i8x8 < 4; i8x8++ )
 811             {
 812                 int i_decimate_8x8 = 0;
 813                 int cbp = 0;
 814
 815                 /* encode one 4x4 block */
 816                 for( i4x4 = 0; i4x4 < 4; i4x4++ )
 817                 {
 818                     idx = i8x8 * 4 + i4x4;
 819
 820                     if( h->mb.b_noise_reduction )
 821                         h->quantf.denoise_dct( dct4x4[idx], h->nr_residual_sum[0], h->nr_offset[0], 16 );
 822                     nz = x264_quant_4x4( h, dct4x4[idx], i_qp, DCT_LUMA_4x4, 0, idx );
 823                     h->mb.cache.non_zero_count[x264_scan8[idx]] = nz;
 824
 825                     if( nz )
 826                     {
 827                         h->zigzagf.scan_4x4( h->dct.luma4x4[idx], dct4x4[idx] );
 828                         h->quantf.dequant_4x4( dct4x4[idx], h->dequant4_mf[CQM_4PY], i_qp );
 829                         if( b_decimate && i_decimate_8x8 < 6 )
 830                             i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[idx] );
 831                         cbp = 1;
 832                     }
 833                 }
 834
 835                 /* decimate this 8x8 block */
 836                 i_decimate_mb += i_decimate_8x8;
 837                 if( b_decimate )
 838                 {
 839                     if( i_decimate_8x8 < 4 )
 840                         STORE_8x8_NNZ(i8x8,0)
 841                     else
 842                         h->mb.i_cbp_luma |= 1<<i8x8;
 843                 }
 844                 else if( cbp )
 845                 {
 846                     h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 847                     h->mb.i_cbp_luma |= 1<<i8x8;
 848                 }
 849             }
 850
 851             if( b_decimate )
 852             {
 853                 if( i_decimate_mb < 6 )
 854                 {
 855                     h->mb.i_cbp_luma = 0;
 856                     CLEAR_16x16_NNZ
 857                 }
 858                 else
 859                 {
 860                     for( i8x8 = 0; i8x8 < 4; i8x8++ )
 861                         if( h->mb.i_cbp_luma&(1<<i8x8) )
 862                             h->dctf.add8x8_idct( &h->mb.pic.p_fdec[0][(i8x8&1)*8 + (i8x8>>1)*8*FDEC_STRIDE], &dct4x4[i8x8*4] );
 863                 }
 864             }
 865         }
 866     }
 867
 868     /* encode chroma */
 869     if( IS_INTRA( h->mb.i_type ) )
 870     {
 871         const int i_mode = h->mb.i_chroma_pred_mode;
 872         if( h->mb.b_lossless )
 873             x264_predict_lossless_8x8_chroma( h, i_mode );
 874         else
 875         {
 876             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[1] );
 877             h->predict_8x8c[i_mode]( h->mb.pic.p_fdec[2] );
 878         }
 879     }
 880
 881     /* encode the 8x8 blocks */
 882     x264_mb_encode_8x8_chroma( h, !IS_INTRA( h->mb.i_type ), h->mb.i_chroma_qp );
 883
 884     if( h->param.b_cabac )
 885     {
 886         i_cbp_dc = h->mb.cache.non_zero_count[x264_scan8[24]]
 887                  | h->mb.cache.non_zero_count[x264_scan8[25]] << 1
 888                  | h->mb.cache.non_zero_count[x264_scan8[26]] << 2;
 889     }
 890
 891     /* store cbp */
 892     h->mb.cbp[h->mb.i_mb_xy] = (i_cbp_dc << 8) | (h->mb.i_cbp_chroma << 4) | h->mb.i_cbp_luma;
 893
 894     /* Check for P_SKIP
 895      * XXX: in the me perhaps we should take x264_mb_predict_mv_pskip into account
 896      *      (if multiple mv give same result)*/
 897     if( !b_force_no_skip )
 898     {
 899         if( h->mb.i_type == P_L0 && h->mb.i_partition == D_16x16 &&
 900             !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) &&
 901             M32( h->mb.cache.mv[0][x264_scan8[0]] ) == M32( h->mb.cache.pskip_mv )
 902             && h->mb.cache.ref[0][x264_scan8[0]] == 0 )
 903         {
 904             h->mb.i_type = P_SKIP;
 905         }
 906
 907         /* Check for B_SKIP */
 908         if( h->mb.i_type == B_DIRECT && !(h->mb.i_cbp_luma | h->mb.i_cbp_chroma) )
 909         {
 910             h->mb.i_type = B_SKIP;
 911         }
 912     }
 913 }
 914
 915 /*****************************************************************************
 916  * x264_macroblock_probe_skip:
 917  *  Check if the current MB could be encoded as a [PB]_SKIP (it supposes you use
 918  *  the previous QP
 919  *****************************************************************************/
 920 int x264_macroblock_probe_skip( x264_t *h, int b_bidir )
 921 {
 922     ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
 923     ALIGNED_ARRAY_16( int16_t, dct2x2,[4] );
 924     ALIGNED_ARRAY_16( int16_t, dctscan,[16] );
 925
 926     int i_qp = h->mb.i_qp;
 927     int mvp[2];
 928     int ch, thresh, ssd;
 929
 930     int i8x8, i4x4;
 931     int i_decimate_mb;
 932
 933     if( !b_bidir )
 934     {
 935         /* Get the MV */
 936         mvp[0] = x264_clip3( h->mb.cache.pskip_mv[0], h->mb.mv_min[0], h->mb.mv_max[0] );
 937         mvp[1] = x264_clip3( h->mb.cache.pskip_mv[1], h->mb.mv_min[1], h->mb.mv_max[1] );
 938
 939         /* Motion compensation */
 940         h->mc.mc_luma( h->mb.pic.p_fdec[0],    FDEC_STRIDE,
 941                        h->mb.pic.p_fref[0][0], h->mb.pic.i_stride[0],
 942                        mvp[0], mvp[1], 16, 16, &h->sh.weight[0][0] );
 943     }
 944
 945     for( i8x8 = 0, i_decimate_mb = 0; i8x8 < 4; i8x8++ )
 946     {
 947         int fenc_offset = (i8x8&1) * 8 + (i8x8>>1) * FENC_STRIDE * 8;
 948         int fdec_offset = (i8x8&1) * 8 + (i8x8>>1) * FDEC_STRIDE * 8;
 949         /* get luma diff */
 950         h->dctf.sub8x8_dct( dct4x4, h->mb.pic.p_fenc[0] + fenc_offset,
 951                                     h->mb.pic.p_fdec[0] + fdec_offset );
 952         /* encode one 4x4 block */
 953         for( i4x4 = 0; i4x4 < 4; i4x4++ )
 954         {
 955             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PY][i_qp], h->quant4_bias[CQM_4PY][i_qp] ) )
 956                 continue;
 957             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
 958             i_decimate_mb += h->quantf.decimate_score16( dctscan );
 959             if( i_decimate_mb >= 6 )
 960                 return 0;
 961         }
 962     }
 963
 964     /* encode chroma */
 965     i_qp = h->mb.i_chroma_qp;
 966     thresh = (x264_lambda2_tab[i_qp] + 32) >> 6;
 967
 968     for( ch = 0; ch < 2; ch++ )
 969     {
 970         uint8_t  *p_src = h->mb.pic.p_fenc[1+ch];
 971         uint8_t  *p_dst = h->mb.pic.p_fdec[1+ch];
 972
 973         if( !b_bidir )
 974         {
 975             h->mc.mc_chroma( h->mb.pic.p_fdec[1+ch],       FDEC_STRIDE,
 976                              h->mb.pic.p_fref[0][0][4+ch], h->mb.pic.i_stride[1+ch],
 977                              mvp[0], mvp[1], 8, 8 );
 978
 979             if( h->sh.weight[0][1+ch].weightfn )
 980                 h->sh.weight[0][1+ch].weightfn[8>>2]( h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
 981                                                       h->mb.pic.p_fdec[1+ch], FDEC_STRIDE,
 982                                                       &h->sh.weight[0][1+ch], 8 );
 983         }
 984
 985         /* there is almost never a termination during chroma, but we can't avoid the check entirely */
 986         /* so instead we check SSD and skip the actual check if the score is low enough. */
 987         ssd = h->pixf.ssd[PIXEL_8x8]( p_dst, FDEC_STRIDE, p_src, FENC_STRIDE );
 988         if( ssd < thresh )
 989             continue;
 990
 991         h->dctf.sub8x8_dct( dct4x4, p_src, p_dst );
 992
 993         /* calculate dct DC */
 994         dct2x2dc( dct2x2, dct4x4 );
 995         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 ) )
 996             return 0;
 997
 998         /* If there wasn't a termination in DC, we can check against a much higher threshold. */
 999         if( ssd < thresh*4 )
1000             continue;
1001
1002         /* calculate dct coeffs */
1003         for( i4x4 = 0, i_decimate_mb = 0; i4x4 < 4; i4x4++ )
1004         {
1005             if( !h->quantf.quant_4x4( dct4x4[i4x4], h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] ) )
1006                 continue;
1007             h->zigzagf.scan_4x4( dctscan, dct4x4[i4x4] );
1008             i_decimate_mb += h->quantf.decimate_score15( dctscan );
1009             if( i_decimate_mb >= 7 )
1010                 return 0;
1011         }
1012     }
1013
1014     h->mb.b_skip_mc = 1;
1015     return 1;
1016 }
1017
1018 /****************************************************************************
1019  * DCT-domain noise reduction / adaptive deadzone
1020  * from libavcodec
1021  ****************************************************************************/
1022
1023 void x264_noise_reduction_update( x264_t *h )
1024 {
1025     int cat, i;
1026     for( cat = 0; cat < 2; cat++ )
1027     {
1028         int size = cat ? 64 : 16;
1029         const uint16_t *weight = cat ? x264_dct8_weight2_tab : x264_dct4_weight2_tab;
1030
1031         if( h->nr_count[cat] > (cat ? (1<<16) : (1<<18)) )
1032         {
1033             for( i = 0; i < size; i++ )
1034                 h->nr_residual_sum[cat][i] >>= 1;
1035             h->nr_count[cat] >>= 1;
1036         }
1037
1038         for( i = 0; i < size; i++ )
1039             h->nr_offset[cat][i] =
1040                 ((uint64_t)h->param.analyse.i_noise_reduction * h->nr_count[cat]
1041                  + h->nr_residual_sum[cat][i]/2)
1042               / ((uint64_t)h->nr_residual_sum[cat][i] * weight[i]/256 + 1);
1043     }
1044 }
1045
1046 /*****************************************************************************
1047  * RD only; 4 calls to this do not make up for one macroblock_encode.
1048  * doesn't transform chroma dc.
1049  *****************************************************************************/
1050 void x264_macroblock_encode_p8x8( x264_t *h, int i8 )
1051 {
1052     int i_qp = h->mb.i_qp;
1053     uint8_t *p_fenc = h->mb.pic.p_fenc[0] + (i8&1)*8 + (i8>>1)*8*FENC_STRIDE;
1054     uint8_t *p_fdec = h->mb.pic.p_fdec[0] + (i8&1)*8 + (i8>>1)*8*FDEC_STRIDE;
1055     int b_decimate = h->sh.i_type == SLICE_TYPE_B || h->param.analyse.b_dct_decimate;
1056     int nnz8x8 = 0;
1057     int ch, nz;
1058
1059     if( !h->mb.b_skip_mc )
1060         x264_mb_mc_8x8( h, i8 );
1061
1062     if( h->mb.b_lossless )
1063     {
1064         int i4;
1065         if( h->mb.b_transform_8x8 )
1066         {
1067             nnz8x8 = h->zigzagf.sub_8x8( h->dct.luma8x8[i8], p_fenc, p_fdec );
1068             STORE_8x8_NNZ(i8,nnz8x8);
1069         }
1070         else
1071         {
1072             for( i4 = i8*4; i4 < i8*4+4; i4++ )
1073             {
1074                 int nz;
1075                 nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4],
1076                                     h->mb.pic.p_fenc[0]+block_idx_xy_fenc[i4],
1077                                     h->mb.pic.p_fdec[0]+block_idx_xy_fdec[i4] );
1078                 h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1079                 nnz8x8 |= nz;
1080             }
1081         }
1082         for( ch = 0; ch < 2; ch++ )
1083         {
1084             int16_t dc;
1085             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1086             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1087             nz = h->zigzagf.sub_4x4ac( h->dct.luma4x4[16+i8+ch*4], p_fenc, p_fdec, &dc );
1088             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1089         }
1090     }
1091     else
1092     {
1093         if( h->mb.b_transform_8x8 )
1094         {
1095             ALIGNED_ARRAY_16( int16_t, dct8x8,[64] );
1096             h->dctf.sub8x8_dct8( dct8x8, p_fenc, p_fdec );
1097             nnz8x8 = x264_quant_8x8( h, dct8x8, i_qp, 0, i8 );
1098             if( nnz8x8 )
1099             {
1100                 h->zigzagf.scan_8x8( h->dct.luma8x8[i8], dct8x8 );
1101
1102                 if( b_decimate && !h->mb.b_trellis )
1103                     nnz8x8 = 4 <= h->quantf.decimate_score64( h->dct.luma8x8[i8] );
1104
1105                 if( nnz8x8 )
1106                 {
1107                     h->quantf.dequant_8x8( dct8x8, h->dequant8_mf[CQM_8PY], i_qp );
1108                     h->dctf.add8x8_idct8( p_fdec, dct8x8 );
1109                     STORE_8x8_NNZ(i8,1);
1110                 }
1111                 else
1112                     STORE_8x8_NNZ(i8,0);
1113             }
1114             else
1115                 STORE_8x8_NNZ(i8,0);
1116         }
1117         else
1118         {
1119             int i4;
1120             int i_decimate_8x8 = 0;
1121             ALIGNED_ARRAY_16( int16_t, dct4x4,[4],[16] );
1122             h->dctf.sub8x8_dct( dct4x4, p_fenc, p_fdec );
1123             for( i4 = 0; i4 < 4; i4++ )
1124             {
1125                 nz = x264_quant_4x4( h, dct4x4[i4], i_qp, DCT_LUMA_4x4, 0, i8*4+i4 );
1126                 h->mb.cache.non_zero_count[x264_scan8[i8*4+i4]] = nz;
1127                 if( nz )
1128                 {
1129                     h->zigzagf.scan_4x4( h->dct.luma4x4[i8*4+i4], dct4x4[i4] );
1130                     h->quantf.dequant_4x4( dct4x4[i4], h->dequant4_mf[CQM_4PY], i_qp );
1131                     if( b_decimate )
1132                         i_decimate_8x8 += h->quantf.decimate_score16( h->dct.luma4x4[i8*4+i4] );
1133                     nnz8x8 = 1;
1134                 }
1135             }
1136
1137             if( b_decimate && i_decimate_8x8 < 4 )
1138                 nnz8x8 = 0;
1139
1140             if( nnz8x8 )
1141                 h->dctf.add8x8_idct( p_fdec, dct4x4 );
1142             else
1143                 STORE_8x8_NNZ(i8,0);
1144         }
1145
1146         i_qp = h->mb.i_chroma_qp;
1147
1148         for( ch = 0; ch < 2; ch++ )
1149         {
1150             ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
1151             p_fenc = h->mb.pic.p_fenc[1+ch] + (i8&1)*4 + (i8>>1)*4*FENC_STRIDE;
1152             p_fdec = h->mb.pic.p_fdec[1+ch] + (i8&1)*4 + (i8>>1)*4*FDEC_STRIDE;
1153
1154             h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1155             dct4x4[0] = 0;
1156
1157             if( h->mb.b_trellis )
1158                 nz = x264_quant_4x4_trellis( h, dct4x4, CQM_4PC, i_qp, DCT_CHROMA_AC, 0, 1, 0 );
1159             else
1160                 nz = h->quantf.quant_4x4( dct4x4, h->quant4_mf[CQM_4PC][i_qp], h->quant4_bias[CQM_4PC][i_qp] );
1161
1162             h->mb.cache.non_zero_count[x264_scan8[16+i8+ch*4]] = nz;
1163             if( nz )
1164             {
1165                 h->zigzagf.scan_4x4( h->dct.luma4x4[16+i8+ch*4], dct4x4 );
1166                 h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PC], i_qp );
1167                 h->dctf.add4x4_idct( p_fdec, dct4x4 );
1168             }
1169         }
1170     }
1171     h->mb.i_cbp_luma &= ~(1 << i8);
1172     h->mb.i_cbp_luma |= nnz8x8 << i8;
1173     h->mb.i_cbp_chroma = 0x02;
1174 }
1175
1176 /*****************************************************************************
1177  * RD only, luma only
1178  *****************************************************************************/
1179 void x264_macroblock_encode_p4x4( x264_t *h, int i4 )
1180 {
1181     int i_qp = h->mb.i_qp;
1182     uint8_t *p_fenc = &h->mb.pic.p_fenc[0][block_idx_xy_fenc[i4]];
1183     uint8_t *p_fdec = &h->mb.pic.p_fdec[0][block_idx_xy_fdec[i4]];
1184     int nz;
1185
1186     /* Don't need motion compensation as this function is only used in qpel-RD, which caches pixel data. */
1187
1188     if( h->mb.b_lossless )
1189     {
1190         nz = h->zigzagf.sub_4x4( h->dct.luma4x4[i4], p_fenc, p_fdec );
1191         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1192     }
1193     else
1194     {
1195         ALIGNED_ARRAY_16( int16_t, dct4x4,[16] );
1196         h->dctf.sub4x4_dct( dct4x4, p_fenc, p_fdec );
1197         nz = x264_quant_4x4( h, dct4x4, i_qp, DCT_LUMA_4x4, 0, i4 );
1198         h->mb.cache.non_zero_count[x264_scan8[i4]] = nz;
1199         if( nz )
1200         {
1201             h->zigzagf.scan_4x4( h->dct.luma4x4[i4], dct4x4 );
1202             h->quantf.dequant_4x4( dct4x4, h->dequant4_mf[CQM_4PY], i_qp );
1203             h->dctf.add4x4_idct( p_fdec, dct4x4 );
1204         }
1205     }
1206 }