git.sesse.net Git - x264/blob - common/quant.c

   1 /*****************************************************************************
   2  * quant.c: h264 encoder library
   3  *****************************************************************************
   4  * Copyright (C) 2005-2008 x264 project
   5  *
   6  * Authors: Loren Merritt <lorenm@u.washington.edu>
   7  *          Christian Heine <sennindemokrit@gmx.net>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
  22  *****************************************************************************/
  23
  24 #include "common.h"
  25
  26 #ifdef HAVE_MMX
  27 #include "x86/quant.h"
  28 #endif
  29 #ifdef ARCH_PPC
  30 #   include "ppc/quant.h"
  31 #endif
  32
  33 #define QUANT_ONE( coef, mf, f ) \
  34 { \
  35     if( (coef) > 0 ) \
  36         (coef) = (f + (coef)) * (mf) >> 16; \
  37     else \
  38         (coef) = - ((f - (coef)) * (mf) >> 16); \
  39     nz |= (coef); \
  40 }
  41
  42 static int quant_8x8( int16_t dct[8][8], uint16_t mf[64], uint16_t bias[64] )
  43 {
  44     int i, nz = 0;
  45     for( i = 0; i < 64; i++ )
  46         QUANT_ONE( dct[0][i], mf[i], bias[i] );
  47     return !!nz;
  48 }
  49
  50 static int quant_4x4( int16_t dct[4][4], uint16_t mf[16], uint16_t bias[16] )
  51 {
  52     int i, nz = 0;
  53     for( i = 0; i < 16; i++ )
  54         QUANT_ONE( dct[0][i], mf[i], bias[i] );
  55     return !!nz;
  56 }
  57
  58 static int quant_4x4_dc( int16_t dct[4][4], int mf, int bias )
  59 {
  60     int i, nz = 0;
  61     for( i = 0; i < 16; i++ )
  62         QUANT_ONE( dct[0][i], mf, bias );
  63     return !!nz;
  64 }
  65
  66 static int quant_2x2_dc( int16_t dct[2][2], int mf, int bias )
  67 {
  68     int nz = 0;
  69     QUANT_ONE( dct[0][0], mf, bias );
  70     QUANT_ONE( dct[0][1], mf, bias );
  71     QUANT_ONE( dct[0][2], mf, bias );
  72     QUANT_ONE( dct[0][3], mf, bias );
  73     return !!nz;
  74 }
  75
  76 #define DEQUANT_SHL( x ) \
  77     dct[y][x] = ( dct[y][x] * dequant_mf[i_mf][y][x] ) << i_qbits
  78
  79 #define DEQUANT_SHR( x ) \
  80     dct[y][x] = ( dct[y][x] * dequant_mf[i_mf][y][x] + f ) >> (-i_qbits)
  81
  82 static void dequant_4x4( int16_t dct[4][4], int dequant_mf[6][4][4], int i_qp )
  83 {
  84     const int i_mf = i_qp%6;
  85     const int i_qbits = i_qp/6 - 4;
  86     int y;
  87
  88     if( i_qbits >= 0 )
  89     {
  90         for( y = 0; y < 4; y++ )
  91         {
  92             DEQUANT_SHL( 0 );
  93             DEQUANT_SHL( 1 );
  94             DEQUANT_SHL( 2 );
  95             DEQUANT_SHL( 3 );
  96         }
  97     }
  98     else
  99     {
 100         const int f = 1 << (-i_qbits-1);
 101         for( y = 0; y < 4; y++ )
 102         {
 103             DEQUANT_SHR( 0 );
 104             DEQUANT_SHR( 1 );
 105             DEQUANT_SHR( 2 );
 106             DEQUANT_SHR( 3 );
 107         }
 108     }
 109 }
 110
 111 static void dequant_8x8( int16_t dct[8][8], int dequant_mf[6][8][8], int i_qp )
 112 {
 113     const int i_mf = i_qp%6;
 114     const int i_qbits = i_qp/6 - 6;
 115     int y;
 116
 117     if( i_qbits >= 0 )
 118     {
 119         for( y = 0; y < 8; y++ )
 120         {
 121             DEQUANT_SHL( 0 );
 122             DEQUANT_SHL( 1 );
 123             DEQUANT_SHL( 2 );
 124             DEQUANT_SHL( 3 );
 125             DEQUANT_SHL( 4 );
 126             DEQUANT_SHL( 5 );
 127             DEQUANT_SHL( 6 );
 128             DEQUANT_SHL( 7 );
 129         }
 130     }
 131     else
 132     {
 133         const int f = 1 << (-i_qbits-1);
 134         for( y = 0; y < 8; y++ )
 135         {
 136             DEQUANT_SHR( 0 );
 137             DEQUANT_SHR( 1 );
 138             DEQUANT_SHR( 2 );
 139             DEQUANT_SHR( 3 );
 140             DEQUANT_SHR( 4 );
 141             DEQUANT_SHR( 5 );
 142             DEQUANT_SHR( 6 );
 143             DEQUANT_SHR( 7 );
 144         }
 145     }
 146 }
 147
 148 static void dequant_4x4_dc( int16_t dct[4][4], int dequant_mf[6][4][4], int i_qp )
 149 {
 150     const int i_qbits = i_qp/6 - 6;
 151     int y;
 152
 153     if( i_qbits >= 0 )
 154     {
 155         const int i_dmf = dequant_mf[i_qp%6][0][0] << i_qbits;
 156
 157         for( y = 0; y < 4; y++ )
 158         {
 159             dct[y][0] *= i_dmf;
 160             dct[y][1] *= i_dmf;
 161             dct[y][2] *= i_dmf;
 162             dct[y][3] *= i_dmf;
 163         }
 164     }
 165     else
 166     {
 167         const int i_dmf = dequant_mf[i_qp%6][0][0];
 168         const int f = 1 << (-i_qbits-1);
 169
 170         for( y = 0; y < 4; y++ )
 171         {
 172             dct[y][0] = ( dct[y][0] * i_dmf + f ) >> (-i_qbits);
 173             dct[y][1] = ( dct[y][1] * i_dmf + f ) >> (-i_qbits);
 174             dct[y][2] = ( dct[y][2] * i_dmf + f ) >> (-i_qbits);
 175             dct[y][3] = ( dct[y][3] * i_dmf + f ) >> (-i_qbits);
 176         }
 177     }
 178 }
 179
 180 static void x264_denoise_dct( int16_t *dct, uint32_t *sum, uint16_t *offset, int size )
 181 {
 182     int i;
 183     for( i=1; i<size; i++ )
 184     {
 185         int level = dct[i];
 186         int sign = level>>15;
 187         level = (level+sign)^sign;
 188         sum[i] += level;
 189         level -= offset[i];
 190         dct[i] = level<0 ? 0 : (level^sign)-sign;
 191     }
 192 }
 193
 194 /* (ref: JVT-B118)
 195  * x264_mb_decimate_score: given dct coeffs it returns a score to see if we could empty this dct coeffs
 196  * to 0 (low score means set it to null)
 197  * Used in inter macroblock (luma and chroma)
 198  *  luma: for a 8x8 block: if score < 4 -> null
 199  *        for the complete mb: if score < 6 -> null
 200  *  chroma: for the complete mb: if score < 7 -> null
 201  */
 202
 203 const uint8_t x264_decimate_table4[16] = {
 204     3,2,2,1,1,1,0,0,0,0,0,0,0,0,0,0 };
 205 const uint8_t x264_decimate_table8[64] = {
 206     3,3,3,3,2,2,2,2,2,2,2,2,1,1,1,1,
 207     1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,
 208     0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
 209     0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 };
 210
 211 static int ALWAYS_INLINE x264_decimate_score_internal( int16_t *dct, int i_max )
 212 {
 213     const uint8_t *ds_table = (i_max == 64) ? x264_decimate_table8 : x264_decimate_table4;
 214     int i_score = 0;
 215     int idx = i_max - 1;
 216
 217     /* Yes, dct[idx-1] is guaranteed to be 32-bit aligned.  idx>=0 instead of 1 works correctly for the same reason */
 218     while( idx >= 0 && *(uint32_t*)&dct[idx-1] == 0 )
 219         idx -= 2;
 220     if( idx >= 0 && dct[idx] == 0 )
 221         idx--;
 222     while( idx >= 0 )
 223     {
 224         int i_run;
 225
 226         if( (unsigned)(dct[idx--] + 1) > 2 )
 227             return 9;
 228
 229         i_run = 0;
 230         while( idx >= 0 && dct[idx] == 0 )
 231         {
 232             idx--;
 233             i_run++;
 234         }
 235         i_score += ds_table[i_run];
 236     }
 237
 238     return i_score;
 239 }
 240
 241 static int x264_decimate_score15( int16_t *dct )
 242 {
 243     return x264_decimate_score_internal( dct+1, 15 );
 244 }
 245 static int x264_decimate_score16( int16_t *dct )
 246 {
 247     return x264_decimate_score_internal( dct, 16 );
 248 }
 249 static int x264_decimate_score64( int16_t *dct )
 250 {
 251     return x264_decimate_score_internal( dct, 64 );
 252 }
 253
 254 static int ALWAYS_INLINE x264_coeff_last_internal( int16_t *l, int i_count )
 255 {
 256     int i_last;
 257     for( i_last = i_count-1; i_last >= 3; i_last -= 4 )
 258         if( *(uint64_t*)(l+i_last-3) )
 259             break;
 260     while( i_last >= 0 && l[i_last] == 0 )
 261         i_last--;
 262     return i_last;
 263 }
 264
 265 static int x264_coeff_last4( int16_t *l )
 266 {
 267     return x264_coeff_last_internal( l, 4 );
 268 }
 269 static int x264_coeff_last15( int16_t *l )
 270 {
 271     return x264_coeff_last_internal( l, 15 );
 272 }
 273 static int x264_coeff_last16( int16_t *l )
 274 {
 275     return x264_coeff_last_internal( l, 16 );
 276 }
 277 static int x264_coeff_last64( int16_t *l )
 278 {
 279     return x264_coeff_last_internal( l, 64 );
 280 }
 281
 282 #define level_run(num)\
 283 static int x264_coeff_level_run##num( int16_t *dct, x264_run_level_t *runlevel )\
 284 {\
 285     int i_last = runlevel->last = x264_coeff_last##num(dct);\
 286     int i_total = 0;\
 287     do\
 288     {\
 289         int r = 0;\
 290         runlevel->level[i_total] = dct[i_last];\
 291         while( --i_last >= 0 && dct[i_last] == 0 )\
 292             r++;\
 293         runlevel->run[i_total++] = r;\
 294     } while( i_last >= 0 );\
 295     return i_total;\
 296 }
 297
 298 level_run(4)
 299 level_run(15)
 300 level_run(16)
 301
 302
 303 void x264_quant_init( x264_t *h, int cpu, x264_quant_function_t *pf )
 304 {
 305     pf->quant_8x8 = quant_8x8;
 306     pf->quant_4x4 = quant_4x4;
 307     pf->quant_4x4_dc = quant_4x4_dc;
 308     pf->quant_2x2_dc = quant_2x2_dc;
 309
 310     pf->dequant_4x4 = dequant_4x4;
 311     pf->dequant_4x4_dc = dequant_4x4_dc;
 312     pf->dequant_8x8 = dequant_8x8;
 313
 314     pf->denoise_dct = x264_denoise_dct;
 315     pf->decimate_score15 = x264_decimate_score15;
 316     pf->decimate_score16 = x264_decimate_score16;
 317     pf->decimate_score64 = x264_decimate_score64;
 318
 319     pf->coeff_last[DCT_CHROMA_DC] = x264_coeff_last4;
 320     pf->coeff_last[  DCT_LUMA_AC] = x264_coeff_last15;
 321     pf->coeff_last[ DCT_LUMA_4x4] = x264_coeff_last16;
 322     pf->coeff_last[ DCT_LUMA_8x8] = x264_coeff_last64;
 323     pf->coeff_level_run[DCT_CHROMA_DC] = x264_coeff_level_run4;
 324     pf->coeff_level_run[  DCT_LUMA_AC] = x264_coeff_level_run15;
 325     pf->coeff_level_run[ DCT_LUMA_4x4] = x264_coeff_level_run16;
 326
 327 #ifdef HAVE_MMX
 328     if( cpu&X264_CPU_MMX )
 329     {
 330 #ifdef ARCH_X86
 331         pf->quant_4x4 = x264_quant_4x4_mmx;
 332         pf->quant_8x8 = x264_quant_8x8_mmx;
 333         pf->dequant_4x4 = x264_dequant_4x4_mmx;
 334         pf->dequant_4x4_dc = x264_dequant_4x4dc_mmxext;
 335         pf->dequant_8x8 = x264_dequant_8x8_mmx;
 336         if( h->param.i_cqm_preset == X264_CQM_FLAT )
 337         {
 338             pf->dequant_4x4 = x264_dequant_4x4_flat16_mmx;
 339             pf->dequant_8x8 = x264_dequant_8x8_flat16_mmx;
 340         }
 341         pf->denoise_dct = x264_denoise_dct_mmx;
 342 #endif
 343     }
 344
 345     if( cpu&X264_CPU_MMXEXT )
 346     {
 347         pf->quant_2x2_dc = x264_quant_2x2_dc_mmxext;
 348 #ifdef ARCH_X86
 349         pf->quant_4x4_dc = x264_quant_4x4_dc_mmxext;
 350         pf->decimate_score15 = x264_decimate_score15_mmxext;
 351         pf->decimate_score16 = x264_decimate_score16_mmxext;
 352         pf->decimate_score64 = x264_decimate_score64_mmxext;
 353         pf->coeff_last[  DCT_LUMA_AC] = x264_coeff_last15_mmxext;
 354         pf->coeff_last[ DCT_LUMA_4x4] = x264_coeff_last16_mmxext;
 355         pf->coeff_last[ DCT_LUMA_8x8] = x264_coeff_last64_mmxext;
 356         pf->coeff_level_run[  DCT_LUMA_AC] = x264_coeff_level_run15_mmxext;
 357         pf->coeff_level_run[ DCT_LUMA_4x4] = x264_coeff_level_run16_mmxext;
 358 #endif
 359         pf->coeff_last[DCT_CHROMA_DC] = x264_coeff_last4_mmxext;
 360         pf->coeff_level_run[DCT_CHROMA_DC] = x264_coeff_level_run4_mmxext;
 361         if( cpu&X264_CPU_LZCNT )
 362         {
 363             pf->coeff_last[DCT_CHROMA_DC] = x264_coeff_last4_mmxext_lzcnt;
 364             pf->coeff_level_run[DCT_CHROMA_DC] = x264_coeff_level_run4_mmxext_lzcnt;
 365         }
 366     }
 367
 368     if( cpu&X264_CPU_SSE2 )
 369     {
 370         pf->quant_4x4_dc = x264_quant_4x4_dc_sse2;
 371         pf->quant_4x4 = x264_quant_4x4_sse2;
 372         pf->quant_8x8 = x264_quant_8x8_sse2;
 373         pf->dequant_4x4 = x264_dequant_4x4_sse2;
 374         pf->dequant_4x4_dc = x264_dequant_4x4dc_sse2;
 375         pf->dequant_8x8 = x264_dequant_8x8_sse2;
 376         if( h->param.i_cqm_preset == X264_CQM_FLAT )
 377         {
 378             pf->dequant_4x4 = x264_dequant_4x4_flat16_sse2;
 379             pf->dequant_8x8 = x264_dequant_8x8_flat16_sse2;
 380         }
 381         pf->denoise_dct = x264_denoise_dct_sse2;
 382         pf->decimate_score15 = x264_decimate_score15_sse2;
 383         pf->decimate_score16 = x264_decimate_score16_sse2;
 384         pf->decimate_score64 = x264_decimate_score64_sse2;
 385         pf->coeff_last[ DCT_LUMA_AC] = x264_coeff_last15_sse2;
 386         pf->coeff_last[DCT_LUMA_4x4] = x264_coeff_last16_sse2;
 387         pf->coeff_last[DCT_LUMA_8x8] = x264_coeff_last64_sse2;
 388         pf->coeff_level_run[ DCT_LUMA_AC] = x264_coeff_level_run15_sse2;
 389         pf->coeff_level_run[DCT_LUMA_4x4] = x264_coeff_level_run16_sse2;
 390         if( cpu&X264_CPU_LZCNT )
 391         {
 392             pf->coeff_last[ DCT_LUMA_AC] = x264_coeff_last15_sse2_lzcnt;
 393             pf->coeff_last[DCT_LUMA_4x4] = x264_coeff_last16_sse2_lzcnt;
 394             pf->coeff_last[DCT_LUMA_8x8] = x264_coeff_last64_sse2_lzcnt;
 395             pf->coeff_level_run[ DCT_LUMA_AC] = x264_coeff_level_run15_sse2_lzcnt;
 396             pf->coeff_level_run[DCT_LUMA_4x4] = x264_coeff_level_run16_sse2_lzcnt;
 397         }
 398     }
 399
 400     if( cpu&X264_CPU_SSSE3 )
 401     {
 402         pf->quant_2x2_dc = x264_quant_2x2_dc_ssse3;
 403         pf->quant_4x4_dc = x264_quant_4x4_dc_ssse3;
 404         pf->quant_4x4 = x264_quant_4x4_ssse3;
 405         pf->quant_8x8 = x264_quant_8x8_ssse3;
 406         pf->denoise_dct = x264_denoise_dct_ssse3;
 407         pf->decimate_score15 = x264_decimate_score15_ssse3;
 408         pf->decimate_score16 = x264_decimate_score16_ssse3;
 409         pf->decimate_score64 = x264_decimate_score64_ssse3;
 410     }
 411
 412     if( cpu&X264_CPU_SSE4 )
 413     {
 414         pf->quant_4x4_dc = x264_quant_4x4_dc_sse4;
 415         pf->quant_4x4 = x264_quant_4x4_sse4;
 416         pf->quant_8x8 = x264_quant_8x8_sse4;
 417     }
 418 #endif // HAVE_MMX
 419
 420 #ifdef ARCH_PPC
 421     if( cpu&X264_CPU_ALTIVEC ) {
 422         pf->quant_2x2_dc = x264_quant_2x2_dc_altivec;
 423         pf->quant_4x4_dc = x264_quant_4x4_dc_altivec;
 424         pf->quant_4x4 = x264_quant_4x4_altivec;
 425         pf->quant_8x8 = x264_quant_8x8_altivec;
 426
 427         pf->dequant_4x4 = x264_dequant_4x4_altivec;
 428         pf->dequant_8x8 = x264_dequant_8x8_altivec;
 429     }
 430 #endif
 431     pf->coeff_last[  DCT_LUMA_DC] = pf->coeff_last[DCT_LUMA_4x4];
 432     pf->coeff_last[DCT_CHROMA_AC] = pf->coeff_last[ DCT_LUMA_AC];
 433     pf->coeff_level_run[  DCT_LUMA_DC] = pf->coeff_level_run[DCT_LUMA_4x4];
 434     pf->coeff_level_run[DCT_CHROMA_AC] = pf->coeff_level_run[ DCT_LUMA_AC];
 435 }