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 #ifdef ARCH_ARM
  33 #   include "arm/quant.h"
  34 #endif
  35
  36 #define QUANT_ONE( coef, mf, f ) \
  37 { \
  38     if( (coef) > 0 ) \
  39         (coef) = (f + (coef)) * (mf) >> 16; \
  40     else \
  41         (coef) = - ((f - (coef)) * (mf) >> 16); \
  42     nz |= (coef); \
  43 }
  44
  45 static int quant_8x8( int16_t dct[8][8], uint16_t mf[64], uint16_t bias[64] )
  46 {
  47     int i, nz = 0;
  48     for( i = 0; i < 64; i++ )
  49         QUANT_ONE( dct[0][i], mf[i], bias[i] );
  50     return !!nz;
  51 }
  52
  53 static int quant_4x4( int16_t dct[4][4], uint16_t mf[16], uint16_t bias[16] )
  54 {
  55     int i, nz = 0;
  56     for( i = 0; i < 16; i++ )
  57         QUANT_ONE( dct[0][i], mf[i], bias[i] );
  58     return !!nz;
  59 }
  60
  61 static int quant_4x4_dc( int16_t dct[4][4], int mf, int bias )
  62 {
  63     int i, nz = 0;
  64     for( i = 0; i < 16; i++ )
  65         QUANT_ONE( dct[0][i], mf, bias );
  66     return !!nz;
  67 }
  68
  69 static int quant_2x2_dc( int16_t dct[2][2], int mf, int bias )
  70 {
  71     int nz = 0;
  72     QUANT_ONE( dct[0][0], mf, bias );
  73     QUANT_ONE( dct[0][1], mf, bias );
  74     QUANT_ONE( dct[0][2], mf, bias );
  75     QUANT_ONE( dct[0][3], mf, bias );
  76     return !!nz;
  77 }
  78
  79 #define DEQUANT_SHL( x ) \
  80     dct[y][x] = ( dct[y][x] * dequant_mf[i_mf][y][x] ) << i_qbits
  81
  82 #define DEQUANT_SHR( x ) \
  83     dct[y][x] = ( dct[y][x] * dequant_mf[i_mf][y][x] + f ) >> (-i_qbits)
  84
  85 static void dequant_4x4( int16_t dct[4][4], int dequant_mf[6][4][4], int i_qp )
  86 {
  87     const int i_mf = i_qp%6;
  88     const int i_qbits = i_qp/6 - 4;
  89     int y;
  90
  91     if( i_qbits >= 0 )
  92     {
  93         for( y = 0; y < 4; y++ )
  94         {
  95             DEQUANT_SHL( 0 );
  96             DEQUANT_SHL( 1 );
  97             DEQUANT_SHL( 2 );
  98             DEQUANT_SHL( 3 );
  99         }
 100     }
 101     else
 102     {
 103         const int f = 1 << (-i_qbits-1);
 104         for( y = 0; y < 4; y++ )
 105         {
 106             DEQUANT_SHR( 0 );
 107             DEQUANT_SHR( 1 );
 108             DEQUANT_SHR( 2 );
 109             DEQUANT_SHR( 3 );
 110         }
 111     }
 112 }
 113
 114 static void dequant_8x8( int16_t dct[8][8], int dequant_mf[6][8][8], int i_qp )
 115 {
 116     const int i_mf = i_qp%6;
 117     const int i_qbits = i_qp/6 - 6;
 118     int y;
 119
 120     if( i_qbits >= 0 )
 121     {
 122         for( y = 0; y < 8; y++ )
 123         {
 124             DEQUANT_SHL( 0 );
 125             DEQUANT_SHL( 1 );
 126             DEQUANT_SHL( 2 );
 127             DEQUANT_SHL( 3 );
 128             DEQUANT_SHL( 4 );
 129             DEQUANT_SHL( 5 );
 130             DEQUANT_SHL( 6 );
 131             DEQUANT_SHL( 7 );
 132         }
 133     }
 134     else
 135     {
 136         const int f = 1 << (-i_qbits-1);
 137         for( y = 0; y < 8; y++ )
 138         {
 139             DEQUANT_SHR( 0 );
 140             DEQUANT_SHR( 1 );
 141             DEQUANT_SHR( 2 );
 142             DEQUANT_SHR( 3 );
 143             DEQUANT_SHR( 4 );
 144             DEQUANT_SHR( 5 );
 145             DEQUANT_SHR( 6 );
 146             DEQUANT_SHR( 7 );
 147         }
 148     }
 149 }
 150
 151 static void dequant_4x4_dc( int16_t dct[4][4], int dequant_mf[6][4][4], int i_qp )
 152 {
 153     const int i_qbits = i_qp/6 - 6;
 154     int y;
 155
 156     if( i_qbits >= 0 )
 157     {
 158         const int i_dmf = dequant_mf[i_qp%6][0][0] << i_qbits;
 159
 160         for( y = 0; y < 4; y++ )
 161         {
 162             dct[y][0] *= i_dmf;
 163             dct[y][1] *= i_dmf;
 164             dct[y][2] *= i_dmf;
 165             dct[y][3] *= i_dmf;
 166         }
 167     }
 168     else
 169     {
 170         const int i_dmf = dequant_mf[i_qp%6][0][0];
 171         const int f = 1 << (-i_qbits-1);
 172
 173         for( y = 0; y < 4; y++ )
 174         {
 175             dct[y][0] = ( dct[y][0] * i_dmf + f ) >> (-i_qbits);
 176             dct[y][1] = ( dct[y][1] * i_dmf + f ) >> (-i_qbits);
 177             dct[y][2] = ( dct[y][2] * i_dmf + f ) >> (-i_qbits);
 178             dct[y][3] = ( dct[y][3] * i_dmf + f ) >> (-i_qbits);
 179         }
 180     }
 181 }
 182
 183 static void x264_denoise_dct( int16_t *dct, uint32_t *sum, uint16_t *offset, int size )
 184 {
 185     int i;
 186     for( i=1; i<size; i++ )
 187     {
 188         int level = dct[i];
 189         int sign = level>>15;
 190         level = (level+sign)^sign;
 191         sum[i] += level;
 192         level -= offset[i];
 193         dct[i] = level<0 ? 0 : (level^sign)-sign;
 194     }
 195 }
 196
 197 /* (ref: JVT-B118)
 198  * x264_mb_decimate_score: given dct coeffs it returns a score to see if we could empty this dct coeffs
 199  * to 0 (low score means set it to null)
 200  * Used in inter macroblock (luma and chroma)
 201  *  luma: for a 8x8 block: if score < 4 -> null
 202  *        for the complete mb: if score < 6 -> null
 203  *  chroma: for the complete mb: if score < 7 -> null
 204  */
 205
 206 const uint8_t x264_decimate_table4[16] = {
 207     3,2,2,1,1,1,0,0,0,0,0,0,0,0,0,0 };
 208 const uint8_t x264_decimate_table8[64] = {
 209     3,3,3,3,2,2,2,2,2,2,2,2,1,1,1,1,
 210     1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,
 211     0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
 212     0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 };
 213
 214 static int ALWAYS_INLINE x264_decimate_score_internal( int16_t *dct, int i_max )
 215 {
 216     const uint8_t *ds_table = (i_max == 64) ? x264_decimate_table8 : x264_decimate_table4;
 217     int i_score = 0;
 218     int idx = i_max - 1;
 219
 220     /* Yes, dct[idx-1] is guaranteed to be 32-bit aligned.  idx>=0 instead of 1 works correctly for the same reason */
 221     while( idx >= 0 && *(uint32_t*)&dct[idx-1] == 0 )
 222         idx -= 2;
 223     if( idx >= 0 && dct[idx] == 0 )
 224         idx--;
 225     while( idx >= 0 )
 226     {
 227         int i_run;
 228
 229         if( (unsigned)(dct[idx--] + 1) > 2 )
 230             return 9;
 231
 232         i_run = 0;
 233         while( idx >= 0 && dct[idx] == 0 )
 234         {
 235             idx--;
 236             i_run++;
 237         }
 238         i_score += ds_table[i_run];
 239     }
 240
 241     return i_score;
 242 }
 243
 244 static int x264_decimate_score15( int16_t *dct )
 245 {
 246     return x264_decimate_score_internal( dct+1, 15 );
 247 }
 248 static int x264_decimate_score16( int16_t *dct )
 249 {
 250     return x264_decimate_score_internal( dct, 16 );
 251 }
 252 static int x264_decimate_score64( int16_t *dct )
 253 {
 254     return x264_decimate_score_internal( dct, 64 );
 255 }
 256
 257 static int ALWAYS_INLINE x264_coeff_last_internal( int16_t *l, int i_count )
 258 {
 259     int i_last;
 260     for( i_last = i_count-1; i_last >= 3; i_last -= 4 )
 261         if( *(uint64_t*)(l+i_last-3) )
 262             break;
 263     while( i_last >= 0 && l[i_last] == 0 )
 264         i_last--;
 265     return i_last;
 266 }
 267
 268 static int x264_coeff_last4( int16_t *l )
 269 {
 270     return x264_coeff_last_internal( l, 4 );
 271 }
 272 static int x264_coeff_last15( int16_t *l )
 273 {
 274     return x264_coeff_last_internal( l, 15 );
 275 }
 276 static int x264_coeff_last16( int16_t *l )
 277 {
 278     return x264_coeff_last_internal( l, 16 );
 279 }
 280 static int x264_coeff_last64( int16_t *l )
 281 {
 282     return x264_coeff_last_internal( l, 64 );
 283 }
 284
 285 #define level_run(num)\
 286 static int x264_coeff_level_run##num( int16_t *dct, x264_run_level_t *runlevel )\
 287 {\
 288     int i_last = runlevel->last = x264_coeff_last##num(dct);\
 289     int i_total = 0;\
 290     do\
 291     {\
 292         int r = 0;\
 293         runlevel->level[i_total] = dct[i_last];\
 294         while( --i_last >= 0 && dct[i_last] == 0 )\
 295             r++;\
 296         runlevel->run[i_total++] = r;\
 297     } while( i_last >= 0 );\
 298     return i_total;\
 299 }
 300
 301 level_run(4)
 302 level_run(15)
 303 level_run(16)
 304
 305
 306 void x264_quant_init( x264_t *h, int cpu, x264_quant_function_t *pf )
 307 {
 308     pf->quant_8x8 = quant_8x8;
 309     pf->quant_4x4 = quant_4x4;
 310     pf->quant_4x4_dc = quant_4x4_dc;
 311     pf->quant_2x2_dc = quant_2x2_dc;
 312
 313     pf->dequant_4x4 = dequant_4x4;
 314     pf->dequant_4x4_dc = dequant_4x4_dc;
 315     pf->dequant_8x8 = dequant_8x8;
 316
 317     pf->denoise_dct = x264_denoise_dct;
 318     pf->decimate_score15 = x264_decimate_score15;
 319     pf->decimate_score16 = x264_decimate_score16;
 320     pf->decimate_score64 = x264_decimate_score64;
 321
 322     pf->coeff_last[DCT_CHROMA_DC] = x264_coeff_last4;
 323     pf->coeff_last[  DCT_LUMA_AC] = x264_coeff_last15;
 324     pf->coeff_last[ DCT_LUMA_4x4] = x264_coeff_last16;
 325     pf->coeff_last[ DCT_LUMA_8x8] = x264_coeff_last64;
 326     pf->coeff_level_run[DCT_CHROMA_DC] = x264_coeff_level_run4;
 327     pf->coeff_level_run[  DCT_LUMA_AC] = x264_coeff_level_run15;
 328     pf->coeff_level_run[ DCT_LUMA_4x4] = x264_coeff_level_run16;
 329
 330 #ifdef HAVE_MMX
 331     if( cpu&X264_CPU_MMX )
 332     {
 333 #ifdef ARCH_X86
 334         pf->quant_4x4 = x264_quant_4x4_mmx;
 335         pf->quant_8x8 = x264_quant_8x8_mmx;
 336         pf->dequant_4x4 = x264_dequant_4x4_mmx;
 337         pf->dequant_4x4_dc = x264_dequant_4x4dc_mmxext;
 338         pf->dequant_8x8 = x264_dequant_8x8_mmx;
 339         if( h->param.i_cqm_preset == X264_CQM_FLAT )
 340         {
 341             pf->dequant_4x4 = x264_dequant_4x4_flat16_mmx;
 342             pf->dequant_8x8 = x264_dequant_8x8_flat16_mmx;
 343         }
 344         pf->denoise_dct = x264_denoise_dct_mmx;
 345 #endif
 346     }
 347
 348     if( cpu&X264_CPU_MMXEXT )
 349     {
 350         pf->quant_2x2_dc = x264_quant_2x2_dc_mmxext;
 351 #ifdef ARCH_X86
 352         pf->quant_4x4_dc = x264_quant_4x4_dc_mmxext;
 353         pf->decimate_score15 = x264_decimate_score15_mmxext;
 354         pf->decimate_score16 = x264_decimate_score16_mmxext;
 355         pf->decimate_score64 = x264_decimate_score64_mmxext;
 356         pf->coeff_last[  DCT_LUMA_AC] = x264_coeff_last15_mmxext;
 357         pf->coeff_last[ DCT_LUMA_4x4] = x264_coeff_last16_mmxext;
 358         pf->coeff_last[ DCT_LUMA_8x8] = x264_coeff_last64_mmxext;
 359         pf->coeff_level_run[  DCT_LUMA_AC] = x264_coeff_level_run15_mmxext;
 360         pf->coeff_level_run[ DCT_LUMA_4x4] = x264_coeff_level_run16_mmxext;
 361 #endif
 362         pf->coeff_last[DCT_CHROMA_DC] = x264_coeff_last4_mmxext;
 363         pf->coeff_level_run[DCT_CHROMA_DC] = x264_coeff_level_run4_mmxext;
 364         if( cpu&X264_CPU_LZCNT )
 365         {
 366             pf->coeff_last[DCT_CHROMA_DC] = x264_coeff_last4_mmxext_lzcnt;
 367             pf->coeff_level_run[DCT_CHROMA_DC] = x264_coeff_level_run4_mmxext_lzcnt;
 368         }
 369     }
 370
 371     if( cpu&X264_CPU_SSE2 )
 372     {
 373         pf->quant_4x4_dc = x264_quant_4x4_dc_sse2;
 374         pf->quant_4x4 = x264_quant_4x4_sse2;
 375         pf->quant_8x8 = x264_quant_8x8_sse2;
 376         pf->dequant_4x4 = x264_dequant_4x4_sse2;
 377         pf->dequant_4x4_dc = x264_dequant_4x4dc_sse2;
 378         pf->dequant_8x8 = x264_dequant_8x8_sse2;
 379         if( h->param.i_cqm_preset == X264_CQM_FLAT )
 380         {
 381             pf->dequant_4x4 = x264_dequant_4x4_flat16_sse2;
 382             pf->dequant_8x8 = x264_dequant_8x8_flat16_sse2;
 383         }
 384         pf->denoise_dct = x264_denoise_dct_sse2;
 385         pf->decimate_score15 = x264_decimate_score15_sse2;
 386         pf->decimate_score16 = x264_decimate_score16_sse2;
 387         pf->decimate_score64 = x264_decimate_score64_sse2;
 388         pf->coeff_last[ DCT_LUMA_AC] = x264_coeff_last15_sse2;
 389         pf->coeff_last[DCT_LUMA_4x4] = x264_coeff_last16_sse2;
 390         pf->coeff_last[DCT_LUMA_8x8] = x264_coeff_last64_sse2;
 391         pf->coeff_level_run[ DCT_LUMA_AC] = x264_coeff_level_run15_sse2;
 392         pf->coeff_level_run[DCT_LUMA_4x4] = x264_coeff_level_run16_sse2;
 393         if( cpu&X264_CPU_LZCNT )
 394         {
 395             pf->coeff_last[ DCT_LUMA_AC] = x264_coeff_last15_sse2_lzcnt;
 396             pf->coeff_last[DCT_LUMA_4x4] = x264_coeff_last16_sse2_lzcnt;
 397             pf->coeff_last[DCT_LUMA_8x8] = x264_coeff_last64_sse2_lzcnt;
 398             pf->coeff_level_run[ DCT_LUMA_AC] = x264_coeff_level_run15_sse2_lzcnt;
 399             pf->coeff_level_run[DCT_LUMA_4x4] = x264_coeff_level_run16_sse2_lzcnt;
 400         }
 401     }
 402
 403     if( cpu&X264_CPU_SSSE3 )
 404     {
 405         pf->quant_2x2_dc = x264_quant_2x2_dc_ssse3;
 406         pf->quant_4x4_dc = x264_quant_4x4_dc_ssse3;
 407         pf->quant_4x4 = x264_quant_4x4_ssse3;
 408         pf->quant_8x8 = x264_quant_8x8_ssse3;
 409         pf->denoise_dct = x264_denoise_dct_ssse3;
 410         pf->decimate_score15 = x264_decimate_score15_ssse3;
 411         pf->decimate_score16 = x264_decimate_score16_ssse3;
 412         pf->decimate_score64 = x264_decimate_score64_ssse3;
 413     }
 414
 415     if( cpu&X264_CPU_SSE4 )
 416     {
 417         pf->quant_4x4_dc = x264_quant_4x4_dc_sse4;
 418         pf->quant_4x4 = x264_quant_4x4_sse4;
 419         pf->quant_8x8 = x264_quant_8x8_sse4;
 420     }
 421 #endif // HAVE_MMX
 422
 423 #ifdef ARCH_PPC
 424     if( cpu&X264_CPU_ALTIVEC ) {
 425         pf->quant_2x2_dc = x264_quant_2x2_dc_altivec;
 426         pf->quant_4x4_dc = x264_quant_4x4_dc_altivec;
 427         pf->quant_4x4 = x264_quant_4x4_altivec;
 428         pf->quant_8x8 = x264_quant_8x8_altivec;
 429
 430         pf->dequant_4x4 = x264_dequant_4x4_altivec;
 431         pf->dequant_8x8 = x264_dequant_8x8_altivec;
 432     }
 433 #endif
 434
 435 #ifdef HAVE_ARMV6
 436     if( cpu&X264_CPU_ARMV6 )
 437         pf->coeff_last[DCT_CHROMA_DC] = x264_coeff_last4_arm;
 438
 439     if( cpu&X264_CPU_NEON )
 440     {
 441         pf->quant_2x2_dc   = x264_quant_2x2_dc_neon;
 442         pf->quant_4x4      = x264_quant_4x4_neon;
 443         pf->quant_4x4_dc   = x264_quant_4x4_dc_neon;
 444         pf->quant_8x8      = x264_quant_8x8_neon;
 445         pf->dequant_4x4    = x264_dequant_4x4_neon;
 446         pf->dequant_4x4_dc = x264_dequant_4x4_dc_neon;
 447         pf->dequant_8x8    = x264_dequant_8x8_neon;
 448         pf->coeff_last[ DCT_LUMA_AC] = x264_coeff_last15_neon;
 449         pf->coeff_last[DCT_LUMA_4x4] = x264_coeff_last16_neon;
 450         pf->coeff_last[DCT_LUMA_8x8] = x264_coeff_last64_neon;
 451     }
 452 #endif
 453     pf->coeff_last[  DCT_LUMA_DC] = pf->coeff_last[DCT_LUMA_4x4];
 454     pf->coeff_last[DCT_CHROMA_AC] = pf->coeff_last[ DCT_LUMA_AC];
 455     pf->coeff_level_run[  DCT_LUMA_DC] = pf->coeff_level_run[DCT_LUMA_4x4];
 456     pf->coeff_level_run[DCT_CHROMA_AC] = pf->coeff_level_run[ DCT_LUMA_AC];
 457 }