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[64], 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[i], mf[i], bias[i] );
  50     return !!nz;
  51 }
  52
  53 static int quant_4x4( int16_t dct[16], 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[i], mf[i], bias[i] );
  58     return !!nz;
  59 }
  60
  61 static int quant_4x4_dc( int16_t dct[16], int mf, int bias )
  62 {
  63     int i, nz = 0;
  64     for( i = 0; i < 16; i++ )
  65         QUANT_ONE( dct[i], mf, bias );
  66     return !!nz;
  67 }
  68
  69 static int quant_2x2_dc( int16_t dct[4], int mf, int bias )
  70 {
  71     int nz = 0;
  72     QUANT_ONE( dct[0], mf, bias );
  73     QUANT_ONE( dct[1], mf, bias );
  74     QUANT_ONE( dct[2], mf, bias );
  75     QUANT_ONE( dct[3], mf, bias );
  76     return !!nz;
  77 }
  78
  79 #define DEQUANT_SHL( x ) \
  80     dct[x] = ( dct[x] * dequant_mf[i_mf][x] ) << i_qbits
  81
  82 #define DEQUANT_SHR( x ) \
  83     dct[x] = ( dct[x] * dequant_mf[i_mf][x] + f ) >> (-i_qbits)
  84
  85 static void dequant_4x4( int16_t dct[16], int dequant_mf[6][16], int i_qp )
  86 {
  87     const int i_mf = i_qp%6;
  88     const int i_qbits = i_qp/6 - 4;
  89     int i;
  90
  91     if( i_qbits >= 0 )
  92     {
  93         for( i = 0; i < 16; i++ )
  94             DEQUANT_SHL( i );
  95     }
  96     else
  97     {
  98         const int f = 1 << (-i_qbits-1);
  99         for( i = 0; i < 16; i++ )
 100             DEQUANT_SHR( i );
 101     }
 102 }
 103
 104 static void dequant_8x8( int16_t dct[64], int dequant_mf[6][64], int i_qp )
 105 {
 106     const int i_mf = i_qp%6;
 107     const int i_qbits = i_qp/6 - 6;
 108     int i;
 109
 110     if( i_qbits >= 0 )
 111     {
 112         for( i = 0; i < 64; i++ )
 113             DEQUANT_SHL( i );
 114     }
 115     else
 116     {
 117         const int f = 1 << (-i_qbits-1);
 118         for( i = 0; i < 64; i++ )
 119             DEQUANT_SHR( i );
 120     }
 121 }
 122
 123 static void dequant_4x4_dc( int16_t dct[16], int dequant_mf[6][16], int i_qp )
 124 {
 125     const int i_qbits = i_qp/6 - 6;
 126     int i;
 127
 128     if( i_qbits >= 0 )
 129     {
 130         const int i_dmf = dequant_mf[i_qp%6][0] << i_qbits;
 131         for( i = 0; i < 16; i++ )
 132             dct[i] *= i_dmf;
 133     }
 134     else
 135     {
 136         const int i_dmf = dequant_mf[i_qp%6][0];
 137         const int f = 1 << (-i_qbits-1);
 138         for( i = 0; i < 16; i++ )
 139             dct[i] = ( dct[i] * i_dmf + f ) >> (-i_qbits);
 140     }
 141 }
 142
 143 static void x264_denoise_dct( int16_t *dct, uint32_t *sum, uint16_t *offset, int size )
 144 {
 145     int i;
 146     for( i=1; i<size; i++ )
 147     {
 148         int level = dct[i];
 149         int sign = level>>15;
 150         level = (level+sign)^sign;
 151         sum[i] += level;
 152         level -= offset[i];
 153         dct[i] = level<0 ? 0 : (level^sign)-sign;
 154     }
 155 }
 156
 157 /* (ref: JVT-B118)
 158  * x264_mb_decimate_score: given dct coeffs it returns a score to see if we could empty this dct coeffs
 159  * to 0 (low score means set it to null)
 160  * Used in inter macroblock (luma and chroma)
 161  *  luma: for a 8x8 block: if score < 4 -> null
 162  *        for the complete mb: if score < 6 -> null
 163  *  chroma: for the complete mb: if score < 7 -> null
 164  */
 165
 166 const uint8_t x264_decimate_table4[16] = {
 167     3,2,2,1,1,1,0,0,0,0,0,0,0,0,0,0 };
 168 const uint8_t x264_decimate_table8[64] = {
 169     3,3,3,3,2,2,2,2,2,2,2,2,1,1,1,1,
 170     1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,
 171     0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
 172     0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 };
 173
 174 static int ALWAYS_INLINE x264_decimate_score_internal( int16_t *dct, int i_max )
 175 {
 176     const uint8_t *ds_table = (i_max == 64) ? x264_decimate_table8 : x264_decimate_table4;
 177     int i_score = 0;
 178     int idx = i_max - 1;
 179
 180     /* Yes, dct[idx-1] is guaranteed to be 32-bit aligned.  idx>=0 instead of 1 works correctly for the same reason */
 181     while( idx >= 0 && M32( &dct[idx-1] ) == 0 )
 182         idx -= 2;
 183     if( idx >= 0 && dct[idx] == 0 )
 184         idx--;
 185     while( idx >= 0 )
 186     {
 187         int i_run;
 188
 189         if( (unsigned)(dct[idx--] + 1) > 2 )
 190             return 9;
 191
 192         i_run = 0;
 193         while( idx >= 0 && dct[idx] == 0 )
 194         {
 195             idx--;
 196             i_run++;
 197         }
 198         i_score += ds_table[i_run];
 199     }
 200
 201     return i_score;
 202 }
 203
 204 static int x264_decimate_score15( int16_t *dct )
 205 {
 206     return x264_decimate_score_internal( dct+1, 15 );
 207 }
 208 static int x264_decimate_score16( int16_t *dct )
 209 {
 210     return x264_decimate_score_internal( dct, 16 );
 211 }
 212 static int x264_decimate_score64( int16_t *dct )
 213 {
 214     return x264_decimate_score_internal( dct, 64 );
 215 }
 216
 217 static int ALWAYS_INLINE x264_coeff_last_internal( int16_t *l, int i_count )
 218 {
 219     int i_last;
 220     for( i_last = i_count-1; i_last >= 3; i_last -= 4 )
 221         if( M64( l+i_last-3 ) )
 222             break;
 223     while( i_last >= 0 && l[i_last] == 0 )
 224         i_last--;
 225     return i_last;
 226 }
 227
 228 static int x264_coeff_last4( int16_t *l )
 229 {
 230     return x264_coeff_last_internal( l, 4 );
 231 }
 232 static int x264_coeff_last15( int16_t *l )
 233 {
 234     return x264_coeff_last_internal( l, 15 );
 235 }
 236 static int x264_coeff_last16( int16_t *l )
 237 {
 238     return x264_coeff_last_internal( l, 16 );
 239 }
 240 static int x264_coeff_last64( int16_t *l )
 241 {
 242     return x264_coeff_last_internal( l, 64 );
 243 }
 244
 245 #define level_run(num)\
 246 static int x264_coeff_level_run##num( int16_t *dct, x264_run_level_t *runlevel )\
 247 {\
 248     int i_last = runlevel->last = x264_coeff_last##num(dct);\
 249     int i_total = 0;\
 250     do\
 251     {\
 252         int r = 0;\
 253         runlevel->level[i_total] = dct[i_last];\
 254         while( --i_last >= 0 && dct[i_last] == 0 )\
 255             r++;\
 256         runlevel->run[i_total++] = r;\
 257     } while( i_last >= 0 );\
 258     return i_total;\
 259 }
 260
 261 level_run(4)
 262 level_run(15)
 263 level_run(16)
 264
 265
 266 void x264_quant_init( x264_t *h, int cpu, x264_quant_function_t *pf )
 267 {
 268     pf->quant_8x8 = quant_8x8;
 269     pf->quant_4x4 = quant_4x4;
 270     pf->quant_4x4_dc = quant_4x4_dc;
 271     pf->quant_2x2_dc = quant_2x2_dc;
 272
 273     pf->dequant_4x4 = dequant_4x4;
 274     pf->dequant_4x4_dc = dequant_4x4_dc;
 275     pf->dequant_8x8 = dequant_8x8;
 276
 277     pf->denoise_dct = x264_denoise_dct;
 278     pf->decimate_score15 = x264_decimate_score15;
 279     pf->decimate_score16 = x264_decimate_score16;
 280     pf->decimate_score64 = x264_decimate_score64;
 281
 282     pf->coeff_last[DCT_CHROMA_DC] = x264_coeff_last4;
 283     pf->coeff_last[  DCT_LUMA_AC] = x264_coeff_last15;
 284     pf->coeff_last[ DCT_LUMA_4x4] = x264_coeff_last16;
 285     pf->coeff_last[ DCT_LUMA_8x8] = x264_coeff_last64;
 286     pf->coeff_level_run[DCT_CHROMA_DC] = x264_coeff_level_run4;
 287     pf->coeff_level_run[  DCT_LUMA_AC] = x264_coeff_level_run15;
 288     pf->coeff_level_run[ DCT_LUMA_4x4] = x264_coeff_level_run16;
 289
 290 #ifdef HAVE_MMX
 291     if( cpu&X264_CPU_MMX )
 292     {
 293 #ifdef ARCH_X86
 294         pf->quant_4x4 = x264_quant_4x4_mmx;
 295         pf->quant_8x8 = x264_quant_8x8_mmx;
 296         pf->dequant_4x4 = x264_dequant_4x4_mmx;
 297         pf->dequant_4x4_dc = x264_dequant_4x4dc_mmxext;
 298         pf->dequant_8x8 = x264_dequant_8x8_mmx;
 299         if( h->param.i_cqm_preset == X264_CQM_FLAT )
 300         {
 301             pf->dequant_4x4 = x264_dequant_4x4_flat16_mmx;
 302             pf->dequant_8x8 = x264_dequant_8x8_flat16_mmx;
 303         }
 304         pf->denoise_dct = x264_denoise_dct_mmx;
 305 #endif
 306     }
 307
 308     if( cpu&X264_CPU_MMXEXT )
 309     {
 310         pf->quant_2x2_dc = x264_quant_2x2_dc_mmxext;
 311 #ifdef ARCH_X86
 312         pf->quant_4x4_dc = x264_quant_4x4_dc_mmxext;
 313         pf->decimate_score15 = x264_decimate_score15_mmxext;
 314         pf->decimate_score16 = x264_decimate_score16_mmxext;
 315         pf->decimate_score64 = x264_decimate_score64_mmxext;
 316         pf->coeff_last[  DCT_LUMA_AC] = x264_coeff_last15_mmxext;
 317         pf->coeff_last[ DCT_LUMA_4x4] = x264_coeff_last16_mmxext;
 318         pf->coeff_last[ DCT_LUMA_8x8] = x264_coeff_last64_mmxext;
 319         pf->coeff_level_run[  DCT_LUMA_AC] = x264_coeff_level_run15_mmxext;
 320         pf->coeff_level_run[ DCT_LUMA_4x4] = x264_coeff_level_run16_mmxext;
 321 #endif
 322         pf->coeff_last[DCT_CHROMA_DC] = x264_coeff_last4_mmxext;
 323         pf->coeff_level_run[DCT_CHROMA_DC] = x264_coeff_level_run4_mmxext;
 324         if( cpu&X264_CPU_LZCNT )
 325         {
 326             pf->coeff_last[DCT_CHROMA_DC] = x264_coeff_last4_mmxext_lzcnt;
 327             pf->coeff_level_run[DCT_CHROMA_DC] = x264_coeff_level_run4_mmxext_lzcnt;
 328         }
 329     }
 330
 331     if( cpu&X264_CPU_SSE2 )
 332     {
 333         pf->quant_4x4_dc = x264_quant_4x4_dc_sse2;
 334         pf->quant_4x4 = x264_quant_4x4_sse2;
 335         pf->quant_8x8 = x264_quant_8x8_sse2;
 336         pf->dequant_4x4 = x264_dequant_4x4_sse2;
 337         pf->dequant_4x4_dc = x264_dequant_4x4dc_sse2;
 338         pf->dequant_8x8 = x264_dequant_8x8_sse2;
 339         if( h->param.i_cqm_preset == X264_CQM_FLAT )
 340         {
 341             pf->dequant_4x4 = x264_dequant_4x4_flat16_sse2;
 342             pf->dequant_8x8 = x264_dequant_8x8_flat16_sse2;
 343         }
 344         pf->denoise_dct = x264_denoise_dct_sse2;
 345         pf->decimate_score15 = x264_decimate_score15_sse2;
 346         pf->decimate_score16 = x264_decimate_score16_sse2;
 347         pf->decimate_score64 = x264_decimate_score64_sse2;
 348         pf->coeff_last[ DCT_LUMA_AC] = x264_coeff_last15_sse2;
 349         pf->coeff_last[DCT_LUMA_4x4] = x264_coeff_last16_sse2;
 350         pf->coeff_last[DCT_LUMA_8x8] = x264_coeff_last64_sse2;
 351         pf->coeff_level_run[ DCT_LUMA_AC] = x264_coeff_level_run15_sse2;
 352         pf->coeff_level_run[DCT_LUMA_4x4] = x264_coeff_level_run16_sse2;
 353         if( cpu&X264_CPU_LZCNT )
 354         {
 355             pf->coeff_last[ DCT_LUMA_AC] = x264_coeff_last15_sse2_lzcnt;
 356             pf->coeff_last[DCT_LUMA_4x4] = x264_coeff_last16_sse2_lzcnt;
 357             pf->coeff_last[DCT_LUMA_8x8] = x264_coeff_last64_sse2_lzcnt;
 358             pf->coeff_level_run[ DCT_LUMA_AC] = x264_coeff_level_run15_sse2_lzcnt;
 359             pf->coeff_level_run[DCT_LUMA_4x4] = x264_coeff_level_run16_sse2_lzcnt;
 360         }
 361     }
 362
 363     if( cpu&X264_CPU_SSSE3 )
 364     {
 365         pf->quant_2x2_dc = x264_quant_2x2_dc_ssse3;
 366         pf->quant_4x4_dc = x264_quant_4x4_dc_ssse3;
 367         pf->quant_4x4 = x264_quant_4x4_ssse3;
 368         pf->quant_8x8 = x264_quant_8x8_ssse3;
 369         pf->denoise_dct = x264_denoise_dct_ssse3;
 370         pf->decimate_score15 = x264_decimate_score15_ssse3;
 371         pf->decimate_score16 = x264_decimate_score16_ssse3;
 372         pf->decimate_score64 = x264_decimate_score64_ssse3;
 373     }
 374
 375     if( cpu&X264_CPU_SSE4 )
 376     {
 377         pf->quant_4x4_dc = x264_quant_4x4_dc_sse4;
 378         pf->quant_4x4 = x264_quant_4x4_sse4;
 379         pf->quant_8x8 = x264_quant_8x8_sse4;
 380     }
 381 #endif // HAVE_MMX
 382
 383 #ifdef HAVE_ALTIVEC
 384     if( cpu&X264_CPU_ALTIVEC ) {
 385         pf->quant_2x2_dc = x264_quant_2x2_dc_altivec;
 386         pf->quant_4x4_dc = x264_quant_4x4_dc_altivec;
 387         pf->quant_4x4 = x264_quant_4x4_altivec;
 388         pf->quant_8x8 = x264_quant_8x8_altivec;
 389
 390         pf->dequant_4x4 = x264_dequant_4x4_altivec;
 391         pf->dequant_8x8 = x264_dequant_8x8_altivec;
 392     }
 393 #endif
 394
 395 #ifdef HAVE_ARMV6
 396     if( cpu&X264_CPU_ARMV6 )
 397         pf->coeff_last[DCT_CHROMA_DC] = x264_coeff_last4_arm;
 398
 399     if( cpu&X264_CPU_NEON )
 400     {
 401         pf->quant_2x2_dc   = x264_quant_2x2_dc_neon;
 402         pf->quant_4x4      = x264_quant_4x4_neon;
 403         pf->quant_4x4_dc   = x264_quant_4x4_dc_neon;
 404         pf->quant_8x8      = x264_quant_8x8_neon;
 405         pf->dequant_4x4    = x264_dequant_4x4_neon;
 406         pf->dequant_4x4_dc = x264_dequant_4x4_dc_neon;
 407         pf->dequant_8x8    = x264_dequant_8x8_neon;
 408         pf->coeff_last[ DCT_LUMA_AC] = x264_coeff_last15_neon;
 409         pf->coeff_last[DCT_LUMA_4x4] = x264_coeff_last16_neon;
 410         pf->coeff_last[DCT_LUMA_8x8] = x264_coeff_last64_neon;
 411     }
 412 #endif
 413     pf->coeff_last[  DCT_LUMA_DC] = pf->coeff_last[DCT_LUMA_4x4];
 414     pf->coeff_last[DCT_CHROMA_AC] = pf->coeff_last[ DCT_LUMA_AC];
 415     pf->coeff_level_run[  DCT_LUMA_DC] = pf->coeff_level_run[DCT_LUMA_4x4];
 416     pf->coeff_level_run[DCT_CHROMA_AC] = pf->coeff_level_run[ DCT_LUMA_AC];
 417 }