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 }
  40
  41 static void quant_8x8( int16_t dct[8][8], uint16_t mf[64], uint16_t bias[64] )
  42 {
  43     int i;
  44     for( i = 0; i < 64; i++ )
  45         QUANT_ONE( dct[0][i], mf[i], bias[i] );
  46 }
  47
  48 static void quant_4x4( int16_t dct[4][4], uint16_t mf[16], uint16_t bias[16] )
  49 {
  50     int i;
  51     for( i = 0; i < 16; i++ )
  52         QUANT_ONE( dct[0][i], mf[i], bias[i] );
  53 }
  54
  55 static void quant_4x4_dc( int16_t dct[4][4], int mf, int bias )
  56 {
  57     int i;
  58     for( i = 0; i < 16; i++ )
  59         QUANT_ONE( dct[0][i], mf, bias );
  60 }
  61
  62 static void quant_2x2_dc( int16_t dct[2][2], int mf, int bias )
  63 {
  64     QUANT_ONE( dct[0][0], mf, bias );
  65     QUANT_ONE( dct[0][1], mf, bias );
  66     QUANT_ONE( dct[0][2], mf, bias );
  67     QUANT_ONE( dct[0][3], mf, bias );
  68 }
  69
  70 #define DEQUANT_SHL( x ) \
  71     dct[y][x] = ( dct[y][x] * dequant_mf[i_mf][y][x] ) << i_qbits
  72
  73 #define DEQUANT_SHR( x ) \
  74     dct[y][x] = ( dct[y][x] * dequant_mf[i_mf][y][x] + f ) >> (-i_qbits)
  75
  76 static void dequant_4x4( int16_t dct[4][4], int dequant_mf[6][4][4], int i_qp )
  77 {
  78     const int i_mf = i_qp%6;
  79     const int i_qbits = i_qp/6 - 4;
  80     int y;
  81
  82     if( i_qbits >= 0 )
  83     {
  84         for( y = 0; y < 4; y++ )
  85         {
  86             DEQUANT_SHL( 0 );
  87             DEQUANT_SHL( 1 );
  88             DEQUANT_SHL( 2 );
  89             DEQUANT_SHL( 3 );
  90         }
  91     }
  92     else
  93     {
  94         const int f = 1 << (-i_qbits-1);
  95         for( y = 0; y < 4; y++ )
  96         {
  97             DEQUANT_SHR( 0 );
  98             DEQUANT_SHR( 1 );
  99             DEQUANT_SHR( 2 );
 100             DEQUANT_SHR( 3 );
 101         }
 102     }
 103 }
 104
 105 static void dequant_8x8( int16_t dct[8][8], int dequant_mf[6][8][8], int i_qp )
 106 {
 107     const int i_mf = i_qp%6;
 108     const int i_qbits = i_qp/6 - 6;
 109     int y;
 110
 111     if( i_qbits >= 0 )
 112     {
 113         for( y = 0; y < 8; y++ )
 114         {
 115             DEQUANT_SHL( 0 );
 116             DEQUANT_SHL( 1 );
 117             DEQUANT_SHL( 2 );
 118             DEQUANT_SHL( 3 );
 119             DEQUANT_SHL( 4 );
 120             DEQUANT_SHL( 5 );
 121             DEQUANT_SHL( 6 );
 122             DEQUANT_SHL( 7 );
 123         }
 124     }
 125     else
 126     {
 127         const int f = 1 << (-i_qbits-1);
 128         for( y = 0; y < 8; y++ )
 129         {
 130             DEQUANT_SHR( 0 );
 131             DEQUANT_SHR( 1 );
 132             DEQUANT_SHR( 2 );
 133             DEQUANT_SHR( 3 );
 134             DEQUANT_SHR( 4 );
 135             DEQUANT_SHR( 5 );
 136             DEQUANT_SHR( 6 );
 137             DEQUANT_SHR( 7 );
 138         }
 139     }
 140 }
 141
 142 static void dequant_4x4_dc( int16_t dct[4][4], int dequant_mf[6][4][4], int i_qp )
 143 {
 144     const int i_qbits = i_qp/6 - 6;
 145     int y;
 146
 147     if( i_qbits >= 0 )
 148     {
 149         const int i_dmf = dequant_mf[i_qp%6][0][0] << i_qbits;
 150
 151         for( y = 0; y < 4; y++ )
 152         {
 153             dct[y][0] *= i_dmf;
 154             dct[y][1] *= i_dmf;
 155             dct[y][2] *= i_dmf;
 156             dct[y][3] *= i_dmf;
 157         }
 158     }
 159     else
 160     {
 161         const int i_dmf = dequant_mf[i_qp%6][0][0];
 162         const int f = 1 << (-i_qbits-1);
 163
 164         for( y = 0; y < 4; y++ )
 165         {
 166             dct[y][0] = ( dct[y][0] * i_dmf + f ) >> (-i_qbits);
 167             dct[y][1] = ( dct[y][1] * i_dmf + f ) >> (-i_qbits);
 168             dct[y][2] = ( dct[y][2] * i_dmf + f ) >> (-i_qbits);
 169             dct[y][3] = ( dct[y][3] * i_dmf + f ) >> (-i_qbits);
 170         }
 171     }
 172 }
 173
 174 static void x264_denoise_dct( int16_t *dct, uint32_t *sum, uint16_t *offset, int size )
 175 {
 176     int i;
 177     for( i=1; i<size; i++ )
 178     {
 179         int level = dct[i];
 180         int sign = level>>15;
 181         level = (level+sign)^sign;
 182         sum[i] += level;
 183         level -= offset[i];
 184         dct[i] = level<0 ? 0 : (level^sign)-sign;
 185     }
 186 }
 187
 188 /* (ref: JVT-B118)
 189  * x264_mb_decimate_score: given dct coeffs it returns a score to see if we could empty this dct coeffs
 190  * to 0 (low score means set it to null)
 191  * Used in inter macroblock (luma and chroma)
 192  *  luma: for a 8x8 block: if score < 4 -> null
 193  *        for the complete mb: if score < 6 -> null
 194  *  chroma: for the complete mb: if score < 7 -> null
 195  */
 196
 197 const uint8_t x264_decimate_table4[16] = {
 198     3,2,2,1,1,1,0,0,0,0,0,0,0,0,0,0 };
 199 const uint8_t x264_decimate_table8[64] = {
 200     3,3,3,3,2,2,2,2,2,2,2,2,1,1,1,1,
 201     1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,
 202     0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
 203     0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 };
 204
 205 static int ALWAYS_INLINE x264_decimate_score_internal( int16_t *dct, int i_max )
 206 {
 207     const uint8_t *ds_table = (i_max == 64) ? x264_decimate_table8 : x264_decimate_table4;
 208     int i_score = 0;
 209     int idx = i_max - 1;
 210
 211     /* Yes, dct[idx-1] is guaranteed to be 32-bit aligned.  idx>=0 instead of 1 works correctly for the same reason */
 212     while( idx >= 0 && *(uint32_t*)&dct[idx-1] == 0 )
 213         idx -= 2;
 214     if( idx >= 0 && dct[idx] == 0 )
 215         idx--;
 216     while( idx >= 0 )
 217     {
 218         int i_run;
 219
 220         if( (unsigned)(dct[idx--] + 1) > 2 )
 221             return 9;
 222
 223         i_run = 0;
 224         while( idx >= 0 && dct[idx] == 0 )
 225         {
 226             idx--;
 227             i_run++;
 228         }
 229         i_score += ds_table[i_run];
 230     }
 231
 232     return i_score;
 233 }
 234
 235 static int x264_decimate_score15( int16_t *dct )
 236 {
 237     return x264_decimate_score_internal( dct+1, 15 );
 238 }
 239 static int x264_decimate_score16( int16_t *dct )
 240 {
 241     return x264_decimate_score_internal( dct, 16 );
 242 }
 243 static int x264_decimate_score64( int16_t *dct )
 244 {
 245     return x264_decimate_score_internal( dct, 64 );
 246 }
 247
 248 static int ALWAYS_INLINE x264_coeff_last_internal( int16_t *l, int i_count )
 249 {
 250     int i_last;
 251     for( i_last = i_count-1; i_last >= 3; i_last -= 4 )
 252         if( *(uint64_t*)(l+i_last-3) )
 253             break;
 254     while( i_last >= 0 && l[i_last] == 0 )
 255         i_last--;
 256     return i_last;
 257 }
 258
 259 static int x264_coeff_last4( int16_t *l )
 260 {
 261     return x264_coeff_last_internal( l, 4 );
 262 }
 263 static int x264_coeff_last15( int16_t *l )
 264 {
 265     return x264_coeff_last_internal( l, 15 );
 266 }
 267 static int x264_coeff_last16( int16_t *l )
 268 {
 269     return x264_coeff_last_internal( l, 16 );
 270 }
 271 static int x264_coeff_last64( int16_t *l )
 272 {
 273     return x264_coeff_last_internal( l, 64 );
 274 }
 275
 276 void x264_quant_init( x264_t *h, int cpu, x264_quant_function_t *pf )
 277 {
 278     pf->quant_8x8 = quant_8x8;
 279     pf->quant_4x4 = quant_4x4;
 280     pf->quant_4x4_dc = quant_4x4_dc;
 281     pf->quant_2x2_dc = quant_2x2_dc;
 282
 283     pf->dequant_4x4 = dequant_4x4;
 284     pf->dequant_4x4_dc = dequant_4x4_dc;
 285     pf->dequant_8x8 = dequant_8x8;
 286
 287     pf->denoise_dct = x264_denoise_dct;
 288     pf->decimate_score15 = x264_decimate_score15;
 289     pf->decimate_score16 = x264_decimate_score16;
 290     pf->decimate_score64 = x264_decimate_score64;
 291
 292     pf->coeff_last[DCT_CHROMA_DC] = x264_coeff_last4;
 293     pf->coeff_last[  DCT_LUMA_AC] = x264_coeff_last15;
 294     pf->coeff_last[ DCT_LUMA_4x4] = x264_coeff_last16;
 295     pf->coeff_last[ DCT_LUMA_8x8] = x264_coeff_last64;
 296
 297 #ifdef HAVE_MMX
 298     if( cpu&X264_CPU_MMX )
 299     {
 300 #ifdef ARCH_X86
 301         pf->quant_4x4 = x264_quant_4x4_mmx;
 302         pf->quant_8x8 = x264_quant_8x8_mmx;
 303         pf->dequant_4x4 = x264_dequant_4x4_mmx;
 304         pf->dequant_4x4_dc = x264_dequant_4x4dc_mmxext;
 305         pf->dequant_8x8 = x264_dequant_8x8_mmx;
 306         if( h->param.i_cqm_preset == X264_CQM_FLAT )
 307         {
 308             pf->dequant_4x4 = x264_dequant_4x4_flat16_mmx;
 309             pf->dequant_8x8 = x264_dequant_8x8_flat16_mmx;
 310         }
 311         pf->denoise_dct = x264_denoise_dct_mmx;
 312 #endif
 313     }
 314
 315     if( cpu&X264_CPU_MMXEXT )
 316     {
 317         pf->quant_2x2_dc = x264_quant_2x2_dc_mmxext;
 318 #ifdef ARCH_X86
 319         pf->quant_4x4_dc = x264_quant_4x4_dc_mmxext;
 320         pf->decimate_score15 = x264_decimate_score15_mmxext;
 321         pf->decimate_score16 = x264_decimate_score16_mmxext;
 322         pf->decimate_score64 = x264_decimate_score64_mmxext;
 323         pf->coeff_last[  DCT_LUMA_AC] = x264_coeff_last15_mmxext;
 324         pf->coeff_last[ DCT_LUMA_4x4] = x264_coeff_last16_mmxext;
 325         pf->coeff_last[ DCT_LUMA_8x8] = x264_coeff_last64_mmxext;
 326 #endif
 327         pf->coeff_last[DCT_CHROMA_DC] = x264_coeff_last4_mmxext;
 328     }
 329
 330     if( cpu&X264_CPU_SSE2 )
 331     {
 332         pf->quant_4x4_dc = x264_quant_4x4_dc_sse2;
 333         pf->quant_4x4 = x264_quant_4x4_sse2;
 334         pf->quant_8x8 = x264_quant_8x8_sse2;
 335         pf->dequant_4x4 = x264_dequant_4x4_sse2;
 336         pf->dequant_4x4_dc = x264_dequant_4x4dc_sse2;
 337         pf->dequant_8x8 = x264_dequant_8x8_sse2;
 338         if( h->param.i_cqm_preset == X264_CQM_FLAT )
 339         {
 340             pf->dequant_4x4 = x264_dequant_4x4_flat16_sse2;
 341             pf->dequant_8x8 = x264_dequant_8x8_flat16_sse2;
 342         }
 343         pf->denoise_dct = x264_denoise_dct_sse2;
 344         pf->decimate_score15 = x264_decimate_score15_sse2;
 345         pf->decimate_score16 = x264_decimate_score16_sse2;
 346         pf->decimate_score64 = x264_decimate_score64_sse2;
 347         pf->coeff_last[ DCT_LUMA_AC] = x264_coeff_last15_sse2;
 348         pf->coeff_last[DCT_LUMA_4x4] = x264_coeff_last16_sse2;
 349         pf->coeff_last[DCT_LUMA_8x8] = x264_coeff_last64_sse2;
 350     }
 351
 352     if( cpu&X264_CPU_SSSE3 )
 353     {
 354         pf->quant_2x2_dc = x264_quant_2x2_dc_ssse3;
 355         pf->quant_4x4_dc = x264_quant_4x4_dc_ssse3;
 356         pf->quant_4x4 = x264_quant_4x4_ssse3;
 357         pf->quant_8x8 = x264_quant_8x8_ssse3;
 358         pf->denoise_dct = x264_denoise_dct_ssse3;
 359         pf->decimate_score15 = x264_decimate_score15_ssse3;
 360         pf->decimate_score16 = x264_decimate_score16_ssse3;
 361         pf->decimate_score64 = x264_decimate_score64_ssse3;
 362     }
 363 #endif // HAVE_MMX
 364
 365 #ifdef ARCH_PPC
 366     if( cpu&X264_CPU_ALTIVEC ) {
 367         pf->quant_2x2_dc = x264_quant_2x2_dc_altivec;
 368         pf->quant_4x4_dc = x264_quant_4x4_dc_altivec;
 369         pf->quant_4x4 = x264_quant_4x4_altivec;
 370         pf->quant_8x8 = x264_quant_8x8_altivec;
 371
 372         pf->dequant_4x4 = x264_dequant_4x4_altivec;
 373         pf->dequant_8x8 = x264_dequant_8x8_altivec;
 374     }
 375 #endif
 376     pf->coeff_last[  DCT_LUMA_DC] = pf->coeff_last[DCT_LUMA_4x4];
 377     pf->coeff_last[DCT_CHROMA_AC] = pf->coeff_last[ DCT_LUMA_AC];
 378 }