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

   1 /*****************************************************************************
   2  * pixel.c: h264 encoder
   3  *****************************************************************************
   4  * Copyright (C) 2003-2008 x264 project
   5  *
   6  * Authors: Loren Merritt <lorenm@u.washington.edu>
   7  *          Laurent Aimar <fenrir@via.ecp.fr>
   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/pixel.h"
  28 #endif
  29 #ifdef ARCH_PPC
  30 #   include "ppc/pixel.h"
  31 #endif
  32 #ifdef ARCH_ARM
  33 #   include "arm/pixel.h"
  34 #endif
  35 #ifdef ARCH_UltraSparc
  36 #   include "sparc/pixel.h"
  37 #endif
  38
  39
  40 /****************************************************************************
  41  * pixel_sad_WxH
  42  ****************************************************************************/
  43 #define PIXEL_SAD_C( name, lx, ly ) \
  44 static int name( uint8_t *pix1, int i_stride_pix1,  \
  45                  uint8_t *pix2, int i_stride_pix2 ) \
  46 {                                                   \
  47     int i_sum = 0;                                  \
  48     int x, y;                                       \
  49     for( y = 0; y < ly; y++ )                       \
  50     {                                               \
  51         for( x = 0; x < lx; x++ )                   \
  52         {                                           \
  53             i_sum += abs( pix1[x] - pix2[x] );      \
  54         }                                           \
  55         pix1 += i_stride_pix1;                      \
  56         pix2 += i_stride_pix2;                      \
  57     }                                               \
  58     return i_sum;                                   \
  59 }
  60
  61
  62 PIXEL_SAD_C( x264_pixel_sad_16x16, 16, 16 )
  63 PIXEL_SAD_C( x264_pixel_sad_16x8,  16,  8 )
  64 PIXEL_SAD_C( x264_pixel_sad_8x16,   8, 16 )
  65 PIXEL_SAD_C( x264_pixel_sad_8x8,    8,  8 )
  66 PIXEL_SAD_C( x264_pixel_sad_8x4,    8,  4 )
  67 PIXEL_SAD_C( x264_pixel_sad_4x8,    4,  8 )
  68 PIXEL_SAD_C( x264_pixel_sad_4x4,    4,  4 )
  69
  70
  71 /****************************************************************************
  72  * pixel_ssd_WxH
  73  ****************************************************************************/
  74 #define PIXEL_SSD_C( name, lx, ly ) \
  75 static int name( uint8_t *pix1, int i_stride_pix1,  \
  76                  uint8_t *pix2, int i_stride_pix2 ) \
  77 {                                                   \
  78     int i_sum = 0;                                  \
  79     int x, y;                                       \
  80     for( y = 0; y < ly; y++ )                       \
  81     {                                               \
  82         for( x = 0; x < lx; x++ )                   \
  83         {                                           \
  84             int d = pix1[x] - pix2[x];              \
  85             i_sum += d*d;                           \
  86         }                                           \
  87         pix1 += i_stride_pix1;                      \
  88         pix2 += i_stride_pix2;                      \
  89     }                                               \
  90     return i_sum;                                   \
  91 }
  92
  93 PIXEL_SSD_C( x264_pixel_ssd_16x16, 16, 16 )
  94 PIXEL_SSD_C( x264_pixel_ssd_16x8,  16,  8 )
  95 PIXEL_SSD_C( x264_pixel_ssd_8x16,   8, 16 )
  96 PIXEL_SSD_C( x264_pixel_ssd_8x8,    8,  8 )
  97 PIXEL_SSD_C( x264_pixel_ssd_8x4,    8,  4 )
  98 PIXEL_SSD_C( x264_pixel_ssd_4x8,    4,  8 )
  99 PIXEL_SSD_C( x264_pixel_ssd_4x4,    4,  4 )
 100
 101 int64_t x264_pixel_ssd_wxh( x264_pixel_function_t *pf, uint8_t *pix1, int i_pix1, uint8_t *pix2, int i_pix2, int i_width, int i_height )
 102 {
 103     int64_t i_ssd = 0;
 104     int x, y;
 105     int align = !(((intptr_t)pix1 | (intptr_t)pix2 | i_pix1 | i_pix2) & 15);
 106
 107 #define SSD(size) i_ssd += pf->ssd[size]( pix1 + y*i_pix1 + x, i_pix1, \
 108                                           pix2 + y*i_pix2 + x, i_pix2 );
 109     for( y = 0; y < i_height-15; y += 16 )
 110     {
 111         x = 0;
 112         if( align )
 113             for( ; x < i_width-15; x += 16 )
 114                 SSD(PIXEL_16x16);
 115         for( ; x < i_width-7; x += 8 )
 116             SSD(PIXEL_8x16);
 117     }
 118     if( y < i_height-7 )
 119         for( x = 0; x < i_width-7; x += 8 )
 120             SSD(PIXEL_8x8);
 121 #undef SSD
 122
 123 #define SSD1 { int d = pix1[y*i_pix1+x] - pix2[y*i_pix2+x]; i_ssd += d*d; }
 124     if( i_width % 8 != 0 )
 125     {
 126         for( y = 0; y < (i_height & ~7); y++ )
 127             for( x = i_width & ~7; x < i_width; x++ )
 128                 SSD1;
 129     }
 130     if( i_height % 8 != 0 )
 131     {
 132         for( y = i_height & ~7; y < i_height; y++ )
 133             for( x = 0; x < i_width; x++ )
 134                 SSD1;
 135     }
 136 #undef SSD1
 137
 138     return i_ssd;
 139 }
 140
 141
 142 /****************************************************************************
 143  * pixel_var_wxh
 144  ****************************************************************************/
 145 #define PIXEL_VAR_C( name, w, shift ) \
 146 static int name( uint8_t *pix, int i_stride ) \
 147 {                                             \
 148     uint32_t var = 0, sum = 0, sqr = 0;       \
 149     int x, y;                                 \
 150     for( y = 0; y < w; y++ )                  \
 151     {                                         \
 152         for( x = 0; x < w; x++ )              \
 153         {                                     \
 154             sum += pix[x];                    \
 155             sqr += pix[x] * pix[x];           \
 156         }                                     \
 157         pix += i_stride;                      \
 158     }                                         \
 159     var = sqr - (sum * sum >> shift);         \
 160     return var;                               \
 161 }
 162
 163 PIXEL_VAR_C( x264_pixel_var_16x16, 16, 8 )
 164 PIXEL_VAR_C( x264_pixel_var_8x8,    8, 6 )
 165
 166 /****************************************************************************
 167  * pixel_var2_wxh
 168  ****************************************************************************/
 169 static int pixel_var2_8x8( uint8_t *pix1, int i_stride1, uint8_t *pix2, int i_stride2, int *ssd )
 170 {
 171     uint32_t var = 0, sum = 0, sqr = 0;
 172     int x, y;
 173     for( y = 0; y < 8; y++ )
 174     {
 175         for( x = 0; x < 8; x++ )
 176         {
 177             int diff = pix1[x] - pix2[x];
 178             sum += diff;
 179             sqr += diff * diff;
 180         }
 181         pix1 += i_stride1;
 182         pix2 += i_stride2;
 183     }
 184     sum = abs(sum);
 185     var = sqr - (sum * sum >> 6);
 186     *ssd = sqr;
 187     return var;
 188 }
 189
 190
 191 #define HADAMARD4(d0,d1,d2,d3,s0,s1,s2,s3) {\
 192     int t0 = s0 + s1;\
 193     int t1 = s0 - s1;\
 194     int t2 = s2 + s3;\
 195     int t3 = s2 - s3;\
 196     d0 = t0 + t2;\
 197     d2 = t0 - t2;\
 198     d1 = t1 + t3;\
 199     d3 = t1 - t3;\
 200 }
 201
 202 // in: a pseudo-simd number of the form x+(y<<16)
 203 // return: abs(x)+(abs(y)<<16)
 204 static ALWAYS_INLINE uint32_t abs2( uint32_t a )
 205 {
 206     uint32_t s = ((a>>15)&0x10001)*0xffff;
 207     return (a+s)^s;
 208 }
 209
 210 /****************************************************************************
 211  * pixel_satd_WxH: sum of 4x4 Hadamard transformed differences
 212  ****************************************************************************/
 213
 214 static NOINLINE int x264_pixel_satd_4x4( uint8_t *pix1, int i_pix1, uint8_t *pix2, int i_pix2 )
 215 {
 216     uint32_t tmp[4][2];
 217     uint32_t a0,a1,a2,a3,b0,b1;
 218     int sum=0, i;
 219     for( i=0; i<4; i++, pix1+=i_pix1, pix2+=i_pix2 )
 220     {
 221         a0 = pix1[0] - pix2[0];
 222         a1 = pix1[1] - pix2[1];
 223         b0 = (a0+a1) + ((a0-a1)<<16);
 224         a2 = pix1[2] - pix2[2];
 225         a3 = pix1[3] - pix2[3];
 226         b1 = (a2+a3) + ((a2-a3)<<16);
 227         tmp[i][0] = b0 + b1;
 228         tmp[i][1] = b0 - b1;
 229     }
 230     for( i=0; i<2; i++ )
 231     {
 232         HADAMARD4( a0,a1,a2,a3, tmp[0][i], tmp[1][i], tmp[2][i], tmp[3][i] );
 233         a0 = abs2(a0) + abs2(a1) + abs2(a2) + abs2(a3);
 234         sum += ((uint16_t)a0) + (a0>>16);
 235     }
 236     return sum >> 1;
 237 }
 238
 239 static NOINLINE int x264_pixel_satd_8x4( uint8_t *pix1, int i_pix1, uint8_t *pix2, int i_pix2 )
 240 {
 241     uint32_t tmp[4][4];
 242     uint32_t a0,a1,a2,a3;
 243     int sum=0, i;
 244     for( i=0; i<4; i++, pix1+=i_pix1, pix2+=i_pix2 )
 245     {
 246         a0 = (pix1[0] - pix2[0]) + ((pix1[4] - pix2[4]) << 16);
 247         a1 = (pix1[1] - pix2[1]) + ((pix1[5] - pix2[5]) << 16);
 248         a2 = (pix1[2] - pix2[2]) + ((pix1[6] - pix2[6]) << 16);
 249         a3 = (pix1[3] - pix2[3]) + ((pix1[7] - pix2[7]) << 16);
 250         HADAMARD4( tmp[i][0], tmp[i][1], tmp[i][2], tmp[i][3], a0,a1,a2,a3 );
 251     }
 252     for( i=0; i<4; i++ )
 253     {
 254         HADAMARD4( a0,a1,a2,a3, tmp[0][i], tmp[1][i], tmp[2][i], tmp[3][i] );
 255         sum += abs2(a0) + abs2(a1) + abs2(a2) + abs2(a3);
 256     }
 257     return (((uint16_t)sum) + ((uint32_t)sum>>16)) >> 1;
 258 }
 259
 260 #define PIXEL_SATD_C( w, h, sub )\
 261 static int x264_pixel_satd_##w##x##h( uint8_t *pix1, int i_pix1, uint8_t *pix2, int i_pix2 )\
 262 {\
 263     int sum = sub( pix1, i_pix1, pix2, i_pix2 )\
 264             + sub( pix1+4*i_pix1, i_pix1, pix2+4*i_pix2, i_pix2 );\
 265     if( w==16 )\
 266         sum+= sub( pix1+8, i_pix1, pix2+8, i_pix2 )\
 267             + sub( pix1+8+4*i_pix1, i_pix1, pix2+8+4*i_pix2, i_pix2 );\
 268     if( h==16 )\
 269         sum+= sub( pix1+8*i_pix1, i_pix1, pix2+8*i_pix2, i_pix2 )\
 270             + sub( pix1+12*i_pix1, i_pix1, pix2+12*i_pix2, i_pix2 );\
 271     if( w==16 && h==16 )\
 272         sum+= sub( pix1+8+8*i_pix1, i_pix1, pix2+8+8*i_pix2, i_pix2 )\
 273             + sub( pix1+8+12*i_pix1, i_pix1, pix2+8+12*i_pix2, i_pix2 );\
 274     return sum;\
 275 }
 276 PIXEL_SATD_C( 16, 16, x264_pixel_satd_8x4 )
 277 PIXEL_SATD_C( 16, 8,  x264_pixel_satd_8x4 )
 278 PIXEL_SATD_C( 8,  16, x264_pixel_satd_8x4 )
 279 PIXEL_SATD_C( 8,  8,  x264_pixel_satd_8x4 )
 280 PIXEL_SATD_C( 4,  8,  x264_pixel_satd_4x4 )
 281
 282
 283 static NOINLINE int sa8d_8x8( uint8_t *pix1, int i_pix1, uint8_t *pix2, int i_pix2 )
 284 {
 285     uint32_t tmp[8][4];
 286     uint32_t a0,a1,a2,a3,a4,a5,a6,a7,b0,b1,b2,b3;
 287     int sum=0, i;
 288     for( i=0; i<8; i++, pix1+=i_pix1, pix2+=i_pix2 )
 289     {
 290         a0 = pix1[0] - pix2[0];
 291         a1 = pix1[1] - pix2[1];
 292         b0 = (a0+a1) + ((a0-a1)<<16);
 293         a2 = pix1[2] - pix2[2];
 294         a3 = pix1[3] - pix2[3];
 295         b1 = (a2+a3) + ((a2-a3)<<16);
 296         a4 = pix1[4] - pix2[4];
 297         a5 = pix1[5] - pix2[5];
 298         b2 = (a4+a5) + ((a4-a5)<<16);
 299         a6 = pix1[6] - pix2[6];
 300         a7 = pix1[7] - pix2[7];
 301         b3 = (a6+a7) + ((a6-a7)<<16);
 302         HADAMARD4( tmp[i][0], tmp[i][1], tmp[i][2], tmp[i][3], b0,b1,b2,b3 );
 303     }
 304     for( i=0; i<4; i++ )
 305     {
 306         HADAMARD4( a0,a1,a2,a3, tmp[0][i], tmp[1][i], tmp[2][i], tmp[3][i] );
 307         HADAMARD4( a4,a5,a6,a7, tmp[4][i], tmp[5][i], tmp[6][i], tmp[7][i] );
 308         b0  = abs2(a0+a4) + abs2(a0-a4);
 309         b0 += abs2(a1+a5) + abs2(a1-a5);
 310         b0 += abs2(a2+a6) + abs2(a2-a6);
 311         b0 += abs2(a3+a7) + abs2(a3-a7);
 312         sum += (uint16_t)b0 + (b0>>16);
 313     }
 314     return sum;
 315 }
 316
 317 static int x264_pixel_sa8d_8x8( uint8_t *pix1, int i_pix1, uint8_t *pix2, int i_pix2 )
 318 {
 319     int sum = sa8d_8x8( pix1, i_pix1, pix2, i_pix2 );
 320     return (sum+2)>>2;
 321 }
 322
 323 static int x264_pixel_sa8d_16x16( uint8_t *pix1, int i_pix1, uint8_t *pix2, int i_pix2 )
 324 {
 325     int sum = sa8d_8x8( pix1, i_pix1, pix2, i_pix2 )
 326             + sa8d_8x8( pix1+8, i_pix1, pix2+8, i_pix2 )
 327             + sa8d_8x8( pix1+8*i_pix1, i_pix1, pix2+8*i_pix2, i_pix2 )
 328             + sa8d_8x8( pix1+8+8*i_pix1, i_pix1, pix2+8+8*i_pix2, i_pix2 );
 329     return (sum+2)>>2;
 330 }
 331
 332
 333 static NOINLINE uint64_t pixel_hadamard_ac( uint8_t *pix, int stride )
 334 {
 335     uint32_t tmp[32];
 336     uint32_t a0,a1,a2,a3,dc;
 337     int sum4=0, sum8=0, i;
 338     for( i=0; i<8; i++, pix+=stride )
 339     {
 340         uint32_t *t = tmp + (i&3) + (i&4)*4;
 341         a0 = (pix[0]+pix[1]) + ((pix[0]-pix[1])<<16);
 342         a1 = (pix[2]+pix[3]) + ((pix[2]-pix[3])<<16);
 343         t[0] = a0 + a1;
 344         t[4] = a0 - a1;
 345         a2 = (pix[4]+pix[5]) + ((pix[4]-pix[5])<<16);
 346         a3 = (pix[6]+pix[7]) + ((pix[6]-pix[7])<<16);
 347         t[8] = a2 + a3;
 348         t[12] = a2 - a3;
 349     }
 350     for( i=0; i<8; i++ )
 351     {
 352         HADAMARD4( a0,a1,a2,a3, tmp[i*4+0], tmp[i*4+1], tmp[i*4+2], tmp[i*4+3] );
 353         tmp[i*4+0] = a0;
 354         tmp[i*4+1] = a1;
 355         tmp[i*4+2] = a2;
 356         tmp[i*4+3] = a3;
 357         sum4 += abs2(a0) + abs2(a1) + abs2(a2) + abs2(a3);
 358     }
 359     for( i=0; i<8; i++ )
 360     {
 361         HADAMARD4( a0,a1,a2,a3, tmp[i], tmp[8+i], tmp[16+i], tmp[24+i] );
 362         sum8 += abs2(a0) + abs2(a1) + abs2(a2) + abs2(a3);
 363     }
 364     dc = (uint16_t)(tmp[0] + tmp[8] + tmp[16] + tmp[24]);
 365     sum4 = (uint16_t)sum4 + ((uint32_t)sum4>>16) - dc;
 366     sum8 = (uint16_t)sum8 + ((uint32_t)sum8>>16) - dc;
 367     return ((uint64_t)sum8<<32) + sum4;
 368 }
 369
 370 #define HADAMARD_AC(w,h) \
 371 static uint64_t x264_pixel_hadamard_ac_##w##x##h( uint8_t *pix, int stride )\
 372 {\
 373     uint64_t sum = pixel_hadamard_ac( pix, stride );\
 374     if( w==16 )\
 375         sum += pixel_hadamard_ac( pix+8, stride );\
 376     if( h==16 )\
 377         sum += pixel_hadamard_ac( pix+8*stride, stride );\
 378     if( w==16 && h==16 )\
 379         sum += pixel_hadamard_ac( pix+8*stride+8, stride );\
 380     return ((sum>>34)<<32) + ((uint32_t)sum>>1);\
 381 }
 382 HADAMARD_AC( 16, 16 )
 383 HADAMARD_AC( 16, 8 )
 384 HADAMARD_AC( 8, 16 )
 385 HADAMARD_AC( 8, 8 )
 386
 387
 388 /****************************************************************************
 389  * pixel_sad_x4
 390  ****************************************************************************/
 391 #define SAD_X( size ) \
 392 static void x264_pixel_sad_x3_##size( uint8_t *fenc, uint8_t *pix0, uint8_t *pix1, uint8_t *pix2, int i_stride, int scores[3] )\
 393 {\
 394     scores[0] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix0, i_stride );\
 395     scores[1] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix1, i_stride );\
 396     scores[2] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix2, i_stride );\
 397 }\
 398 static void x264_pixel_sad_x4_##size( uint8_t *fenc, uint8_t *pix0, uint8_t *pix1, uint8_t *pix2, uint8_t *pix3, int i_stride, int scores[4] )\
 399 {\
 400     scores[0] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix0, i_stride );\
 401     scores[1] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix1, i_stride );\
 402     scores[2] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix2, i_stride );\
 403     scores[3] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix3, i_stride );\
 404 }
 405
 406 SAD_X( 16x16 )
 407 SAD_X( 16x8 )
 408 SAD_X( 8x16 )
 409 SAD_X( 8x8 )
 410 SAD_X( 8x4 )
 411 SAD_X( 4x8 )
 412 SAD_X( 4x4 )
 413
 414 #ifdef ARCH_UltraSparc
 415 SAD_X( 16x16_vis )
 416 SAD_X( 16x8_vis )
 417 SAD_X( 8x16_vis )
 418 SAD_X( 8x8_vis )
 419 #endif
 420
 421 /****************************************************************************
 422  * pixel_satd_x4
 423  * no faster than single satd, but needed for satd to be a drop-in replacement for sad
 424  ****************************************************************************/
 425
 426 #define SATD_X( size, cpu ) \
 427 static void x264_pixel_satd_x3_##size##cpu( uint8_t *fenc, uint8_t *pix0, uint8_t *pix1, uint8_t *pix2, int i_stride, int scores[3] )\
 428 {\
 429     scores[0] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix0, i_stride );\
 430     scores[1] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix1, i_stride );\
 431     scores[2] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix2, i_stride );\
 432 }\
 433 static void x264_pixel_satd_x4_##size##cpu( uint8_t *fenc, uint8_t *pix0, uint8_t *pix1, uint8_t *pix2, uint8_t *pix3, int i_stride, int scores[4] )\
 434 {\
 435     scores[0] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix0, i_stride );\
 436     scores[1] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix1, i_stride );\
 437     scores[2] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix2, i_stride );\
 438     scores[3] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix3, i_stride );\
 439 }
 440 #define SATD_X_DECL6( cpu )\
 441 SATD_X( 16x16, cpu )\
 442 SATD_X( 16x8, cpu )\
 443 SATD_X( 8x16, cpu )\
 444 SATD_X( 8x8, cpu )\
 445 SATD_X( 8x4, cpu )\
 446 SATD_X( 4x8, cpu )
 447 #define SATD_X_DECL7( cpu )\
 448 SATD_X_DECL6( cpu )\
 449 SATD_X( 4x4, cpu )
 450
 451 SATD_X_DECL7()
 452 #ifdef HAVE_MMX
 453 SATD_X_DECL7( _mmxext )
 454 SATD_X_DECL6( _sse2 )
 455 SATD_X_DECL7( _ssse3 )
 456 SATD_X_DECL7( _sse4 )
 457 #endif
 458
 459 #ifdef HAVE_ARMV6
 460 SATD_X_DECL7( _neon )
 461 #endif
 462
 463 /****************************************************************************
 464  * structural similarity metric
 465  ****************************************************************************/
 466 static void ssim_4x4x2_core( const uint8_t *pix1, int stride1,
 467                              const uint8_t *pix2, int stride2,
 468                              int sums[2][4])
 469 {
 470     int x, y, z;
 471     for(z=0; z<2; z++)
 472     {
 473         uint32_t s1=0, s2=0, ss=0, s12=0;
 474         for(y=0; y<4; y++)
 475             for(x=0; x<4; x++)
 476             {
 477                 int a = pix1[x+y*stride1];
 478                 int b = pix2[x+y*stride2];
 479                 s1  += a;
 480                 s2  += b;
 481                 ss  += a*a;
 482                 ss  += b*b;
 483                 s12 += a*b;
 484             }
 485         sums[z][0] = s1;
 486         sums[z][1] = s2;
 487         sums[z][2] = ss;
 488         sums[z][3] = s12;
 489         pix1 += 4;
 490         pix2 += 4;
 491     }
 492 }
 493
 494 static float ssim_end1( int s1, int s2, int ss, int s12 )
 495 {
 496     static const int ssim_c1 = (int)(.01*.01*255*255*64 + .5);
 497     static const int ssim_c2 = (int)(.03*.03*255*255*64*63 + .5);
 498     int vars = ss*64 - s1*s1 - s2*s2;
 499     int covar = s12*64 - s1*s2;
 500     return (float)(2*s1*s2 + ssim_c1) * (float)(2*covar + ssim_c2)\
 501            / ((float)(s1*s1 + s2*s2 + ssim_c1) * (float)(vars + ssim_c2));
 502 }
 503
 504 static float ssim_end4( int sum0[5][4], int sum1[5][4], int width )
 505 {
 506     int i;
 507     float ssim = 0.0;
 508     for( i = 0; i < width; i++ )
 509         ssim += ssim_end1( sum0[i][0] + sum0[i+1][0] + sum1[i][0] + sum1[i+1][0],
 510                            sum0[i][1] + sum0[i+1][1] + sum1[i][1] + sum1[i+1][1],
 511                            sum0[i][2] + sum0[i+1][2] + sum1[i][2] + sum1[i+1][2],
 512                            sum0[i][3] + sum0[i+1][3] + sum1[i][3] + sum1[i+1][3] );
 513     return ssim;
 514 }
 515
 516 float x264_pixel_ssim_wxh( x264_pixel_function_t *pf,
 517                            uint8_t *pix1, int stride1,
 518                            uint8_t *pix2, int stride2,
 519                            int width, int height, void *buf )
 520 {
 521     int x, y, z;
 522     float ssim = 0.0;
 523     int (*sum0)[4] = buf;
 524     int (*sum1)[4] = sum0 + width/4+3;
 525     width >>= 2;
 526     height >>= 2;
 527     z = 0;
 528     for( y = 1; y < height; y++ )
 529     {
 530         for( ; z <= y; z++ )
 531         {
 532             XCHG( void*, sum0, sum1 );
 533             for( x = 0; x < width; x+=2 )
 534                 pf->ssim_4x4x2_core( &pix1[4*(x+z*stride1)], stride1, &pix2[4*(x+z*stride2)], stride2, &sum0[x] );
 535         }
 536         for( x = 0; x < width-1; x += 4 )
 537             ssim += pf->ssim_end4( sum0+x, sum1+x, X264_MIN(4,width-x-1) );
 538     }
 539     return ssim;
 540 }
 541
 542
 543 /****************************************************************************
 544  * successive elimination
 545  ****************************************************************************/
 546 static int x264_pixel_ads4( int enc_dc[4], uint16_t *sums, int delta,
 547                             uint16_t *cost_mvx, int16_t *mvs, int width, int thresh )
 548 {
 549     int nmv=0, i;
 550     for( i=0; i<width; i++, sums++ )
 551     {
 552         int ads = abs( enc_dc[0] - sums[0] )
 553                 + abs( enc_dc[1] - sums[8] )
 554                 + abs( enc_dc[2] - sums[delta] )
 555                 + abs( enc_dc[3] - sums[delta+8] )
 556                 + cost_mvx[i];
 557         if( ads < thresh )
 558             mvs[nmv++] = i;
 559     }
 560     return nmv;
 561 }
 562
 563 static int x264_pixel_ads2( int enc_dc[2], uint16_t *sums, int delta,
 564                             uint16_t *cost_mvx, int16_t *mvs, int width, int thresh )
 565 {
 566     int nmv=0, i;
 567     for( i=0; i<width; i++, sums++ )
 568     {
 569         int ads = abs( enc_dc[0] - sums[0] )
 570                 + abs( enc_dc[1] - sums[delta] )
 571                 + cost_mvx[i];
 572         if( ads < thresh )
 573             mvs[nmv++] = i;
 574     }
 575     return nmv;
 576 }
 577
 578 static int x264_pixel_ads1( int enc_dc[1], uint16_t *sums, int delta,
 579                             uint16_t *cost_mvx, int16_t *mvs, int width, int thresh )
 580 {
 581     int nmv=0, i;
 582     for( i=0; i<width; i++, sums++ )
 583     {
 584         int ads = abs( enc_dc[0] - sums[0] )
 585                 + cost_mvx[i];
 586         if( ads < thresh )
 587             mvs[nmv++] = i;
 588     }
 589     return nmv;
 590 }
 591
 592
 593 /****************************************************************************
 594  * x264_pixel_init:
 595  ****************************************************************************/
 596 void x264_pixel_init( int cpu, x264_pixel_function_t *pixf )
 597 {
 598     memset( pixf, 0, sizeof(*pixf) );
 599
 600 #define INIT2_NAME( name1, name2, cpu ) \
 601     pixf->name1[PIXEL_16x16] = x264_pixel_##name2##_16x16##cpu;\
 602     pixf->name1[PIXEL_16x8]  = x264_pixel_##name2##_16x8##cpu;
 603 #define INIT4_NAME( name1, name2, cpu ) \
 604     INIT2_NAME( name1, name2, cpu ) \
 605     pixf->name1[PIXEL_8x16]  = x264_pixel_##name2##_8x16##cpu;\
 606     pixf->name1[PIXEL_8x8]   = x264_pixel_##name2##_8x8##cpu;
 607 #define INIT5_NAME( name1, name2, cpu ) \
 608     INIT4_NAME( name1, name2, cpu ) \
 609     pixf->name1[PIXEL_8x4]   = x264_pixel_##name2##_8x4##cpu;
 610 #define INIT6_NAME( name1, name2, cpu ) \
 611     INIT5_NAME( name1, name2, cpu ) \
 612     pixf->name1[PIXEL_4x8]   = x264_pixel_##name2##_4x8##cpu;
 613 #define INIT7_NAME( name1, name2, cpu ) \
 614     INIT6_NAME( name1, name2, cpu ) \
 615     pixf->name1[PIXEL_4x4]   = x264_pixel_##name2##_4x4##cpu;
 616 #define INIT2( name, cpu ) INIT2_NAME( name, name, cpu )
 617 #define INIT4( name, cpu ) INIT4_NAME( name, name, cpu )
 618 #define INIT5( name, cpu ) INIT5_NAME( name, name, cpu )
 619 #define INIT6( name, cpu ) INIT6_NAME( name, name, cpu )
 620 #define INIT7( name, cpu ) INIT7_NAME( name, name, cpu )
 621
 622 #define INIT_ADS( cpu ) \
 623     pixf->ads[PIXEL_16x16] = x264_pixel_ads4##cpu;\
 624     pixf->ads[PIXEL_16x8] = x264_pixel_ads2##cpu;\
 625     pixf->ads[PIXEL_8x8] = x264_pixel_ads1##cpu;
 626
 627     INIT7( sad, );
 628     INIT7_NAME( sad_aligned, sad, );
 629     INIT7( sad_x3, );
 630     INIT7( sad_x4, );
 631     INIT7( ssd, );
 632     INIT7( satd, );
 633     INIT7( satd_x3, );
 634     INIT7( satd_x4, );
 635     INIT4( hadamard_ac, );
 636     INIT_ADS( );
 637
 638     pixf->sa8d[PIXEL_16x16] = x264_pixel_sa8d_16x16;
 639     pixf->sa8d[PIXEL_8x8]   = x264_pixel_sa8d_8x8;
 640     pixf->var[PIXEL_16x16] = x264_pixel_var_16x16;
 641     pixf->var[PIXEL_8x8]   = x264_pixel_var_8x8;
 642
 643     pixf->ssim_4x4x2_core = ssim_4x4x2_core;
 644     pixf->ssim_end4 = ssim_end4;
 645     pixf->var2_8x8 = pixel_var2_8x8;
 646
 647 #ifdef HAVE_MMX
 648     if( cpu&X264_CPU_MMX )
 649     {
 650         INIT7( ssd, _mmx );
 651     }
 652
 653     if( cpu&X264_CPU_MMXEXT )
 654     {
 655         INIT7( sad, _mmxext );
 656         INIT7_NAME( sad_aligned, sad, _mmxext );
 657         INIT7( sad_x3, _mmxext );
 658         INIT7( sad_x4, _mmxext );
 659         INIT7( satd, _mmxext );
 660         INIT7( satd_x3, _mmxext );
 661         INIT7( satd_x4, _mmxext );
 662         INIT4( hadamard_ac, _mmxext );
 663         INIT_ADS( _mmxext );
 664         pixf->var[PIXEL_16x16] = x264_pixel_var_16x16_mmxext;
 665         pixf->var[PIXEL_8x8]   = x264_pixel_var_8x8_mmxext;
 666 #ifdef ARCH_X86
 667         pixf->sa8d[PIXEL_16x16] = x264_pixel_sa8d_16x16_mmxext;
 668         pixf->sa8d[PIXEL_8x8]   = x264_pixel_sa8d_8x8_mmxext;
 669         pixf->intra_sa8d_x3_8x8 = x264_intra_sa8d_x3_8x8_mmxext;
 670         pixf->ssim_4x4x2_core  = x264_pixel_ssim_4x4x2_core_mmxext;
 671         pixf->var2_8x8 = x264_pixel_var2_8x8_mmxext;
 672
 673         if( cpu&X264_CPU_CACHELINE_32 )
 674         {
 675             INIT5( sad, _cache32_mmxext );
 676             INIT4( sad_x3, _cache32_mmxext );
 677             INIT4( sad_x4, _cache32_mmxext );
 678         }
 679         else if( cpu&X264_CPU_CACHELINE_64 )
 680         {
 681             INIT5( sad, _cache64_mmxext );
 682             INIT4( sad_x3, _cache64_mmxext );
 683             INIT4( sad_x4, _cache64_mmxext );
 684         }
 685 #else
 686         if( cpu&X264_CPU_CACHELINE_64 )
 687         {
 688             pixf->sad[PIXEL_8x16] = x264_pixel_sad_8x16_cache64_mmxext;
 689             pixf->sad[PIXEL_8x8]  = x264_pixel_sad_8x8_cache64_mmxext;
 690             pixf->sad[PIXEL_8x4]  = x264_pixel_sad_8x4_cache64_mmxext;
 691             pixf->sad_x3[PIXEL_8x16] = x264_pixel_sad_x3_8x16_cache64_mmxext;
 692             pixf->sad_x3[PIXEL_8x8]  = x264_pixel_sad_x3_8x8_cache64_mmxext;
 693             pixf->sad_x4[PIXEL_8x16] = x264_pixel_sad_x4_8x16_cache64_mmxext;
 694             pixf->sad_x4[PIXEL_8x8]  = x264_pixel_sad_x4_8x8_cache64_mmxext;
 695         }
 696 #endif
 697         pixf->intra_satd_x3_16x16 = x264_intra_satd_x3_16x16_mmxext;
 698         pixf->intra_sad_x3_16x16  = x264_intra_sad_x3_16x16_mmxext;
 699         pixf->intra_satd_x3_8x8c  = x264_intra_satd_x3_8x8c_mmxext;
 700         pixf->intra_sad_x3_8x8c   = x264_intra_sad_x3_8x8c_mmxext;
 701         pixf->intra_sad_x3_8x8    = x264_intra_sad_x3_8x8_mmxext;
 702         pixf->intra_satd_x3_4x4   = x264_intra_satd_x3_4x4_mmxext;
 703         pixf->intra_sad_x3_4x4    = x264_intra_sad_x3_4x4_mmxext;
 704     }
 705
 706     if( cpu&X264_CPU_SSE2 )
 707     {
 708         INIT5( ssd, _sse2slow );
 709         INIT2_NAME( sad_aligned, sad, _sse2_aligned );
 710         pixf->var[PIXEL_16x16] = x264_pixel_var_16x16_sse2;
 711         pixf->ssim_4x4x2_core  = x264_pixel_ssim_4x4x2_core_sse2;
 712         pixf->ssim_end4        = x264_pixel_ssim_end4_sse2;
 713         pixf->sa8d[PIXEL_16x16] = x264_pixel_sa8d_16x16_sse2;
 714         pixf->sa8d[PIXEL_8x8]   = x264_pixel_sa8d_8x8_sse2;
 715 #ifdef ARCH_X86_64
 716         pixf->intra_sa8d_x3_8x8 = x264_intra_sa8d_x3_8x8_sse2;
 717 #endif
 718         pixf->var2_8x8 = x264_pixel_var2_8x8_sse2;
 719     }
 720
 721     if( (cpu&X264_CPU_SSE2) && !(cpu&X264_CPU_SSE2_IS_SLOW) )
 722     {
 723         INIT2( sad, _sse2 );
 724         INIT2( sad_x3, _sse2 );
 725         INIT2( sad_x4, _sse2 );
 726         INIT6( satd, _sse2 );
 727         INIT6( satd_x3, _sse2 );
 728         INIT6( satd_x4, _sse2 );
 729         if( !(cpu&X264_CPU_STACK_MOD4) )
 730         {
 731             INIT4( hadamard_ac, _sse2 );
 732         }
 733         INIT_ADS( _sse2 );
 734         pixf->var[PIXEL_8x8] = x264_pixel_var_8x8_sse2;
 735         pixf->intra_sad_x3_16x16 = x264_intra_sad_x3_16x16_sse2;
 736         if( cpu&X264_CPU_CACHELINE_64 )
 737         {
 738             INIT2( ssd, _sse2); /* faster for width 16 on p4 */
 739 #ifdef ARCH_X86
 740             INIT2( sad, _cache64_sse2 );
 741             INIT2( sad_x3, _cache64_sse2 );
 742             INIT2( sad_x4, _cache64_sse2 );
 743 #endif
 744            if( cpu&X264_CPU_SSE2_IS_FAST )
 745            {
 746                pixf->sad_x3[PIXEL_8x16] = x264_pixel_sad_x3_8x16_cache64_sse2;
 747                pixf->sad_x4[PIXEL_8x16] = x264_pixel_sad_x4_8x16_cache64_sse2;
 748            }
 749         }
 750
 751         if( cpu&X264_CPU_SSE_MISALIGN )
 752         {
 753             INIT2( sad_x3, _sse2_misalign );
 754             INIT2( sad_x4, _sse2_misalign );
 755         }
 756     }
 757
 758     if( cpu&X264_CPU_SSE2_IS_FAST && !(cpu&X264_CPU_CACHELINE_64) )
 759     {
 760         pixf->sad_aligned[PIXEL_8x16] = x264_pixel_sad_8x16_sse2;
 761         pixf->sad[PIXEL_8x16] = x264_pixel_sad_8x16_sse2;
 762         pixf->sad_x3[PIXEL_8x16] = x264_pixel_sad_x3_8x16_sse2;
 763         pixf->sad_x3[PIXEL_8x8] = x264_pixel_sad_x3_8x8_sse2;
 764         pixf->sad_x3[PIXEL_8x4] = x264_pixel_sad_x3_8x4_sse2;
 765         pixf->sad_x4[PIXEL_8x16] = x264_pixel_sad_x4_8x16_sse2;
 766         pixf->sad_x4[PIXEL_8x8] = x264_pixel_sad_x4_8x8_sse2;
 767         pixf->sad_x4[PIXEL_8x4] = x264_pixel_sad_x4_8x4_sse2;
 768     }
 769
 770     if( (cpu&X264_CPU_SSE3) && (cpu&X264_CPU_CACHELINE_64) )
 771     {
 772         INIT2( sad, _sse3 );
 773         INIT2( sad_x3, _sse3 );
 774         INIT2( sad_x4, _sse3 );
 775     }
 776
 777     if( cpu&X264_CPU_SSSE3 )
 778     {
 779         INIT7( ssd, _ssse3 );
 780         INIT7( satd, _ssse3 );
 781         INIT7( satd_x3, _ssse3 );
 782         INIT7( satd_x4, _ssse3 );
 783         if( !(cpu&X264_CPU_STACK_MOD4) )
 784         {
 785             INIT4( hadamard_ac, _ssse3 );
 786         }
 787         INIT_ADS( _ssse3 );
 788         pixf->sa8d[PIXEL_16x16]= x264_pixel_sa8d_16x16_ssse3;
 789         pixf->sa8d[PIXEL_8x8]  = x264_pixel_sa8d_8x8_ssse3;
 790         pixf->intra_satd_x3_16x16 = x264_intra_satd_x3_16x16_ssse3;
 791         pixf->intra_sad_x3_16x16  = x264_intra_sad_x3_16x16_ssse3;
 792         pixf->intra_satd_x3_8x8c  = x264_intra_satd_x3_8x8c_ssse3;
 793         pixf->intra_sad_x3_8x8c   = x264_intra_sad_x3_8x8c_ssse3;
 794         pixf->intra_satd_x3_4x4   = x264_intra_satd_x3_4x4_ssse3;
 795 #ifdef ARCH_X86_64
 796         pixf->intra_sa8d_x3_8x8 = x264_intra_sa8d_x3_8x8_ssse3;
 797 #endif
 798         pixf->var2_8x8 = x264_pixel_var2_8x8_ssse3;
 799         if( cpu&X264_CPU_CACHELINE_64 )
 800         {
 801             INIT2( sad, _cache64_ssse3 );
 802             INIT2( sad_x3, _cache64_ssse3 );
 803             INIT2( sad_x4, _cache64_ssse3 );
 804         }
 805         if( !(cpu&X264_CPU_SHUFFLE_IS_FAST) )
 806         {
 807             INIT5( ssd, _sse2 ); /* on conroe, sse2 is faster for width8/16 */
 808         }
 809     }
 810
 811     if( cpu&X264_CPU_SSE4 )
 812     {
 813         INIT7( satd, _sse4 );
 814         INIT7( satd_x3, _sse4 );
 815         INIT7( satd_x4, _sse4 );
 816         if( !(cpu&X264_CPU_STACK_MOD4) )
 817         {
 818             INIT4( hadamard_ac, _sse4 );
 819         }
 820         pixf->sa8d[PIXEL_16x16]= x264_pixel_sa8d_16x16_sse4;
 821         pixf->sa8d[PIXEL_8x8]  = x264_pixel_sa8d_8x8_sse4;
 822     }
 823 #endif //HAVE_MMX
 824
 825 #ifdef HAVE_ARMV6
 826     if( cpu&X264_CPU_ARMV6 )
 827     {
 828         pixf->sad[PIXEL_4x8] = x264_pixel_sad_4x8_armv6;
 829         pixf->sad[PIXEL_4x4] = x264_pixel_sad_4x4_armv6;
 830         pixf->sad_aligned[PIXEL_4x8] = x264_pixel_sad_4x8_armv6;
 831         pixf->sad_aligned[PIXEL_4x4] = x264_pixel_sad_4x4_armv6;
 832     }
 833     if( cpu&X264_CPU_NEON )
 834     {
 835         INIT5( sad, _neon );
 836         INIT5( sad_aligned, _neon );
 837         INIT7( sad_x3, _neon );
 838         INIT7( sad_x4, _neon );
 839         INIT7( ssd, _neon );
 840         INIT7( satd, _neon );
 841         INIT7( satd_x3, _neon );
 842         INIT7( satd_x4, _neon );
 843         INIT4( hadamard_ac, _neon );
 844         pixf->sa8d[PIXEL_8x8]   = x264_pixel_sa8d_8x8_neon;
 845         pixf->sa8d[PIXEL_16x16] = x264_pixel_sa8d_16x16_neon;
 846         pixf->var[PIXEL_8x8]    = x264_pixel_var_8x8_neon;
 847         pixf->var[PIXEL_16x16]  = x264_pixel_var_16x16_neon;
 848         pixf->var2_8x8          = x264_pixel_var2_8x8_neon;
 849
 850         pixf->ssim_4x4x2_core   = x264_pixel_ssim_4x4x2_core_neon;
 851         pixf->ssim_end4         = x264_pixel_ssim_end4_neon;
 852
 853         if( cpu&X264_CPU_FAST_NEON_MRC )
 854         {
 855             pixf->sad[PIXEL_4x8] = x264_pixel_sad_4x8_neon;
 856             pixf->sad[PIXEL_4x4] = x264_pixel_sad_4x4_neon;
 857             pixf->sad_aligned[PIXEL_4x8] = x264_pixel_sad_aligned_4x8_neon;
 858             pixf->sad_aligned[PIXEL_4x4] = x264_pixel_sad_aligned_4x4_neon;
 859         }
 860         else    // really just scheduled for dual issue / A8
 861         {
 862             INIT5( sad_aligned, _neon_dual );
 863         }
 864     }
 865 #endif
 866 #ifdef ARCH_PPC
 867     if( cpu&X264_CPU_ALTIVEC )
 868     {
 869         x264_pixel_altivec_init( pixf );
 870     }
 871 #endif
 872 #ifdef ARCH_UltraSparc
 873     INIT4( sad, _vis );
 874     INIT4( sad_x3, _vis );
 875     INIT4( sad_x4, _vis );
 876 #endif
 877
 878     pixf->ads[PIXEL_8x16] =
 879     pixf->ads[PIXEL_8x4] =
 880     pixf->ads[PIXEL_4x8] = pixf->ads[PIXEL_16x8];
 881     pixf->ads[PIXEL_4x4] = pixf->ads[PIXEL_8x8];
 882 }
 883