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( pixel *pix1, int i_stride_pix1,  \
  45                  pixel *pix2, int i_stride_pix2 ) \
  46 {                                                   \
  47     int i_sum = 0;                                  \
  48     for( int y = 0; y < ly; y++ )                   \
  49     {                                               \
  50         for( int x = 0; x < lx; x++ )               \
  51         {                                           \
  52             i_sum += abs( pix1[x] - pix2[x] );      \
  53         }                                           \
  54         pix1 += i_stride_pix1;                      \
  55         pix2 += i_stride_pix2;                      \
  56     }                                               \
  57     return i_sum;                                   \
  58 }
  59
  60
  61 PIXEL_SAD_C( x264_pixel_sad_16x16, 16, 16 )
  62 PIXEL_SAD_C( x264_pixel_sad_16x8,  16,  8 )
  63 PIXEL_SAD_C( x264_pixel_sad_8x16,   8, 16 )
  64 PIXEL_SAD_C( x264_pixel_sad_8x8,    8,  8 )
  65 PIXEL_SAD_C( x264_pixel_sad_8x4,    8,  4 )
  66 PIXEL_SAD_C( x264_pixel_sad_4x8,    4,  8 )
  67 PIXEL_SAD_C( x264_pixel_sad_4x4,    4,  4 )
  68
  69
  70 /****************************************************************************
  71  * pixel_ssd_WxH
  72  ****************************************************************************/
  73 #define PIXEL_SSD_C( name, lx, ly ) \
  74 static int name( pixel *pix1, int i_stride_pix1,  \
  75                  pixel *pix2, int i_stride_pix2 ) \
  76 {                                                   \
  77     int i_sum = 0;                                  \
  78     for( int y = 0; y < ly; y++ )                   \
  79     {                                               \
  80         for( int x = 0; x < lx; x++ )               \
  81         {                                           \
  82             int d = pix1[x] - pix2[x];              \
  83             i_sum += d*d;                           \
  84         }                                           \
  85         pix1 += i_stride_pix1;                      \
  86         pix2 += i_stride_pix2;                      \
  87     }                                               \
  88     return i_sum;                                   \
  89 }
  90
  91 PIXEL_SSD_C( x264_pixel_ssd_16x16, 16, 16 )
  92 PIXEL_SSD_C( x264_pixel_ssd_16x8,  16,  8 )
  93 PIXEL_SSD_C( x264_pixel_ssd_8x16,   8, 16 )
  94 PIXEL_SSD_C( x264_pixel_ssd_8x8,    8,  8 )
  95 PIXEL_SSD_C( x264_pixel_ssd_8x4,    8,  4 )
  96 PIXEL_SSD_C( x264_pixel_ssd_4x8,    4,  8 )
  97 PIXEL_SSD_C( x264_pixel_ssd_4x4,    4,  4 )
  98
  99 int64_t x264_pixel_ssd_wxh( x264_pixel_function_t *pf, pixel *pix1, int i_pix1, pixel *pix2, int i_pix2, int i_width, int i_height )
 100 {
 101     int64_t i_ssd = 0;
 102     int y;
 103     int align = !(((intptr_t)pix1 | (intptr_t)pix2 | i_pix1 | i_pix2) & 15);
 104
 105 #define SSD(size) i_ssd += pf->ssd[size]( pix1 + y*i_pix1 + x, i_pix1, \
 106                                           pix2 + y*i_pix2 + x, i_pix2 );
 107     for( y = 0; y < i_height-15; y += 16 )
 108     {
 109         int x = 0;
 110         if( align )
 111             for( ; x < i_width-15; x += 16 )
 112                 SSD(PIXEL_16x16);
 113         for( ; x < i_width-7; x += 8 )
 114             SSD(PIXEL_8x16);
 115     }
 116     if( y < i_height-7 )
 117         for( int x = 0; x < i_width-7; x += 8 )
 118             SSD(PIXEL_8x8);
 119 #undef SSD
 120
 121 #define SSD1 { int d = pix1[y*i_pix1+x] - pix2[y*i_pix2+x]; i_ssd += d*d; }
 122     if( i_width & 7 )
 123     {
 124         for( y = 0; y < (i_height & ~7); y++ )
 125             for( int x = i_width & ~7; x < i_width; x++ )
 126                 SSD1;
 127     }
 128     if( i_height & 7 )
 129     {
 130         for( y = i_height & ~7; y < i_height; y++ )
 131             for( int x = 0; x < i_width; x++ )
 132                 SSD1;
 133     }
 134 #undef SSD1
 135
 136     return i_ssd;
 137 }
 138
 139
 140 /****************************************************************************
 141  * pixel_var_wxh
 142  ****************************************************************************/
 143 #define PIXEL_VAR_C( name, w ) \
 144 static uint64_t name( pixel *pix, int i_stride ) \
 145 {                                             \
 146     uint32_t sum = 0, sqr = 0;                \
 147     for( int y = 0; y < w; y++ )              \
 148     {                                         \
 149         for( int x = 0; x < w; x++ )          \
 150         {                                     \
 151             sum += pix[x];                    \
 152             sqr += pix[x] * pix[x];           \
 153         }                                     \
 154         pix += i_stride;                      \
 155     }                                         \
 156     return sum + ((uint64_t)sqr << 32);       \
 157 }
 158
 159 PIXEL_VAR_C( x264_pixel_var_16x16, 16 )
 160 PIXEL_VAR_C( x264_pixel_var_8x8,    8 )
 161
 162 /****************************************************************************
 163  * pixel_var2_wxh
 164  ****************************************************************************/
 165 static int pixel_var2_8x8( pixel *pix1, int i_stride1, pixel *pix2, int i_stride2, int *ssd )
 166 {
 167     uint32_t var = 0, sum = 0, sqr = 0;
 168     for( int y = 0; y < 8; y++ )
 169     {
 170         for( int x = 0; x < 8; x++ )
 171         {
 172             int diff = pix1[x] - pix2[x];
 173             sum += diff;
 174             sqr += diff * diff;
 175         }
 176         pix1 += i_stride1;
 177         pix2 += i_stride2;
 178     }
 179     sum = abs(sum);
 180     var = sqr - (sum * sum >> 6);
 181     *ssd = sqr;
 182     return var;
 183 }
 184
 185
 186 #define HADAMARD4(d0, d1, d2, d3, s0, s1, s2, s3) {\
 187     int t0 = s0 + s1;\
 188     int t1 = s0 - s1;\
 189     int t2 = s2 + s3;\
 190     int t3 = s2 - s3;\
 191     d0 = t0 + t2;\
 192     d2 = t0 - t2;\
 193     d1 = t1 + t3;\
 194     d3 = t1 - t3;\
 195 }
 196
 197 // in: a pseudo-simd number of the form x+(y<<16)
 198 // return: abs(x)+(abs(y)<<16)
 199 static ALWAYS_INLINE uint32_t abs2( uint32_t a )
 200 {
 201     uint32_t s = ((a>>15)&0x10001)*0xffff;
 202     return (a+s)^s;
 203 }
 204
 205 /****************************************************************************
 206  * pixel_satd_WxH: sum of 4x4 Hadamard transformed differences
 207  ****************************************************************************/
 208
 209 static NOINLINE int x264_pixel_satd_4x4( pixel *pix1, int i_pix1, pixel *pix2, int i_pix2 )
 210 {
 211     uint32_t tmp[4][2];
 212     uint32_t a0, a1, a2, a3, b0, b1;
 213     int sum = 0;
 214     for( int i = 0; i < 4; i++, pix1 += i_pix1, pix2 += i_pix2 )
 215     {
 216         a0 = pix1[0] - pix2[0];
 217         a1 = pix1[1] - pix2[1];
 218         b0 = (a0+a1) + ((a0-a1)<<16);
 219         a2 = pix1[2] - pix2[2];
 220         a3 = pix1[3] - pix2[3];
 221         b1 = (a2+a3) + ((a2-a3)<<16);
 222         tmp[i][0] = b0 + b1;
 223         tmp[i][1] = b0 - b1;
 224     }
 225     for( int i = 0; i < 2; i++ )
 226     {
 227         HADAMARD4( a0, a1, a2, a3, tmp[0][i], tmp[1][i], tmp[2][i], tmp[3][i] );
 228         a0 = abs2(a0) + abs2(a1) + abs2(a2) + abs2(a3);
 229         sum += ((uint16_t)a0) + (a0>>16);
 230     }
 231     return sum >> 1;
 232 }
 233
 234 static NOINLINE int x264_pixel_satd_8x4( pixel *pix1, int i_pix1, pixel *pix2, int i_pix2 )
 235 {
 236     uint32_t tmp[4][4];
 237     uint32_t a0, a1, a2, a3;
 238     int sum = 0;
 239     for( int i = 0; i < 4; i++, pix1 += i_pix1, pix2 += i_pix2 )
 240     {
 241         a0 = (pix1[0] - pix2[0]) + ((pix1[4] - pix2[4]) << 16);
 242         a1 = (pix1[1] - pix2[1]) + ((pix1[5] - pix2[5]) << 16);
 243         a2 = (pix1[2] - pix2[2]) + ((pix1[6] - pix2[6]) << 16);
 244         a3 = (pix1[3] - pix2[3]) + ((pix1[7] - pix2[7]) << 16);
 245         HADAMARD4( tmp[i][0], tmp[i][1], tmp[i][2], tmp[i][3], a0,a1,a2,a3 );
 246     }
 247     for( int i = 0; i < 4; i++ )
 248     {
 249         HADAMARD4( a0, a1, a2, a3, tmp[0][i], tmp[1][i], tmp[2][i], tmp[3][i] );
 250         sum += abs2(a0) + abs2(a1) + abs2(a2) + abs2(a3);
 251     }
 252     return (((uint16_t)sum) + ((uint32_t)sum>>16)) >> 1;
 253 }
 254
 255 #define PIXEL_SATD_C( w, h, sub )\
 256 static int x264_pixel_satd_##w##x##h( pixel *pix1, int i_pix1, pixel *pix2, int i_pix2 )\
 257 {\
 258     int sum = sub( pix1, i_pix1, pix2, i_pix2 )\
 259             + sub( pix1+4*i_pix1, i_pix1, pix2+4*i_pix2, i_pix2 );\
 260     if( w==16 )\
 261         sum+= sub( pix1+8, i_pix1, pix2+8, i_pix2 )\
 262             + sub( pix1+8+4*i_pix1, i_pix1, pix2+8+4*i_pix2, i_pix2 );\
 263     if( h==16 )\
 264         sum+= sub( pix1+8*i_pix1, i_pix1, pix2+8*i_pix2, i_pix2 )\
 265             + sub( pix1+12*i_pix1, i_pix1, pix2+12*i_pix2, i_pix2 );\
 266     if( w==16 && h==16 )\
 267         sum+= sub( pix1+8+8*i_pix1, i_pix1, pix2+8+8*i_pix2, i_pix2 )\
 268             + sub( pix1+8+12*i_pix1, i_pix1, pix2+8+12*i_pix2, i_pix2 );\
 269     return sum;\
 270 }
 271 PIXEL_SATD_C( 16, 16, x264_pixel_satd_8x4 )
 272 PIXEL_SATD_C( 16, 8,  x264_pixel_satd_8x4 )
 273 PIXEL_SATD_C( 8,  16, x264_pixel_satd_8x4 )
 274 PIXEL_SATD_C( 8,  8,  x264_pixel_satd_8x4 )
 275 PIXEL_SATD_C( 4,  8,  x264_pixel_satd_4x4 )
 276
 277
 278 static NOINLINE int sa8d_8x8( pixel *pix1, int i_pix1, pixel *pix2, int i_pix2 )
 279 {
 280     uint32_t tmp[8][4];
 281     uint32_t a0, a1, a2, a3, a4, a5, a6, a7, b0, b1, b2, b3;
 282     int sum = 0;
 283     for( int i = 0; i < 8; i++, pix1 += i_pix1, pix2 += i_pix2 )
 284     {
 285         a0 = pix1[0] - pix2[0];
 286         a1 = pix1[1] - pix2[1];
 287         b0 = (a0+a1) + ((a0-a1)<<16);
 288         a2 = pix1[2] - pix2[2];
 289         a3 = pix1[3] - pix2[3];
 290         b1 = (a2+a3) + ((a2-a3)<<16);
 291         a4 = pix1[4] - pix2[4];
 292         a5 = pix1[5] - pix2[5];
 293         b2 = (a4+a5) + ((a4-a5)<<16);
 294         a6 = pix1[6] - pix2[6];
 295         a7 = pix1[7] - pix2[7];
 296         b3 = (a6+a7) + ((a6-a7)<<16);
 297         HADAMARD4( tmp[i][0], tmp[i][1], tmp[i][2], tmp[i][3], b0,b1,b2,b3 );
 298     }
 299     for( int i = 0; i < 4; i++ )
 300     {
 301         HADAMARD4( a0, a1, a2, a3, tmp[0][i], tmp[1][i], tmp[2][i], tmp[3][i] );
 302         HADAMARD4( a4, a5, a6, a7, tmp[4][i], tmp[5][i], tmp[6][i], tmp[7][i] );
 303         b0  = abs2(a0+a4) + abs2(a0-a4);
 304         b0 += abs2(a1+a5) + abs2(a1-a5);
 305         b0 += abs2(a2+a6) + abs2(a2-a6);
 306         b0 += abs2(a3+a7) + abs2(a3-a7);
 307         sum += (uint16_t)b0 + (b0>>16);
 308     }
 309     return sum;
 310 }
 311
 312 static int x264_pixel_sa8d_8x8( pixel *pix1, int i_pix1, pixel *pix2, int i_pix2 )
 313 {
 314     int sum = sa8d_8x8( pix1, i_pix1, pix2, i_pix2 );
 315     return (sum+2)>>2;
 316 }
 317
 318 static int x264_pixel_sa8d_16x16( pixel *pix1, int i_pix1, pixel *pix2, int i_pix2 )
 319 {
 320     int sum = sa8d_8x8( pix1, i_pix1, pix2, i_pix2 )
 321             + sa8d_8x8( pix1+8, i_pix1, pix2+8, i_pix2 )
 322             + sa8d_8x8( pix1+8*i_pix1, i_pix1, pix2+8*i_pix2, i_pix2 )
 323             + sa8d_8x8( pix1+8+8*i_pix1, i_pix1, pix2+8+8*i_pix2, i_pix2 );
 324     return (sum+2)>>2;
 325 }
 326
 327
 328 static NOINLINE uint64_t pixel_hadamard_ac( pixel *pix, int stride )
 329 {
 330     uint32_t tmp[32];
 331     uint32_t a0, a1, a2, a3, dc;
 332     int sum4 = 0, sum8 = 0;
 333     for( int i = 0; i < 8; i++, pix+=stride )
 334     {
 335         uint32_t *t = tmp + (i&3) + (i&4)*4;
 336         a0 = (pix[0]+pix[1]) + ((pix[0]-pix[1])<<16);
 337         a1 = (pix[2]+pix[3]) + ((pix[2]-pix[3])<<16);
 338         t[0] = a0 + a1;
 339         t[4] = a0 - a1;
 340         a2 = (pix[4]+pix[5]) + ((pix[4]-pix[5])<<16);
 341         a3 = (pix[6]+pix[7]) + ((pix[6]-pix[7])<<16);
 342         t[8] = a2 + a3;
 343         t[12] = a2 - a3;
 344     }
 345     for( int i = 0; i < 8; i++ )
 346     {
 347         HADAMARD4( a0, a1, a2, a3, tmp[i*4+0], tmp[i*4+1], tmp[i*4+2], tmp[i*4+3] );
 348         tmp[i*4+0] = a0;
 349         tmp[i*4+1] = a1;
 350         tmp[i*4+2] = a2;
 351         tmp[i*4+3] = a3;
 352         sum4 += abs2(a0) + abs2(a1) + abs2(a2) + abs2(a3);
 353     }
 354     for( int i = 0; i < 8; i++ )
 355     {
 356         HADAMARD4( a0,a1,a2,a3, tmp[i], tmp[8+i], tmp[16+i], tmp[24+i] );
 357         sum8 += abs2(a0) + abs2(a1) + abs2(a2) + abs2(a3);
 358     }
 359     dc = (uint16_t)(tmp[0] + tmp[8] + tmp[16] + tmp[24]);
 360     sum4 = (uint16_t)sum4 + ((uint32_t)sum4>>16) - dc;
 361     sum8 = (uint16_t)sum8 + ((uint32_t)sum8>>16) - dc;
 362     return ((uint64_t)sum8<<32) + sum4;
 363 }
 364
 365 #define HADAMARD_AC(w,h) \
 366 static uint64_t x264_pixel_hadamard_ac_##w##x##h( pixel *pix, int stride )\
 367 {\
 368     uint64_t sum = pixel_hadamard_ac( pix, stride );\
 369     if( w==16 )\
 370         sum += pixel_hadamard_ac( pix+8, stride );\
 371     if( h==16 )\
 372         sum += pixel_hadamard_ac( pix+8*stride, stride );\
 373     if( w==16 && h==16 )\
 374         sum += pixel_hadamard_ac( pix+8*stride+8, stride );\
 375     return ((sum>>34)<<32) + ((uint32_t)sum>>1);\
 376 }
 377 HADAMARD_AC( 16, 16 )
 378 HADAMARD_AC( 16, 8 )
 379 HADAMARD_AC( 8, 16 )
 380 HADAMARD_AC( 8, 8 )
 381
 382
 383 /****************************************************************************
 384  * pixel_sad_x4
 385  ****************************************************************************/
 386 #define SAD_X( size ) \
 387 static void x264_pixel_sad_x3_##size( pixel *fenc, pixel *pix0, pixel *pix1, pixel *pix2, int i_stride, int scores[3] )\
 388 {\
 389     scores[0] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix0, i_stride );\
 390     scores[1] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix1, i_stride );\
 391     scores[2] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix2, i_stride );\
 392 }\
 393 static void x264_pixel_sad_x4_##size( pixel *fenc, pixel *pix0, pixel *pix1, pixel *pix2, pixel *pix3, int i_stride, int scores[4] )\
 394 {\
 395     scores[0] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix0, i_stride );\
 396     scores[1] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix1, i_stride );\
 397     scores[2] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix2, i_stride );\
 398     scores[3] = x264_pixel_sad_##size( fenc, FENC_STRIDE, pix3, i_stride );\
 399 }
 400
 401 SAD_X( 16x16 )
 402 SAD_X( 16x8 )
 403 SAD_X( 8x16 )
 404 SAD_X( 8x8 )
 405 SAD_X( 8x4 )
 406 SAD_X( 4x8 )
 407 SAD_X( 4x4 )
 408
 409 #ifdef ARCH_UltraSparc
 410 SAD_X( 16x16_vis )
 411 SAD_X( 16x8_vis )
 412 SAD_X( 8x16_vis )
 413 SAD_X( 8x8_vis )
 414 #endif
 415
 416 /****************************************************************************
 417  * pixel_satd_x4
 418  * no faster than single satd, but needed for satd to be a drop-in replacement for sad
 419  ****************************************************************************/
 420
 421 #define SATD_X( size, cpu ) \
 422 static void x264_pixel_satd_x3_##size##cpu( pixel *fenc, pixel *pix0, pixel *pix1, pixel *pix2, int i_stride, int scores[3] )\
 423 {\
 424     scores[0] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix0, i_stride );\
 425     scores[1] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix1, i_stride );\
 426     scores[2] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix2, i_stride );\
 427 }\
 428 static void x264_pixel_satd_x4_##size##cpu( pixel *fenc, pixel *pix0, pixel *pix1, pixel *pix2, pixel *pix3, int i_stride, int scores[4] )\
 429 {\
 430     scores[0] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix0, i_stride );\
 431     scores[1] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix1, i_stride );\
 432     scores[2] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix2, i_stride );\
 433     scores[3] = x264_pixel_satd_##size##cpu( fenc, FENC_STRIDE, pix3, i_stride );\
 434 }
 435 #define SATD_X_DECL6( cpu )\
 436 SATD_X( 16x16, cpu )\
 437 SATD_X( 16x8, cpu )\
 438 SATD_X( 8x16, cpu )\
 439 SATD_X( 8x8, cpu )\
 440 SATD_X( 8x4, cpu )\
 441 SATD_X( 4x8, cpu )
 442 #define SATD_X_DECL7( cpu )\
 443 SATD_X_DECL6( cpu )\
 444 SATD_X( 4x4, cpu )
 445
 446 SATD_X_DECL7()
 447 #ifdef HAVE_MMX
 448 SATD_X_DECL7( _mmxext )
 449 SATD_X_DECL6( _sse2 )
 450 SATD_X_DECL7( _ssse3 )
 451 SATD_X_DECL7( _sse4 )
 452 #endif
 453
 454 #ifdef HAVE_ARMV6
 455 SATD_X_DECL7( _neon )
 456 #endif
 457
 458 /****************************************************************************
 459  * structural similarity metric
 460  ****************************************************************************/
 461 static void ssim_4x4x2_core( const pixel *pix1, int stride1,
 462                              const pixel *pix2, int stride2,
 463                              int sums[2][4])
 464 {
 465     for( int z = 0; z < 2; z++ )
 466     {
 467         uint32_t s1 = 0, s2 = 0, ss = 0, s12 = 0;
 468         for( int y = 0; y < 4; y++ )
 469             for( int x = 0; x < 4; x++ )
 470             {
 471                 int a = pix1[x+y*stride1];
 472                 int b = pix2[x+y*stride2];
 473                 s1  += a;
 474                 s2  += b;
 475                 ss  += a*a;
 476                 ss  += b*b;
 477                 s12 += a*b;
 478             }
 479         sums[z][0] = s1;
 480         sums[z][1] = s2;
 481         sums[z][2] = ss;
 482         sums[z][3] = s12;
 483         pix1 += 4;
 484         pix2 += 4;
 485     }
 486 }
 487
 488 static float ssim_end1( int s1, int s2, int ss, int s12 )
 489 {
 490     static const int ssim_c1 = (int)(.01*.01*255*255*64 + .5);
 491     static const int ssim_c2 = (int)(.03*.03*255*255*64*63 + .5);
 492     int vars = ss*64 - s1*s1 - s2*s2;
 493     int covar = s12*64 - s1*s2;
 494     return (float)(2*s1*s2 + ssim_c1) * (float)(2*covar + ssim_c2)
 495          / ((float)(s1*s1 + s2*s2 + ssim_c1) * (float)(vars + ssim_c2));
 496 }
 497
 498 static float ssim_end4( int sum0[5][4], int sum1[5][4], int width )
 499 {
 500     float ssim = 0.0;
 501     for( int i = 0; i < width; i++ )
 502         ssim += ssim_end1( sum0[i][0] + sum0[i+1][0] + sum1[i][0] + sum1[i+1][0],
 503                            sum0[i][1] + sum0[i+1][1] + sum1[i][1] + sum1[i+1][1],
 504                            sum0[i][2] + sum0[i+1][2] + sum1[i][2] + sum1[i+1][2],
 505                            sum0[i][3] + sum0[i+1][3] + sum1[i][3] + sum1[i+1][3] );
 506     return ssim;
 507 }
 508
 509 float x264_pixel_ssim_wxh( x264_pixel_function_t *pf,
 510                            pixel *pix1, int stride1,
 511                            pixel *pix2, int stride2,
 512                            int width, int height, void *buf )
 513 {
 514     int z = 0;
 515     float ssim = 0.0;
 516     int (*sum0)[4] = buf;
 517     int (*sum1)[4] = sum0 + width/4+3;
 518     width >>= 2;
 519     height >>= 2;
 520     for( int y = 1; y < height; y++ )
 521     {
 522         for( ; z <= y; z++ )
 523         {
 524             XCHG( void*, sum0, sum1 );
 525             for( int x = 0; x < width; x+=2 )
 526                 pf->ssim_4x4x2_core( &pix1[4*(x+z*stride1)], stride1, &pix2[4*(x+z*stride2)], stride2, &sum0[x] );
 527         }
 528         for( int x = 0; x < width-1; x += 4 )
 529             ssim += pf->ssim_end4( sum0+x, sum1+x, X264_MIN(4,width-x-1) );
 530     }
 531     return ssim;
 532 }
 533
 534
 535 /****************************************************************************
 536  * successive elimination
 537  ****************************************************************************/
 538 static int x264_pixel_ads4( int enc_dc[4], uint16_t *sums, int delta,
 539                             uint16_t *cost_mvx, int16_t *mvs, int width, int thresh )
 540 {
 541     int nmv = 0;
 542     for( int i = 0; i < width; i++, sums++ )
 543     {
 544         int ads = abs( enc_dc[0] - sums[0] )
 545                 + abs( enc_dc[1] - sums[8] )
 546                 + abs( enc_dc[2] - sums[delta] )
 547                 + abs( enc_dc[3] - sums[delta+8] )
 548                 + cost_mvx[i];
 549         if( ads < thresh )
 550             mvs[nmv++] = i;
 551     }
 552     return nmv;
 553 }
 554
 555 static int x264_pixel_ads2( int enc_dc[2], uint16_t *sums, int delta,
 556                             uint16_t *cost_mvx, int16_t *mvs, int width, int thresh )
 557 {
 558     int nmv = 0;
 559     for( int i = 0; i < width; i++, sums++ )
 560     {
 561         int ads = abs( enc_dc[0] - sums[0] )
 562                 + abs( enc_dc[1] - sums[delta] )
 563                 + cost_mvx[i];
 564         if( ads < thresh )
 565             mvs[nmv++] = i;
 566     }
 567     return nmv;
 568 }
 569
 570 static int x264_pixel_ads1( int enc_dc[1], uint16_t *sums, int delta,
 571                             uint16_t *cost_mvx, int16_t *mvs, int width, int thresh )
 572 {
 573     int nmv = 0;
 574     for( int i = 0; i<width; i++, sums++ )
 575     {
 576         int ads = abs( enc_dc[0] - sums[0] )
 577                 + cost_mvx[i];
 578         if( ads < thresh )
 579             mvs[nmv++] = i;
 580     }
 581     return nmv;
 582 }
 583
 584
 585 /****************************************************************************
 586  * x264_pixel_init:
 587  ****************************************************************************/
 588 void x264_pixel_init( int cpu, x264_pixel_function_t *pixf )
 589 {
 590     memset( pixf, 0, sizeof(*pixf) );
 591
 592 #define INIT2_NAME( name1, name2, cpu ) \
 593     pixf->name1[PIXEL_16x16] = x264_pixel_##name2##_16x16##cpu;\
 594     pixf->name1[PIXEL_16x8]  = x264_pixel_##name2##_16x8##cpu;
 595 #define INIT4_NAME( name1, name2, cpu ) \
 596     INIT2_NAME( name1, name2, cpu ) \
 597     pixf->name1[PIXEL_8x16]  = x264_pixel_##name2##_8x16##cpu;\
 598     pixf->name1[PIXEL_8x8]   = x264_pixel_##name2##_8x8##cpu;
 599 #define INIT5_NAME( name1, name2, cpu ) \
 600     INIT4_NAME( name1, name2, cpu ) \
 601     pixf->name1[PIXEL_8x4]   = x264_pixel_##name2##_8x4##cpu;
 602 #define INIT6_NAME( name1, name2, cpu ) \
 603     INIT5_NAME( name1, name2, cpu ) \
 604     pixf->name1[PIXEL_4x8]   = x264_pixel_##name2##_4x8##cpu;
 605 #define INIT7_NAME( name1, name2, cpu ) \
 606     INIT6_NAME( name1, name2, cpu ) \
 607     pixf->name1[PIXEL_4x4]   = x264_pixel_##name2##_4x4##cpu;
 608 #define INIT2( name, cpu ) INIT2_NAME( name, name, cpu )
 609 #define INIT4( name, cpu ) INIT4_NAME( name, name, cpu )
 610 #define INIT5( name, cpu ) INIT5_NAME( name, name, cpu )
 611 #define INIT6( name, cpu ) INIT6_NAME( name, name, cpu )
 612 #define INIT7( name, cpu ) INIT7_NAME( name, name, cpu )
 613
 614 #define INIT_ADS( cpu ) \
 615     pixf->ads[PIXEL_16x16] = x264_pixel_ads4##cpu;\
 616     pixf->ads[PIXEL_16x8] = x264_pixel_ads2##cpu;\
 617     pixf->ads[PIXEL_8x8] = x264_pixel_ads1##cpu;
 618
 619     INIT7( sad, );
 620     INIT7_NAME( sad_aligned, sad, );
 621     INIT7( sad_x3, );
 622     INIT7( sad_x4, );
 623     INIT7( ssd, );
 624     INIT7( satd, );
 625     INIT7( satd_x3, );
 626     INIT7( satd_x4, );
 627     INIT4( hadamard_ac, );
 628     INIT_ADS( );
 629
 630     pixf->sa8d[PIXEL_16x16] = x264_pixel_sa8d_16x16;
 631     pixf->sa8d[PIXEL_8x8]   = x264_pixel_sa8d_8x8;
 632     pixf->var[PIXEL_16x16] = x264_pixel_var_16x16;
 633     pixf->var[PIXEL_8x8]   = x264_pixel_var_8x8;
 634
 635     pixf->ssim_4x4x2_core = ssim_4x4x2_core;
 636     pixf->ssim_end4 = ssim_end4;
 637     pixf->var2_8x8 = pixel_var2_8x8;
 638
 639 #ifdef HAVE_MMX
 640     if( cpu&X264_CPU_MMX )
 641     {
 642         INIT7( ssd, _mmx );
 643     }
 644
 645     if( cpu&X264_CPU_MMXEXT )
 646     {
 647         INIT7( sad, _mmxext );
 648         INIT7_NAME( sad_aligned, sad, _mmxext );
 649         INIT7( sad_x3, _mmxext );
 650         INIT7( sad_x4, _mmxext );
 651         INIT7( satd, _mmxext );
 652         INIT7( satd_x3, _mmxext );
 653         INIT7( satd_x4, _mmxext );
 654         INIT4( hadamard_ac, _mmxext );
 655         INIT_ADS( _mmxext );
 656         pixf->var[PIXEL_16x16] = x264_pixel_var_16x16_mmxext;
 657         pixf->var[PIXEL_8x8]   = x264_pixel_var_8x8_mmxext;
 658 #ifdef ARCH_X86
 659         pixf->sa8d[PIXEL_16x16] = x264_pixel_sa8d_16x16_mmxext;
 660         pixf->sa8d[PIXEL_8x8]   = x264_pixel_sa8d_8x8_mmxext;
 661         pixf->intra_sa8d_x3_8x8 = x264_intra_sa8d_x3_8x8_mmxext;
 662         pixf->ssim_4x4x2_core  = x264_pixel_ssim_4x4x2_core_mmxext;
 663         pixf->var2_8x8 = x264_pixel_var2_8x8_mmxext;
 664
 665         if( cpu&X264_CPU_CACHELINE_32 )
 666         {
 667             INIT5( sad, _cache32_mmxext );
 668             INIT4( sad_x3, _cache32_mmxext );
 669             INIT4( sad_x4, _cache32_mmxext );
 670         }
 671         else if( cpu&X264_CPU_CACHELINE_64 )
 672         {
 673             INIT5( sad, _cache64_mmxext );
 674             INIT4( sad_x3, _cache64_mmxext );
 675             INIT4( sad_x4, _cache64_mmxext );
 676         }
 677 #else
 678         if( cpu&X264_CPU_CACHELINE_64 )
 679         {
 680             pixf->sad[PIXEL_8x16] = x264_pixel_sad_8x16_cache64_mmxext;
 681             pixf->sad[PIXEL_8x8]  = x264_pixel_sad_8x8_cache64_mmxext;
 682             pixf->sad[PIXEL_8x4]  = x264_pixel_sad_8x4_cache64_mmxext;
 683             pixf->sad_x3[PIXEL_8x16] = x264_pixel_sad_x3_8x16_cache64_mmxext;
 684             pixf->sad_x3[PIXEL_8x8]  = x264_pixel_sad_x3_8x8_cache64_mmxext;
 685             pixf->sad_x4[PIXEL_8x16] = x264_pixel_sad_x4_8x16_cache64_mmxext;
 686             pixf->sad_x4[PIXEL_8x8]  = x264_pixel_sad_x4_8x8_cache64_mmxext;
 687         }
 688 #endif
 689         pixf->intra_satd_x3_16x16 = x264_intra_satd_x3_16x16_mmxext;
 690         pixf->intra_sad_x3_16x16  = x264_intra_sad_x3_16x16_mmxext;
 691         pixf->intra_satd_x3_8x8c  = x264_intra_satd_x3_8x8c_mmxext;
 692         pixf->intra_sad_x3_8x8c   = x264_intra_sad_x3_8x8c_mmxext;
 693         pixf->intra_sad_x3_8x8    = x264_intra_sad_x3_8x8_mmxext;
 694         pixf->intra_satd_x3_4x4   = x264_intra_satd_x3_4x4_mmxext;
 695         pixf->intra_sad_x3_4x4    = x264_intra_sad_x3_4x4_mmxext;
 696     }
 697
 698     if( cpu&X264_CPU_SSE2 )
 699     {
 700         INIT5( ssd, _sse2slow );
 701         INIT2_NAME( sad_aligned, sad, _sse2_aligned );
 702         pixf->var[PIXEL_16x16] = x264_pixel_var_16x16_sse2;
 703         pixf->ssim_4x4x2_core  = x264_pixel_ssim_4x4x2_core_sse2;
 704         pixf->ssim_end4        = x264_pixel_ssim_end4_sse2;
 705         pixf->sa8d[PIXEL_16x16] = x264_pixel_sa8d_16x16_sse2;
 706         pixf->sa8d[PIXEL_8x8]   = x264_pixel_sa8d_8x8_sse2;
 707 #ifdef ARCH_X86_64
 708         pixf->intra_sa8d_x3_8x8 = x264_intra_sa8d_x3_8x8_sse2;
 709 #endif
 710         pixf->var2_8x8 = x264_pixel_var2_8x8_sse2;
 711     }
 712
 713     if( (cpu&X264_CPU_SSE2) && !(cpu&X264_CPU_SSE2_IS_SLOW) )
 714     {
 715         INIT2( sad, _sse2 );
 716         INIT2( sad_x3, _sse2 );
 717         INIT2( sad_x4, _sse2 );
 718         INIT6( satd, _sse2 );
 719         INIT6( satd_x3, _sse2 );
 720         INIT6( satd_x4, _sse2 );
 721         if( !(cpu&X264_CPU_STACK_MOD4) )
 722         {
 723             INIT4( hadamard_ac, _sse2 );
 724         }
 725         INIT_ADS( _sse2 );
 726         pixf->var[PIXEL_8x8] = x264_pixel_var_8x8_sse2;
 727         pixf->intra_sad_x3_16x16 = x264_intra_sad_x3_16x16_sse2;
 728         if( cpu&X264_CPU_CACHELINE_64 )
 729         {
 730             INIT2( ssd, _sse2); /* faster for width 16 on p4 */
 731 #ifdef ARCH_X86
 732             INIT2( sad, _cache64_sse2 );
 733             INIT2( sad_x3, _cache64_sse2 );
 734             INIT2( sad_x4, _cache64_sse2 );
 735 #endif
 736            if( cpu&X264_CPU_SSE2_IS_FAST )
 737            {
 738                pixf->sad_x3[PIXEL_8x16] = x264_pixel_sad_x3_8x16_cache64_sse2;
 739                pixf->sad_x4[PIXEL_8x16] = x264_pixel_sad_x4_8x16_cache64_sse2;
 740            }
 741         }
 742
 743         if( cpu&X264_CPU_SSE_MISALIGN )
 744         {
 745             INIT2( sad_x3, _sse2_misalign );
 746             INIT2( sad_x4, _sse2_misalign );
 747         }
 748     }
 749
 750     if( cpu&X264_CPU_SSE2_IS_FAST && !(cpu&X264_CPU_CACHELINE_64) )
 751     {
 752         pixf->sad_aligned[PIXEL_8x16] = x264_pixel_sad_8x16_sse2;
 753         pixf->sad[PIXEL_8x16] = x264_pixel_sad_8x16_sse2;
 754         pixf->sad_x3[PIXEL_8x16] = x264_pixel_sad_x3_8x16_sse2;
 755         pixf->sad_x3[PIXEL_8x8] = x264_pixel_sad_x3_8x8_sse2;
 756         pixf->sad_x3[PIXEL_8x4] = x264_pixel_sad_x3_8x4_sse2;
 757         pixf->sad_x4[PIXEL_8x16] = x264_pixel_sad_x4_8x16_sse2;
 758         pixf->sad_x4[PIXEL_8x8] = x264_pixel_sad_x4_8x8_sse2;
 759         pixf->sad_x4[PIXEL_8x4] = x264_pixel_sad_x4_8x4_sse2;
 760     }
 761
 762     if( (cpu&X264_CPU_SSE3) && (cpu&X264_CPU_CACHELINE_64) )
 763     {
 764         INIT2( sad, _sse3 );
 765         INIT2( sad_x3, _sse3 );
 766         INIT2( sad_x4, _sse3 );
 767     }
 768
 769     if( cpu&X264_CPU_SSSE3 )
 770     {
 771         if( !(cpu&X264_CPU_STACK_MOD4) )
 772         {
 773             INIT4( hadamard_ac, _ssse3 );
 774         }
 775         INIT_ADS( _ssse3 );
 776         if( !(cpu&X264_CPU_SLOW_ATOM) )
 777         {
 778             INIT7( ssd, _ssse3 );
 779             pixf->sa8d[PIXEL_16x16]= x264_pixel_sa8d_16x16_ssse3;
 780             pixf->sa8d[PIXEL_8x8]  = x264_pixel_sa8d_8x8_ssse3;
 781             INIT7( satd, _ssse3 );
 782             INIT7( satd_x3, _ssse3 );
 783             INIT7( satd_x4, _ssse3 );
 784         }
 785         pixf->intra_satd_x3_16x16 = x264_intra_satd_x3_16x16_ssse3;
 786         pixf->intra_sad_x3_16x16  = x264_intra_sad_x3_16x16_ssse3;
 787         pixf->intra_satd_x3_8x8c  = x264_intra_satd_x3_8x8c_ssse3;
 788         pixf->intra_sad_x3_8x8c   = x264_intra_sad_x3_8x8c_ssse3;
 789         pixf->intra_satd_x3_4x4   = x264_intra_satd_x3_4x4_ssse3;
 790 #ifdef ARCH_X86_64
 791         pixf->intra_sa8d_x3_8x8 = x264_intra_sa8d_x3_8x8_ssse3;
 792 #endif
 793         pixf->var2_8x8 = x264_pixel_var2_8x8_ssse3;
 794         if( cpu&X264_CPU_CACHELINE_64 )
 795         {
 796             INIT2( sad, _cache64_ssse3 );
 797             INIT2( sad_x3, _cache64_ssse3 );
 798             INIT2( sad_x4, _cache64_ssse3 );
 799         }
 800         if( cpu&X264_CPU_SLOW_ATOM || !(cpu&X264_CPU_SHUFFLE_IS_FAST) )
 801         {
 802             INIT5( ssd, _sse2 ); /* on conroe, sse2 is faster for width8/16 */
 803         }
 804     }
 805
 806     if( cpu&X264_CPU_SSE4 )
 807     {
 808         INIT7( satd, _sse4 );
 809         INIT7( satd_x3, _sse4 );
 810         INIT7( satd_x4, _sse4 );
 811         if( !(cpu&X264_CPU_STACK_MOD4) )
 812         {
 813             INIT4( hadamard_ac, _sse4 );
 814         }
 815         pixf->sa8d[PIXEL_16x16]= x264_pixel_sa8d_16x16_sse4;
 816         pixf->sa8d[PIXEL_8x8]  = x264_pixel_sa8d_8x8_sse4;
 817     }
 818 #endif //HAVE_MMX
 819
 820 #ifdef HAVE_ARMV6
 821     if( cpu&X264_CPU_ARMV6 )
 822     {
 823         pixf->sad[PIXEL_4x8] = x264_pixel_sad_4x8_armv6;
 824         pixf->sad[PIXEL_4x4] = x264_pixel_sad_4x4_armv6;
 825         pixf->sad_aligned[PIXEL_4x8] = x264_pixel_sad_4x8_armv6;
 826         pixf->sad_aligned[PIXEL_4x4] = x264_pixel_sad_4x4_armv6;
 827     }
 828     if( cpu&X264_CPU_NEON )
 829     {
 830         INIT5( sad, _neon );
 831         INIT5( sad_aligned, _neon );
 832         INIT7( sad_x3, _neon );
 833         INIT7( sad_x4, _neon );
 834         INIT7( ssd, _neon );
 835         INIT7( satd, _neon );
 836         INIT7( satd_x3, _neon );
 837         INIT7( satd_x4, _neon );
 838         INIT4( hadamard_ac, _neon );
 839         pixf->sa8d[PIXEL_8x8]   = x264_pixel_sa8d_8x8_neon;
 840         pixf->sa8d[PIXEL_16x16] = x264_pixel_sa8d_16x16_neon;
 841         pixf->var[PIXEL_8x8]    = x264_pixel_var_8x8_neon;
 842         pixf->var[PIXEL_16x16]  = x264_pixel_var_16x16_neon;
 843         pixf->var2_8x8          = x264_pixel_var2_8x8_neon;
 844
 845         pixf->ssim_4x4x2_core   = x264_pixel_ssim_4x4x2_core_neon;
 846         pixf->ssim_end4         = x264_pixel_ssim_end4_neon;
 847
 848         if( cpu&X264_CPU_FAST_NEON_MRC )
 849         {
 850             pixf->sad[PIXEL_4x8] = x264_pixel_sad_4x8_neon;
 851             pixf->sad[PIXEL_4x4] = x264_pixel_sad_4x4_neon;
 852             pixf->sad_aligned[PIXEL_4x8] = x264_pixel_sad_aligned_4x8_neon;
 853             pixf->sad_aligned[PIXEL_4x4] = x264_pixel_sad_aligned_4x4_neon;
 854         }
 855         else    // really just scheduled for dual issue / A8
 856         {
 857             INIT5( sad_aligned, _neon_dual );
 858         }
 859     }
 860 #endif
 861 #ifdef HAVE_ALTIVEC
 862     if( cpu&X264_CPU_ALTIVEC )
 863     {
 864         x264_pixel_altivec_init( pixf );
 865     }
 866 #endif
 867 #ifdef ARCH_UltraSparc
 868     INIT4( sad, _vis );
 869     INIT4( sad_x3, _vis );
 870     INIT4( sad_x4, _vis );
 871 #endif
 872
 873     pixf->ads[PIXEL_8x16] =
 874     pixf->ads[PIXEL_8x4] =
 875     pixf->ads[PIXEL_4x8] = pixf->ads[PIXEL_16x8];
 876     pixf->ads[PIXEL_4x4] = pixf->ads[PIXEL_8x8];
 877 }
 878