git.sesse.net Git - ffmpeg/blob - postproc/postprocess_template.c

   1 /*
   2     Copyright (C) 2001-2002 Michael Niedermayer (michaelni@gmx.at)
   3
   4     This program is free software; you can redistribute it and/or modify
   5     it under the terms of the GNU General Public License as published by
   6     the Free Software Foundation; either version 2 of the License, or
   7     (at your option) any later version.
   8
   9     This program is distributed in the hope that it will be useful,
  10     but WITHOUT ANY WARRANTY; without even the implied warranty of
  11     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12     GNU General Public License for more details.
  13
  14     You should have received a copy of the GNU General Public License
  15     along with this program; if not, write to the Free Software
  16     Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  17 */
  18
  19 #undef PAVGB
  20 #undef PMINUB
  21 #undef PMAXUB
  22
  23 #ifdef HAVE_MMX2
  24 #define PAVGB(a,b) "pavgb " #a ", " #b " \n\t"
  25 #elif defined (HAVE_3DNOW)
  26 #define PAVGB(a,b) "pavgusb " #a ", " #b " \n\t"
  27 #endif
  28
  29 #ifdef HAVE_MMX2
  30 #define PMINUB(a,b,t) "pminub " #a ", " #b " \n\t"
  31 #elif defined (HAVE_MMX)
  32 #define PMINUB(b,a,t) \
  33         "movq " #a ", " #t " \n\t"\
  34         "psubusb " #b ", " #t " \n\t"\
  35         "psubb " #t ", " #a " \n\t"
  36 #endif
  37
  38 #ifdef HAVE_MMX2
  39 #define PMAXUB(a,b) "pmaxub " #a ", " #b " \n\t"
  40 #elif defined (HAVE_MMX)
  41 #define PMAXUB(a,b) \
  42         "psubusb " #a ", " #b " \n\t"\
  43         "paddb " #a ", " #b " \n\t"
  44 #endif
  45
  46
  47 //FIXME? |255-0| = 1 (shouldnt be a problem ...)
  48 /**
  49  * Check if the middle 8x8 Block in the given 8x16 block is flat
  50  */
  51 static inline int RENAME(isVertDC)(uint8_t src[], int stride){
  52         int numEq= 0;
  53 #ifndef HAVE_MMX
  54         int y;
  55 #endif
  56         src+= stride*4; // src points to begin of the 8x8 Block
  57 #ifdef HAVE_MMX
  58 asm volatile(
  59                 "leal (%1, %2), %%eax                           \n\t"
  60                 "leal (%%eax, %2, 4), %%ebx                     \n\t"
  61 //      0       1       2       3       4       5       6       7       8       9
  62 //      %1      eax     eax+%2  eax+2%2 %1+4%2  ebx     ebx+%2  ebx+2%2 %1+8%2  ebx+4%2
  63                 "movq "MANGLE(mmxDCOffset)", %%mm7              \n\t" // mm7 = 0x7F
  64                 "movq "MANGLE(mmxDCThreshold)", %%mm6           \n\t" // mm6 = 0x7D
  65                 "movq (%1), %%mm0                               \n\t"
  66                 "movq (%%eax), %%mm1                            \n\t"
  67                 "psubb %%mm1, %%mm0                             \n\t" // mm0 = differnece
  68                 "paddb %%mm7, %%mm0                             \n\t"
  69                 "pcmpgtb %%mm6, %%mm0                           \n\t"
  70
  71                 "movq (%%eax,%2), %%mm2                         \n\t"
  72                 "psubb %%mm2, %%mm1                             \n\t"
  73                 "paddb %%mm7, %%mm1                             \n\t"
  74                 "pcmpgtb %%mm6, %%mm1                           \n\t"
  75                 "paddb %%mm1, %%mm0                             \n\t"
  76
  77                 "movq (%%eax, %2, 2), %%mm1                     \n\t"
  78                 "psubb %%mm1, %%mm2                             \n\t"
  79                 "paddb %%mm7, %%mm2                             \n\t"
  80                 "pcmpgtb %%mm6, %%mm2                           \n\t"
  81                 "paddb %%mm2, %%mm0                             \n\t"
  82
  83                 "movq (%1, %2, 4), %%mm2                        \n\t"
  84                 "psubb %%mm2, %%mm1                             \n\t"
  85                 "paddb %%mm7, %%mm1                             \n\t"
  86                 "pcmpgtb %%mm6, %%mm1                           \n\t"
  87                 "paddb %%mm1, %%mm0                             \n\t"
  88
  89                 "movq (%%ebx), %%mm1                            \n\t"
  90                 "psubb %%mm1, %%mm2                             \n\t"
  91                 "paddb %%mm7, %%mm2                             \n\t"
  92                 "pcmpgtb %%mm6, %%mm2                           \n\t"
  93                 "paddb %%mm2, %%mm0                             \n\t"
  94
  95                 "movq (%%ebx, %2), %%mm2                        \n\t"
  96                 "psubb %%mm2, %%mm1                             \n\t"
  97                 "paddb %%mm7, %%mm1                             \n\t"
  98                 "pcmpgtb %%mm6, %%mm1                           \n\t"
  99                 "paddb %%mm1, %%mm0                             \n\t"
 100
 101                 "movq (%%ebx, %2, 2), %%mm1                     \n\t"
 102                 "psubb %%mm1, %%mm2                             \n\t"
 103                 "paddb %%mm7, %%mm2                             \n\t"
 104                 "pcmpgtb %%mm6, %%mm2                           \n\t"
 105                 "paddb %%mm2, %%mm0                             \n\t"
 106
 107                 "                                               \n\t"
 108 #ifdef HAVE_MMX2
 109                 "pxor %%mm7, %%mm7                              \n\t"
 110                 "psadbw %%mm7, %%mm0                            \n\t"
 111 #else
 112                 "movq %%mm0, %%mm1                              \n\t"
 113                 "psrlw $8, %%mm0                                \n\t"
 114                 "paddb %%mm1, %%mm0                             \n\t"
 115                 "movq %%mm0, %%mm1                              \n\t"
 116                 "psrlq $16, %%mm0                               \n\t"
 117                 "paddb %%mm1, %%mm0                             \n\t"
 118                 "movq %%mm0, %%mm1                              \n\t"
 119                 "psrlq $32, %%mm0                               \n\t"
 120                 "paddb %%mm1, %%mm0                             \n\t"
 121 #endif
 122                 "movd %%mm0, %0                                 \n\t"
 123                 : "=r" (numEq)
 124                 : "r" (src), "r" (stride)
 125                 : "%ebx"
 126                 );
 127         numEq= (-numEq) &0xFF;
 128
 129 #else
 130         for(y=0; y<BLOCK_SIZE-1; y++)
 131         {
 132                 if(((src[0] - src[0+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
 133                 if(((src[1] - src[1+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
 134                 if(((src[2] - src[2+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
 135                 if(((src[3] - src[3+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
 136                 if(((src[4] - src[4+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
 137                 if(((src[5] - src[5+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
 138                 if(((src[6] - src[6+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
 139                 if(((src[7] - src[7+stride] + dcOffset)&0xFFFF) < dcThreshold) numEq++;
 140                 src+= stride;
 141         }
 142 #endif
 143 /*      if(abs(numEq - asmEq) > 0)
 144         {
 145                 printf("\nasm:%d  c:%d\n", asmEq, numEq);
 146                 for(int y=0; y<8; y++)
 147                 {
 148                         for(int x=0; x<8; x++)
 149                         {
 150                                 printf("%d ", temp[x + y*stride]);
 151                         }
 152                         printf("\n");
 153                 }
 154         }
 155 */
 156 //      for(int i=0; i<numEq/8; i++) src[i]=255;
 157         return (numEq > vFlatnessThreshold) ? 1 : 0;
 158 }
 159
 160 static inline int RENAME(isVertMinMaxOk)(uint8_t src[], int stride, int QP)
 161 {
 162 #ifdef HAVE_MMX
 163         int isOk;
 164         src+= stride*3;
 165         asm volatile(
 166 //              "int $3 \n\t"
 167                 "movq (%1, %2), %%mm0                           \n\t"
 168                 "movq (%1, %2, 8), %%mm1                        \n\t"
 169                 "movq %%mm0, %%mm2                              \n\t"
 170                 "psubusb %%mm1, %%mm0                           \n\t"
 171                 "psubusb %%mm2, %%mm1                           \n\t"
 172                 "por %%mm1, %%mm0                               \n\t" // ABS Diff
 173
 174                 "movq "MANGLE(pQPb)", %%mm7                     \n\t" // QP,..., QP
 175                 "paddusb %%mm7, %%mm7                           \n\t" // 2QP ... 2QP
 176                 "psubusb %%mm7, %%mm0                           \n\t" // Diff <= 2QP -> 0
 177                 "pcmpeqd "MANGLE(b00)", %%mm0                   \n\t"
 178                 "psrlq $16, %%mm0                               \n\t"
 179                 "pcmpeqd "MANGLE(bFF)", %%mm0                   \n\t"
 180 //              "movd %%mm0, (%1, %2, 4)\n\t"
 181                 "movd %%mm0, %0                                 \n\t"
 182                 : "=r" (isOk)
 183                 : "r" (src), "r" (stride)
 184                 );
 185         return isOk;
 186 #else
 187
 188         int isOk2= 1;
 189         int x;
 190         src+= stride*3;
 191         for(x=0; x<BLOCK_SIZE; x++)
 192         {
 193                 if(abs((int)src[x + stride] - (int)src[x + (stride<<3)]) > 2*QP) isOk2=0;
 194         }
 195 /*      if(isOk && !isOk2 || !isOk && isOk2)
 196         {
 197                 printf("\nasm:%d  c:%d QP:%d\n", isOk, isOk2, QP);
 198                 for(int y=0; y<9; y++)
 199                 {
 200                         for(int x=0; x<8; x++)
 201                         {
 202                                 printf("%d ", src[x + y*stride]);
 203                         }
 204                         printf("\n");
 205                 }
 206         } */
 207
 208         return isOk2;
 209 #endif
 210
 211 }
 212
 213 /**
 214  * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)
 215  * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16
 216  */
 217 static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, int QP)
 218 {
 219 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
 220         src+= stride*3;
 221         asm volatile(   //"movv %0 %1 %2\n\t"
 222                 "movq "MANGLE(pQPb)", %%mm0                     \n\t"  // QP,..., QP
 223
 224                 "movq (%0), %%mm6                               \n\t"
 225                 "movq (%0, %1), %%mm5                           \n\t"
 226                 "movq %%mm5, %%mm1                              \n\t"
 227                 "movq %%mm6, %%mm2                              \n\t"
 228                 "psubusb %%mm6, %%mm5                           \n\t"
 229                 "psubusb %%mm1, %%mm2                           \n\t"
 230                 "por %%mm5, %%mm2                               \n\t" // ABS Diff of lines
 231                 "psubusb %%mm0, %%mm2                           \n\t" // diff <= QP -> 0
 232                 "pcmpeqb "MANGLE(b00)", %%mm2                   \n\t" // diff <= QP -> FF
 233
 234                 "pand %%mm2, %%mm6                              \n\t"
 235                 "pandn %%mm1, %%mm2                             \n\t"
 236                 "por %%mm2, %%mm6                               \n\t"// First Line to Filter
 237
 238                 "movq (%0, %1, 8), %%mm5                        \n\t"
 239                 "leal (%0, %1, 4), %%eax                        \n\t"
 240                 "leal (%0, %1, 8), %%ebx                        \n\t"
 241                 "subl %1, %%ebx                                 \n\t"
 242                 "addl %1, %0                                    \n\t" // %0 points to line 1 not 0
 243                 "movq (%0, %1, 8), %%mm7                        \n\t"
 244                 "movq %%mm5, %%mm1                              \n\t"
 245                 "movq %%mm7, %%mm2                              \n\t"
 246                 "psubusb %%mm7, %%mm5                           \n\t"
 247                 "psubusb %%mm1, %%mm2                           \n\t"
 248                 "por %%mm5, %%mm2                               \n\t" // ABS Diff of lines
 249                 "psubusb %%mm0, %%mm2                           \n\t" // diff <= QP -> 0
 250                 "pcmpeqb "MANGLE(b00)", %%mm2                   \n\t" // diff <= QP -> FF
 251
 252                 "pand %%mm2, %%mm7                              \n\t"
 253                 "pandn %%mm1, %%mm2                             \n\t"
 254                 "por %%mm2, %%mm7                               \n\t" // First Line to Filter
 255
 256
 257                 //      1       2       3       4       5       6       7       8
 258                 //      %0      %0+%1   %0+2%1  eax     %0+4%1  eax+2%1 ebx     eax+4%1
 259                 // 6 4 2 2 1 1
 260                 // 6 4 4 2
 261                 // 6 8 2
 262
 263                 "movq (%0, %1), %%mm0                           \n\t" //  1
 264                 "movq %%mm0, %%mm1                              \n\t" //  1
 265                 PAVGB(%%mm6, %%mm0)                                   //1 1     /2
 266                 PAVGB(%%mm6, %%mm0)                                   //3 1     /4
 267
 268                 "movq (%0, %1, 4), %%mm2                        \n\t" //     1
 269                 "movq %%mm2, %%mm5                              \n\t" //     1
 270                 PAVGB((%%eax), %%mm2)                                 //    11  /2
 271                 PAVGB((%0, %1, 2), %%mm2)                             //   211  /4
 272                 "movq %%mm2, %%mm3                              \n\t" //   211  /4
 273                 "movq (%0), %%mm4                               \n\t" // 1
 274                 PAVGB(%%mm4, %%mm3)                                   // 4 211  /8
 275                 PAVGB(%%mm0, %%mm3)                                   //642211  /16
 276                 "movq %%mm3, (%0)                               \n\t" // X
 277                 // mm1=2 mm2=3(211) mm4=1 mm5=5 mm6=0 mm7=9
 278                 "movq %%mm1, %%mm0                              \n\t" //  1
 279                 PAVGB(%%mm6, %%mm0)                                   //1 1     /2
 280                 "movq %%mm4, %%mm3                              \n\t" // 1
 281                 PAVGB((%0,%1,2), %%mm3)                               // 1 1    /2
 282                 PAVGB((%%eax,%1,2), %%mm5)                            //     11 /2
 283                 PAVGB((%%eax), %%mm5)                                 //    211 /4
 284                 PAVGB(%%mm5, %%mm3)                                   // 2 2211 /8
 285                 PAVGB(%%mm0, %%mm3)                                   //4242211 /16
 286                 "movq %%mm3, (%0,%1)                            \n\t" //  X
 287                 // mm1=2 mm2=3(211) mm4=1 mm5=4(211) mm6=0 mm7=9
 288                 PAVGB(%%mm4, %%mm6)                                   //11      /2
 289                 "movq (%%ebx), %%mm0                            \n\t" //       1
 290                 PAVGB((%%eax, %1, 2), %%mm0)                          //      11/2
 291                 "movq %%mm0, %%mm3                              \n\t" //      11/2
 292                 PAVGB(%%mm1, %%mm0)                                   //  2   11/4
 293                 PAVGB(%%mm6, %%mm0)                                   //222   11/8
 294                 PAVGB(%%mm2, %%mm0)                                   //22242211/16
 295                 "movq (%0, %1, 2), %%mm2                        \n\t" //   1
 296                 "movq %%mm0, (%0, %1, 2)                        \n\t" //   X
 297                 // mm1=2 mm2=3 mm3=6(11) mm4=1 mm5=4(211) mm6=0(11) mm7=9
 298                 "movq (%%eax, %1, 4), %%mm0                     \n\t" //        1
 299                 PAVGB((%%ebx), %%mm0)                                 //       11       /2
 300                 PAVGB(%%mm0, %%mm6)                                   //11     11       /4
 301                 PAVGB(%%mm1, %%mm4)                                   // 11             /2
 302                 PAVGB(%%mm2, %%mm1)                                   //  11            /2
 303                 PAVGB(%%mm1, %%mm6)                                   //1122   11       /8
 304                 PAVGB(%%mm5, %%mm6)                                   //112242211       /16
 305                 "movq (%%eax), %%mm5                            \n\t" //    1
 306                 "movq %%mm6, (%%eax)                            \n\t" //    X
 307                 // mm0=7(11) mm1=2(11) mm2=3 mm3=6(11) mm4=1(11) mm5=4 mm7=9
 308                 "movq (%%eax, %1, 4), %%mm6                     \n\t" //        1
 309                 PAVGB(%%mm7, %%mm6)                                   //        11      /2
 310                 PAVGB(%%mm4, %%mm6)                                   // 11     11      /4
 311                 PAVGB(%%mm3, %%mm6)                                   // 11   2211      /8
 312                 PAVGB(%%mm5, %%mm2)                                   //   11           /2
 313                 "movq (%0, %1, 4), %%mm4                        \n\t" //     1
 314                 PAVGB(%%mm4, %%mm2)                                   //   112          /4
 315                 PAVGB(%%mm2, %%mm6)                                   // 112242211      /16
 316                 "movq %%mm6, (%0, %1, 4)                        \n\t" //     X
 317                 // mm0=7(11) mm1=2(11) mm2=3(112) mm3=6(11) mm4=5 mm5=4 mm7=9
 318                 PAVGB(%%mm7, %%mm1)                                   //  11     2      /4
 319                 PAVGB(%%mm4, %%mm5)                                   //    11          /2
 320                 PAVGB(%%mm5, %%mm0)                                   //    11 11       /4
 321                 "movq (%%eax, %1, 2), %%mm6                     \n\t" //      1
 322                 PAVGB(%%mm6, %%mm1)                                   //  11  4  2      /8
 323                 PAVGB(%%mm0, %%mm1)                                   //  11224222      /16
 324                 "movq %%mm1, (%%eax, %1, 2)                     \n\t" //      X
 325                 // mm2=3(112) mm3=6(11) mm4=5 mm5=4(11) mm6=6 mm7=9
 326                 PAVGB((%%ebx), %%mm2)                                 //   112 4        /8
 327                 "movq (%%eax, %1, 4), %%mm0                     \n\t" //        1
 328                 PAVGB(%%mm0, %%mm6)                                   //      1 1       /2
 329                 PAVGB(%%mm7, %%mm6)                                   //      1 12      /4
 330                 PAVGB(%%mm2, %%mm6)                                   //   1122424      /4
 331                 "movq %%mm6, (%%ebx)                            \n\t" //       X
 332                 // mm0=8 mm3=6(11) mm4=5 mm5=4(11) mm7=9
 333                 PAVGB(%%mm7, %%mm5)                                   //    11   2      /4
 334                 PAVGB(%%mm7, %%mm5)                                   //    11   6      /8
 335
 336                 PAVGB(%%mm3, %%mm0)                                   //      112       /4
 337                 PAVGB(%%mm0, %%mm5)                                   //    112246      /16
 338                 "movq %%mm5, (%%eax, %1, 4)                     \n\t" //        X
 339                 "subl %1, %0                                    \n\t"
 340
 341                 :
 342                 : "r" (src), "r" (stride)
 343                 : "%eax", "%ebx"
 344         );
 345 #else
 346         const int l1= stride;
 347         const int l2= stride + l1;
 348         const int l3= stride + l2;
 349         const int l4= stride + l3;
 350         const int l5= stride + l4;
 351         const int l6= stride + l5;
 352         const int l7= stride + l6;
 353         const int l8= stride + l7;
 354         const int l9= stride + l8;
 355         int x;
 356         src+= stride*3;
 357         for(x=0; x<BLOCK_SIZE; x++)
 358         {
 359                 const int first= ABS(src[0] - src[l1]) < QP ? src[0] : src[l1];
 360                 const int last= ABS(src[l8] - src[l9]) < QP ? src[l9] : src[l8];
 361
 362                 int sums[9];
 363                 sums[0] = first + src[l1];
 364                 sums[1] = src[l1] + src[l2];
 365                 sums[2] = src[l2] + src[l3];
 366                 sums[3] = src[l3] + src[l4];
 367                 sums[4] = src[l4] + src[l5];
 368                 sums[5] = src[l5] + src[l6];
 369                 sums[6] = src[l6] + src[l7];
 370                 sums[7] = src[l7] + src[l8];
 371                 sums[8] = src[l8] + last;
 372
 373                 src[l1]= ((sums[0]<<2) + ((first + sums[2])<<1) + sums[4] + 8)>>4;
 374                 src[l2]= ((src[l2]<<2) + ((first + sums[0] + sums[3])<<1) + sums[5] + 8)>>4;
 375                 src[l3]= ((src[l3]<<2) + ((first + sums[1] + sums[4])<<1) + sums[6] + 8)>>4;
 376                 src[l4]= ((src[l4]<<2) + ((sums[2] + sums[5])<<1) + sums[0] + sums[7] + 8)>>4;
 377                 src[l5]= ((src[l5]<<2) + ((sums[3] + sums[6])<<1) + sums[1] + sums[8] + 8)>>4;
 378                 src[l6]= ((src[l6]<<2) + ((last + sums[7] + sums[4])<<1) + sums[2] + 8)>>4;
 379                 src[l7]= (((last + src[l7])<<2) + ((src[l8] + sums[5])<<1) + sums[3] + 8)>>4;
 380                 src[l8]= ((sums[8]<<2) + ((last + sums[6])<<1) + sums[4] + 8)>>4;
 381
 382                 src++;
 383         }
 384
 385 #endif
 386 }
 387
 388 /**
 389  * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar
 390  * values are correctly clipped (MMX2)
 391  * values are wraparound (C)
 392  * conclusion: its fast, but introduces ugly horizontal patterns if there is a continious gradient
 393         0 8 16 24
 394         x = 8
 395         x/2 = 4
 396         x/8 = 1
 397         1 12 12 23
 398  */
 399 static inline void RENAME(vertRK1Filter)(uint8_t *src, int stride, int QP)
 400 {
 401 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
 402         src+= stride*3;
 403 // FIXME rounding
 404         asm volatile(
 405                 "pxor %%mm7, %%mm7                              \n\t" // 0
 406                 "movq "MANGLE(b80)", %%mm6                      \n\t" // MIN_SIGNED_BYTE
 407                 "leal (%0, %1), %%eax                           \n\t"
 408                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
 409 //      0       1       2       3       4       5       6       7       8       9
 410 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
 411                 "movq "MANGLE(pQPb)", %%mm0                     \n\t" // QP,..., QP
 412                 "movq %%mm0, %%mm1                              \n\t" // QP,..., QP
 413                 "paddusb "MANGLE(b02)", %%mm0                   \n\t"
 414                 "psrlw $2, %%mm0                                \n\t"
 415                 "pand "MANGLE(b3F)", %%mm0                      \n\t" // QP/4,..., QP/4
 416                 "paddusb %%mm1, %%mm0                           \n\t" // QP*1.25 ...
 417                 "movq (%0, %1, 4), %%mm2                        \n\t" // line 4
 418                 "movq (%%ebx), %%mm3                            \n\t" // line 5
 419                 "movq %%mm2, %%mm4                              \n\t" // line 4
 420                 "pcmpeqb %%mm5, %%mm5                           \n\t" // -1
 421                 "pxor %%mm2, %%mm5                              \n\t" // -line 4 - 1
 422                 PAVGB(%%mm3, %%mm5)
 423                 "paddb %%mm6, %%mm5                             \n\t" // (l5-l4)/2
 424                 "psubusb %%mm3, %%mm4                           \n\t"
 425                 "psubusb %%mm2, %%mm3                           \n\t"
 426                 "por %%mm3, %%mm4                               \n\t" // |l4 - l5|
 427                 "psubusb %%mm0, %%mm4                           \n\t"
 428                 "pcmpeqb %%mm7, %%mm4                           \n\t"
 429                 "pand %%mm4, %%mm5                              \n\t" // d/2
 430
 431 //              "paddb %%mm6, %%mm2                             \n\t" // line 4 + 0x80
 432                 "paddb %%mm5, %%mm2                             \n\t"
 433 //              "psubb %%mm6, %%mm2                             \n\t"
 434                 "movq %%mm2, (%0,%1, 4)                         \n\t"
 435
 436                 "movq (%%ebx), %%mm2                            \n\t"
 437 //              "paddb %%mm6, %%mm2                             \n\t" // line 5 + 0x80
 438                 "psubb %%mm5, %%mm2                             \n\t"
 439 //              "psubb %%mm6, %%mm2                             \n\t"
 440                 "movq %%mm2, (%%ebx)                            \n\t"
 441
 442                 "paddb %%mm6, %%mm5                             \n\t"
 443                 "psrlw $2, %%mm5                                \n\t"
 444                 "pand "MANGLE(b3F)", %%mm5                      \n\t"
 445                 "psubb "MANGLE(b20)", %%mm5                     \n\t" // (l5-l4)/8
 446
 447                 "movq (%%eax, %1, 2), %%mm2                     \n\t"
 448                 "paddb %%mm6, %%mm2                             \n\t" // line 3 + 0x80
 449                 "paddsb %%mm5, %%mm2                            \n\t"
 450                 "psubb %%mm6, %%mm2                             \n\t"
 451                 "movq %%mm2, (%%eax, %1, 2)                     \n\t"
 452
 453                 "movq (%%ebx, %1), %%mm2                        \n\t"
 454                 "paddb %%mm6, %%mm2                             \n\t" // line 6 + 0x80
 455                 "psubsb %%mm5, %%mm2                            \n\t"
 456                 "psubb %%mm6, %%mm2                             \n\t"
 457                 "movq %%mm2, (%%ebx, %1)                        \n\t"
 458
 459                 :
 460                 : "r" (src), "r" (stride)
 461                 : "%eax", "%ebx"
 462         );
 463 #else
 464         const int l1= stride;
 465         const int l2= stride + l1;
 466         const int l3= stride + l2;
 467         const int l4= stride + l3;
 468         const int l5= stride + l4;
 469         const int l6= stride + l5;
 470 //      const int l7= stride + l6;
 471 //      const int l8= stride + l7;
 472 //      const int l9= stride + l8;
 473         int x;
 474         const int QP15= QP + (QP>>2);
 475         src+= stride*3;
 476         for(x=0; x<BLOCK_SIZE; x++)
 477         {
 478                 const int v = (src[x+l5] - src[x+l4]);
 479                 if(ABS(v) < QP15)
 480                 {
 481                         src[x+l3] +=v>>3;
 482                         src[x+l4] +=v>>1;
 483                         src[x+l5] -=v>>1;
 484                         src[x+l6] -=v>>3;
 485
 486                 }
 487         }
 488
 489 #endif
 490 }
 491
 492 /**
 493  * Experimental Filter 1
 494  * will not damage linear gradients
 495  * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
 496  * can only smooth blocks at the expected locations (it cant smooth them if they did move)
 497  * MMX2 version does correct clipping C version doesnt
 498  */
 499 static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, int QP)
 500 {
 501 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
 502         src+= stride*3;
 503
 504         asm volatile(
 505                 "pxor %%mm7, %%mm7                              \n\t" // 0
 506                 "leal (%0, %1), %%eax                           \n\t"
 507                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
 508 //      0       1       2       3       4       5       6       7       8       9
 509 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
 510                 "movq (%%eax, %1, 2), %%mm0                     \n\t" // line 3
 511                 "movq (%0, %1, 4), %%mm1                        \n\t" // line 4
 512                 "movq %%mm1, %%mm2                              \n\t" // line 4
 513                 "psubusb %%mm0, %%mm1                           \n\t"
 514                 "psubusb %%mm2, %%mm0                           \n\t"
 515                 "por %%mm1, %%mm0                               \n\t" // |l2 - l3|
 516                 "movq (%%ebx), %%mm3                            \n\t" // line 5
 517                 "movq (%%ebx, %1), %%mm4                        \n\t" // line 6
 518                 "movq %%mm3, %%mm5                              \n\t" // line 5
 519                 "psubusb %%mm4, %%mm3                           \n\t"
 520                 "psubusb %%mm5, %%mm4                           \n\t"
 521                 "por %%mm4, %%mm3                               \n\t" // |l5 - l6|
 522                 PAVGB(%%mm3, %%mm0)                                   // (|l2 - l3| + |l5 - l6|)/2
 523                 "movq %%mm2, %%mm1                              \n\t" // line 4
 524                 "psubusb %%mm5, %%mm2                           \n\t"
 525                 "movq %%mm2, %%mm4                              \n\t"
 526                 "pcmpeqb %%mm7, %%mm2                           \n\t" // (l4 - l5) <= 0 ? -1 : 0
 527                 "psubusb %%mm1, %%mm5                           \n\t"
 528                 "por %%mm5, %%mm4                               \n\t" // |l4 - l5|
 529                 "psubusb %%mm0, %%mm4           \n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
 530                 "movq %%mm4, %%mm3                              \n\t" // d
 531                 "movq "MANGLE(pQPb)", %%mm0                     \n\t"
 532                 "paddusb %%mm0, %%mm0                           \n\t"
 533                 "psubusb %%mm0, %%mm4                           \n\t"
 534                 "pcmpeqb %%mm7, %%mm4                           \n\t" // d <= QP ? -1 : 0
 535                 "psubusb "MANGLE(b01)", %%mm3                   \n\t"
 536                 "pand %%mm4, %%mm3                              \n\t" // d <= QP ? d : 0
 537
 538                 PAVGB(%%mm7, %%mm3)                                   // d/2
 539                 "movq %%mm3, %%mm1                              \n\t" // d/2
 540                 PAVGB(%%mm7, %%mm3)                                   // d/4
 541                 PAVGB(%%mm1, %%mm3)                                   // 3*d/8
 542
 543                 "movq (%0, %1, 4), %%mm0                        \n\t" // line 4
 544                 "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
 545                 "psubusb %%mm3, %%mm0                           \n\t"
 546                 "pxor %%mm2, %%mm0                              \n\t"
 547                 "movq %%mm0, (%0, %1, 4)                        \n\t" // line 4
 548
 549                 "movq (%%ebx), %%mm0                            \n\t" // line 5
 550                 "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
 551                 "paddusb %%mm3, %%mm0                           \n\t"
 552                 "pxor %%mm2, %%mm0                              \n\t"
 553                 "movq %%mm0, (%%ebx)                            \n\t" // line 5
 554
 555                 PAVGB(%%mm7, %%mm1)                                   // d/4
 556
 557                 "movq (%%eax, %1, 2), %%mm0                     \n\t" // line 3
 558                 "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
 559                 "psubusb %%mm1, %%mm0                           \n\t"
 560                 "pxor %%mm2, %%mm0                              \n\t"
 561                 "movq %%mm0, (%%eax, %1, 2)                     \n\t" // line 3
 562
 563                 "movq (%%ebx, %1), %%mm0                        \n\t" // line 6
 564                 "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
 565                 "paddusb %%mm1, %%mm0                           \n\t"
 566                 "pxor %%mm2, %%mm0                              \n\t"
 567                 "movq %%mm0, (%%ebx, %1)                        \n\t" // line 6
 568
 569                 PAVGB(%%mm7, %%mm1)                                   // d/8
 570
 571                 "movq (%%eax, %1), %%mm0                        \n\t" // line 2
 572                 "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
 573                 "psubusb %%mm1, %%mm0                           \n\t"
 574                 "pxor %%mm2, %%mm0                              \n\t"
 575                 "movq %%mm0, (%%eax, %1)                        \n\t" // line 2
 576
 577                 "movq (%%ebx, %1, 2), %%mm0                     \n\t" // line 7
 578                 "pxor %%mm2, %%mm0                              \n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
 579                 "paddusb %%mm1, %%mm0                           \n\t"
 580                 "pxor %%mm2, %%mm0                              \n\t"
 581                 "movq %%mm0, (%%ebx, %1, 2)                     \n\t" // line 7
 582
 583                 :
 584                 : "r" (src), "r" (stride)
 585                 : "%eax", "%ebx"
 586         );
 587 #else
 588
 589         const int l1= stride;
 590         const int l2= stride + l1;
 591         const int l3= stride + l2;
 592         const int l4= stride + l3;
 593         const int l5= stride + l4;
 594         const int l6= stride + l5;
 595         const int l7= stride + l6;
 596 //      const int l8= stride + l7;
 597 //      const int l9= stride + l8;
 598         int x;
 599
 600         src+= stride*3;
 601         for(x=0; x<BLOCK_SIZE; x++)
 602         {
 603                 int a= src[l3] - src[l4];
 604                 int b= src[l4] - src[l5];
 605                 int c= src[l5] - src[l6];
 606
 607                 int d= ABS(b) - ((ABS(a) + ABS(c))>>1);
 608                 d= MAX(d, 0);
 609
 610                 if(d < QP*2)
 611                 {
 612                         int v = d * SIGN(-b);
 613
 614                         src[l2] +=v>>3;
 615                         src[l3] +=v>>2;
 616                         src[l4] +=(3*v)>>3;
 617                         src[l5] -=(3*v)>>3;
 618                         src[l6] -=v>>2;
 619                         src[l7] -=v>>3;
 620
 621                 }
 622                 src++;
 623         }
 624         /*
 625         const int l1= stride;
 626         const int l2= stride + l1;
 627         const int l3= stride + l2;
 628         const int l4= stride + l3;
 629         const int l5= stride + l4;
 630         const int l6= stride + l5;
 631         const int l7= stride + l6;
 632         const int l8= stride + l7;
 633         const int l9= stride + l8;
 634         for(int x=0; x<BLOCK_SIZE; x++)
 635         {
 636                 int v2= src[l2];
 637                 int v3= src[l3];
 638                 int v4= src[l4];
 639                 int v5= src[l5];
 640                 int v6= src[l6];
 641                 int v7= src[l7];
 642
 643                 if(ABS(v4-v5)<QP &&  ABS(v4-v5) - (ABS(v3-v4) + ABS(v5-v6))>0 )
 644                 {
 645                         src[l3] = (6*v2 + 4*v3 + 3*v4 + 2*v5 + v6         )/16;
 646                         src[l4] = (3*v2 + 3*v3 + 4*v4 + 3*v5 + 2*v6 + v7  )/16;
 647                         src[l5] = (1*v2 + 2*v3 + 3*v4 + 4*v5 + 3*v6 + 3*v7)/16;
 648                         src[l6] = (       1*v3 + 2*v4 + 3*v5 + 4*v6 + 6*v7)/16;
 649                 }
 650                 src++;
 651         }
 652 */
 653 #endif
 654 }
 655
 656 static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, int QP)
 657 {
 658 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
 659 /*
 660         uint8_t tmp[16];
 661         const int l1= stride;
 662         const int l2= stride + l1;
 663         const int l3= stride + l2;
 664         const int l4= (int)tmp - (int)src - stride*3;
 665         const int l5= (int)tmp - (int)src - stride*3 + 8;
 666         const int l6= stride*3 + l3;
 667         const int l7= stride + l6;
 668         const int l8= stride + l7;
 669
 670         memcpy(tmp, src+stride*7, 8);
 671         memcpy(tmp+8, src+stride*8, 8);
 672 */
 673         src+= stride*4;
 674         asm volatile(
 675
 676 #if 0 //sligtly more accurate and slightly slower
 677                 "pxor %%mm7, %%mm7                              \n\t" // 0
 678                 "leal (%0, %1), %%eax                           \n\t"
 679                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
 680 //      0       1       2       3       4       5       6       7
 681 //      %0      %0+%1   %0+2%1  eax+2%1 %0+4%1  eax+4%1 ebx+%1  ebx+2%1
 682 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1
 683
 684
 685                 "movq (%0, %1, 2), %%mm0                        \n\t" // l2
 686                 "movq (%0), %%mm1                               \n\t" // l0
 687                 "movq %%mm0, %%mm2                              \n\t" // l2
 688                 PAVGB(%%mm7, %%mm0)                                   // ~l2/2
 689                 PAVGB(%%mm1, %%mm0)                                   // ~(l2 + 2l0)/4
 690                 PAVGB(%%mm2, %%mm0)                                   // ~(5l2 + 2l0)/8
 691
 692                 "movq (%%eax), %%mm1                            \n\t" // l1
 693                 "movq (%%eax, %1, 2), %%mm3                     \n\t" // l3
 694                 "movq %%mm1, %%mm4                              \n\t" // l1
 695                 PAVGB(%%mm7, %%mm1)                                   // ~l1/2
 696                 PAVGB(%%mm3, %%mm1)                                   // ~(l1 + 2l3)/4
 697                 PAVGB(%%mm4, %%mm1)                                   // ~(5l1 + 2l3)/8
 698
 699                 "movq %%mm0, %%mm4                              \n\t" // ~(5l2 + 2l0)/8
 700                 "psubusb %%mm1, %%mm0                           \n\t"
 701                 "psubusb %%mm4, %%mm1                           \n\t"
 702                 "por %%mm0, %%mm1                               \n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8
 703 // mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0
 704
 705                 "movq (%0, %1, 4), %%mm0                        \n\t" // l4
 706                 "movq %%mm0, %%mm4                              \n\t" // l4
 707                 PAVGB(%%mm7, %%mm0)                                   // ~l4/2
 708                 PAVGB(%%mm2, %%mm0)                                   // ~(l4 + 2l2)/4
 709                 PAVGB(%%mm4, %%mm0)                                   // ~(5l4 + 2l2)/8
 710
 711                 "movq (%%ebx), %%mm2                            \n\t" // l5
 712                 "movq %%mm3, %%mm5                              \n\t" // l3
 713                 PAVGB(%%mm7, %%mm3)                                   // ~l3/2
 714                 PAVGB(%%mm2, %%mm3)                                   // ~(l3 + 2l5)/4
 715                 PAVGB(%%mm5, %%mm3)                                   // ~(5l3 + 2l5)/8
 716
 717                 "movq %%mm0, %%mm6                              \n\t" // ~(5l4 + 2l2)/8
 718                 "psubusb %%mm3, %%mm0                           \n\t"
 719                 "psubusb %%mm6, %%mm3                           \n\t"
 720                 "por %%mm0, %%mm3                               \n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8
 721                 "pcmpeqb %%mm7, %%mm0                           \n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5)
 722 // mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0
 723
 724                 "movq (%%ebx, %1), %%mm6                        \n\t" // l6
 725                 "movq %%mm6, %%mm5                              \n\t" // l6
 726                 PAVGB(%%mm7, %%mm6)                                   // ~l6/2
 727                 PAVGB(%%mm4, %%mm6)                                   // ~(l6 + 2l4)/4
 728                 PAVGB(%%mm5, %%mm6)                                   // ~(5l6 + 2l4)/8
 729
 730                 "movq (%%ebx, %1, 2), %%mm5                     \n\t" // l7
 731                 "movq %%mm2, %%mm4                              \n\t" // l5
 732                 PAVGB(%%mm7, %%mm2)                                   // ~l5/2
 733                 PAVGB(%%mm5, %%mm2)                                   // ~(l5 + 2l7)/4
 734                 PAVGB(%%mm4, %%mm2)                                   // ~(5l5 + 2l7)/8
 735
 736                 "movq %%mm6, %%mm4                              \n\t" // ~(5l6 + 2l4)/8
 737                 "psubusb %%mm2, %%mm6                           \n\t"
 738                 "psubusb %%mm4, %%mm2                           \n\t"
 739                 "por %%mm6, %%mm2                               \n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8
 740 // mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0
 741
 742
 743                 PMINUB(%%mm2, %%mm1, %%mm4)                           // MIN(|lenergy|,|renergy|)/8
 744                 "movq "MANGLE(pQPb)", %%mm4                     \n\t" // QP //FIXME QP+1 ?
 745                 "paddusb "MANGLE(b01)", %%mm4                   \n\t"
 746                 "pcmpgtb %%mm3, %%mm4                           \n\t" // |menergy|/8 < QP
 747                 "psubusb %%mm1, %%mm3                           \n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8
 748                 "pand %%mm4, %%mm3                              \n\t"
 749
 750                 "movq %%mm3, %%mm1                              \n\t"
 751 //              "psubusb "MANGLE(b01)", %%mm3                   \n\t"
 752                 PAVGB(%%mm7, %%mm3)
 753                 PAVGB(%%mm7, %%mm3)
 754                 "paddusb %%mm1, %%mm3                           \n\t"
 755 //              "paddusb "MANGLE(b01)", %%mm3                   \n\t"
 756
 757                 "movq (%%eax, %1, 2), %%mm6                     \n\t" //l3
 758                 "movq (%0, %1, 4), %%mm5                        \n\t" //l4
 759                 "movq (%0, %1, 4), %%mm4                        \n\t" //l4
 760                 "psubusb %%mm6, %%mm5                           \n\t"
 761                 "psubusb %%mm4, %%mm6                           \n\t"
 762                 "por %%mm6, %%mm5                               \n\t" // |l3-l4|
 763                 "pcmpeqb %%mm7, %%mm6                           \n\t" // SIGN(l3-l4)
 764                 "pxor %%mm6, %%mm0                              \n\t"
 765                 "pand %%mm0, %%mm3                              \n\t"
 766                 PMINUB(%%mm5, %%mm3, %%mm0)
 767
 768                 "psubusb "MANGLE(b01)", %%mm3                   \n\t"
 769                 PAVGB(%%mm7, %%mm3)
 770
 771                 "movq (%%eax, %1, 2), %%mm0                     \n\t"
 772                 "movq (%0, %1, 4), %%mm2                        \n\t"
 773                 "pxor %%mm6, %%mm0                              \n\t"
 774                 "pxor %%mm6, %%mm2                              \n\t"
 775                 "psubb %%mm3, %%mm0                             \n\t"
 776                 "paddb %%mm3, %%mm2                             \n\t"
 777                 "pxor %%mm6, %%mm0                              \n\t"
 778                 "pxor %%mm6, %%mm2                              \n\t"
 779                 "movq %%mm0, (%%eax, %1, 2)                     \n\t"
 780                 "movq %%mm2, (%0, %1, 4)                        \n\t"
 781 #endif
 782
 783                 "leal (%0, %1), %%eax                           \n\t"
 784                 "pcmpeqb %%mm6, %%mm6                           \n\t" // -1
 785 //      0       1       2       3       4       5       6       7
 786 //      %0      %0+%1   %0+2%1  eax+2%1 %0+4%1  eax+4%1 ebx+%1  ebx+2%1
 787 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1
 788
 789
 790                 "movq (%%eax, %1, 2), %%mm1                     \n\t" // l3
 791                 "movq (%0, %1, 4), %%mm0                        \n\t" // l4
 792                 "pxor %%mm6, %%mm1                              \n\t" // -l3-1
 793                 PAVGB(%%mm1, %%mm0)                                   // -q+128 = (l4-l3+256)/2
 794 // mm1=-l3-1, mm0=128-q
 795
 796                 "movq (%%eax, %1, 4), %%mm2                     \n\t" // l5
 797                 "movq (%%eax, %1), %%mm3                        \n\t" // l2
 798                 "pxor %%mm6, %%mm2                              \n\t" // -l5-1
 799                 "movq %%mm2, %%mm5                              \n\t" // -l5-1
 800                 "movq "MANGLE(b80)", %%mm4                      \n\t" // 128
 801                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
 802                 PAVGB(%%mm3, %%mm2)                                   // (l2-l5+256)/2
 803                 PAVGB(%%mm0, %%mm4)                                   // ~(l4-l3)/4 + 128
 804                 PAVGB(%%mm2, %%mm4)                                   // ~(l2-l5)/4 +(l4-l3)/8 + 128
 805                 PAVGB(%%mm0, %%mm4)                                   // ~(l2-l5)/8 +5(l4-l3)/16 + 128
 806 // mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1
 807
 808                 "movq (%%eax), %%mm2                            \n\t" // l1
 809                 "pxor %%mm6, %%mm2                              \n\t" // -l1-1
 810                 PAVGB(%%mm3, %%mm2)                                   // (l2-l1+256)/2
 811                 PAVGB((%0), %%mm1)                                    // (l0-l3+256)/2
 812                 "movq "MANGLE(b80)", %%mm3                      \n\t" // 128
 813                 PAVGB(%%mm2, %%mm3)                                   // ~(l2-l1)/4 + 128
 814                 PAVGB(%%mm1, %%mm3)                                   // ~(l0-l3)/4 +(l2-l1)/8 + 128
 815                 PAVGB(%%mm2, %%mm3)                                   // ~(l0-l3)/8 +5(l2-l1)/16 + 128
 816 // mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1
 817
 818                 PAVGB((%%ebx, %1), %%mm5)                             // (l6-l5+256)/2
 819                 "movq (%%ebx, %1, 2), %%mm1                     \n\t" // l7
 820                 "pxor %%mm6, %%mm1                              \n\t" // -l7-1
 821                 PAVGB((%0, %1, 4), %%mm1)                             // (l4-l7+256)/2
 822                 "movq "MANGLE(b80)", %%mm2                      \n\t" // 128
 823                 PAVGB(%%mm5, %%mm2)                                   // ~(l6-l5)/4 + 128
 824                 PAVGB(%%mm1, %%mm2)                                   // ~(l4-l7)/4 +(l6-l5)/8 + 128
 825                 PAVGB(%%mm5, %%mm2)                                   // ~(l4-l7)/8 +5(l6-l5)/16 + 128
 826 // mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128
 827
 828                 "movq "MANGLE(b00)", %%mm1                      \n\t" // 0
 829                 "movq "MANGLE(b00)", %%mm5                      \n\t" // 0
 830                 "psubb %%mm2, %%mm1                             \n\t" // 128 - renergy/16
 831                 "psubb %%mm3, %%mm5                             \n\t" // 128 - lenergy/16
 832                 PMAXUB(%%mm1, %%mm2)                                  // 128 + |renergy/16|
 833                 PMAXUB(%%mm5, %%mm3)                                  // 128 + |lenergy/16|
 834                 PMINUB(%%mm2, %%mm3, %%mm1)                           // 128 + MIN(|lenergy|,|renergy|)/16
 835
 836 // mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128
 837
 838                 "movq "MANGLE(b00)", %%mm7                      \n\t" // 0
 839                 "movq "MANGLE(pQPb)", %%mm2                     \n\t" // QP
 840                 PAVGB(%%mm6, %%mm2)                                   // 128 + QP/2
 841                 "psubb %%mm6, %%mm2                             \n\t"
 842
 843                 "movq %%mm4, %%mm1                              \n\t"
 844                 "pcmpgtb %%mm7, %%mm1                           \n\t" // SIGN(menergy)
 845                 "pxor %%mm1, %%mm4                              \n\t"
 846                 "psubb %%mm1, %%mm4                             \n\t" // 128 + |menergy|/16
 847                 "pcmpgtb %%mm4, %%mm2                           \n\t" // |menergy|/16 < QP/2
 848                 "psubusb %%mm3, %%mm4                           \n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16
 849 // mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16
 850
 851                 "movq %%mm4, %%mm3                              \n\t" // d
 852                 "psubusb "MANGLE(b01)", %%mm4                   \n\t"
 853                 PAVGB(%%mm7, %%mm4)                                   // d/32
 854                 PAVGB(%%mm7, %%mm4)                                   // (d + 32)/64
 855                 "paddb %%mm3, %%mm4                             \n\t" // 5d/64
 856                 "pand %%mm2, %%mm4                              \n\t"
 857
 858                 "movq "MANGLE(b80)", %%mm5                      \n\t" // 128
 859                 "psubb %%mm0, %%mm5                             \n\t" // q
 860                 "paddsb %%mm6, %%mm5                            \n\t" // fix bad rounding
 861                 "pcmpgtb %%mm5, %%mm7                           \n\t" // SIGN(q)
 862                 "pxor %%mm7, %%mm5                              \n\t"
 863
 864                 PMINUB(%%mm5, %%mm4, %%mm3)                           // MIN(|q|, 5d/64)
 865                 "pxor %%mm1, %%mm7                              \n\t" // SIGN(d*q)
 866
 867                 "pand %%mm7, %%mm4                              \n\t"
 868                 "movq (%%eax, %1, 2), %%mm0                     \n\t"
 869                 "movq (%0, %1, 4), %%mm2                        \n\t"
 870                 "pxor %%mm1, %%mm0                              \n\t"
 871                 "pxor %%mm1, %%mm2                              \n\t"
 872                 "paddb %%mm4, %%mm0                             \n\t"
 873                 "psubb %%mm4, %%mm2                             \n\t"
 874                 "pxor %%mm1, %%mm0                              \n\t"
 875                 "pxor %%mm1, %%mm2                              \n\t"
 876                 "movq %%mm0, (%%eax, %1, 2)                     \n\t"
 877                 "movq %%mm2, (%0, %1, 4)                        \n\t"
 878
 879                 :
 880                 : "r" (src), "r" (stride)
 881                 : "%eax", "%ebx"
 882         );
 883
 884 /*
 885         {
 886         int x;
 887         src-= stride;
 888         for(x=0; x<BLOCK_SIZE; x++)
 889         {
 890                 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
 891                 if(ABS(middleEnergy)< 8*QP)
 892                 {
 893                         const int q=(src[l4] - src[l5])/2;
 894                         const int leftEnergy=  5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
 895                         const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
 896
 897                         int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
 898                         d= MAX(d, 0);
 899
 900                         d= (5*d + 32) >> 6;
 901                         d*= SIGN(-middleEnergy);
 902
 903                         if(q>0)
 904                         {
 905                                 d= d<0 ? 0 : d;
 906                                 d= d>q ? q : d;
 907                         }
 908                         else
 909                         {
 910                                 d= d>0 ? 0 : d;
 911                                 d= d<q ? q : d;
 912                         }
 913
 914                         src[l4]-= d;
 915                         src[l5]+= d;
 916                 }
 917                 src++;
 918         }
 919 src-=8;
 920         for(x=0; x<8; x++)
 921         {
 922                 int y;
 923                 for(y=4; y<6; y++)
 924                 {
 925                         int d= src[x+y*stride] - tmp[x+(y-4)*8];
 926                         int ad= ABS(d);
 927                         static int max=0;
 928                         static int sum=0;
 929                         static int num=0;
 930                         static int bias=0;
 931
 932                         if(max<ad) max=ad;
 933                         sum+= ad>3 ? 1 : 0;
 934                         if(ad>3)
 935                         {
 936                                 src[0] = src[7] = src[stride*7] = src[(stride+1)*7]=255;
 937                         }
 938                         if(y==4) bias+=d;
 939                         num++;
 940                         if(num%1000000 == 0)
 941                         {
 942                                 printf(" %d %d %d %d\n", num, sum, max, bias);
 943                         }
 944                 }
 945         }
 946 }
 947 */
 948 #elif defined (HAVE_MMX)
 949         src+= stride*4;
 950
 951         asm volatile(
 952                 "pxor %%mm7, %%mm7                              \n\t"
 953                 "leal (%0, %1), %%eax                           \n\t"
 954                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
 955 //      0       1       2       3       4       5       6       7
 956 //      %0      %0+%1   %0+2%1  eax+2%1 %0+4%1  eax+4%1 ebx+%1  ebx+2%1
 957 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1
 958
 959                 "movq (%0), %%mm0                               \n\t"
 960                 "movq %%mm0, %%mm1                              \n\t"
 961                 "punpcklbw %%mm7, %%mm0                         \n\t" // low part of line 0
 962                 "punpckhbw %%mm7, %%mm1                         \n\t" // high part of line 0
 963
 964                 "movq (%%eax), %%mm2                            \n\t"
 965                 "movq %%mm2, %%mm3                              \n\t"
 966                 "punpcklbw %%mm7, %%mm2                         \n\t" // low part of line 1
 967                 "punpckhbw %%mm7, %%mm3                         \n\t" // high part of line 1
 968
 969                 "movq (%%eax, %1), %%mm4                        \n\t"
 970                 "movq %%mm4, %%mm5                              \n\t"
 971                 "punpcklbw %%mm7, %%mm4                         \n\t" // low part of line 2
 972                 "punpckhbw %%mm7, %%mm5                         \n\t" // high part of line 2
 973
 974                 "paddw %%mm0, %%mm0                             \n\t" // 2L0
 975                 "paddw %%mm1, %%mm1                             \n\t" // 2H0
 976                 "psubw %%mm4, %%mm2                             \n\t" // L1 - L2
 977                 "psubw %%mm5, %%mm3                             \n\t" // H1 - H2
 978                 "psubw %%mm2, %%mm0                             \n\t" // 2L0 - L1 + L2
 979                 "psubw %%mm3, %%mm1                             \n\t" // 2H0 - H1 + H2
 980
 981                 "psllw $2, %%mm2                                \n\t" // 4L1 - 4L2
 982                 "psllw $2, %%mm3                                \n\t" // 4H1 - 4H2
 983                 "psubw %%mm2, %%mm0                             \n\t" // 2L0 - 5L1 + 5L2
 984                 "psubw %%mm3, %%mm1                             \n\t" // 2H0 - 5H1 + 5H2
 985
 986                 "movq (%%eax, %1, 2), %%mm2                     \n\t"
 987                 "movq %%mm2, %%mm3                              \n\t"
 988                 "punpcklbw %%mm7, %%mm2                         \n\t" // L3
 989                 "punpckhbw %%mm7, %%mm3                         \n\t" // H3
 990
 991                 "psubw %%mm2, %%mm0                             \n\t" // 2L0 - 5L1 + 5L2 - L3
 992                 "psubw %%mm3, %%mm1                             \n\t" // 2H0 - 5H1 + 5H2 - H3
 993                 "psubw %%mm2, %%mm0                             \n\t" // 2L0 - 5L1 + 5L2 - 2L3
 994                 "psubw %%mm3, %%mm1                             \n\t" // 2H0 - 5H1 + 5H2 - 2H3
 995                 "movq %%mm0, "MANGLE(temp0)"                    \n\t" // 2L0 - 5L1 + 5L2 - 2L3
 996                 "movq %%mm1, "MANGLE(temp1)"                    \n\t" // 2H0 - 5H1 + 5H2 - 2H3
 997
 998                 "movq (%0, %1, 4), %%mm0                        \n\t"
 999                 "movq %%mm0, %%mm1                              \n\t"
1000                 "punpcklbw %%mm7, %%mm0                         \n\t" // L4
1001                 "punpckhbw %%mm7, %%mm1                         \n\t" // H4
1002
1003                 "psubw %%mm0, %%mm2                             \n\t" // L3 - L4
1004                 "psubw %%mm1, %%mm3                             \n\t" // H3 - H4
1005                 "movq %%mm2, "MANGLE(temp2)"                    \n\t" // L3 - L4
1006                 "movq %%mm3, "MANGLE(temp3)"                    \n\t" // H3 - H4
1007                 "paddw %%mm4, %%mm4                             \n\t" // 2L2
1008                 "paddw %%mm5, %%mm5                             \n\t" // 2H2
1009                 "psubw %%mm2, %%mm4                             \n\t" // 2L2 - L3 + L4
1010                 "psubw %%mm3, %%mm5                             \n\t" // 2H2 - H3 + H4
1011
1012                 "psllw $2, %%mm2                                \n\t" // 4L3 - 4L4
1013                 "psllw $2, %%mm3                                \n\t" // 4H3 - 4H4
1014                 "psubw %%mm2, %%mm4                             \n\t" // 2L2 - 5L3 + 5L4
1015                 "psubw %%mm3, %%mm5                             \n\t" // 2H2 - 5H3 + 5H4
1016 //50 opcodes so far
1017                 "movq (%%ebx), %%mm2                            \n\t"
1018                 "movq %%mm2, %%mm3                              \n\t"
1019                 "punpcklbw %%mm7, %%mm2                         \n\t" // L5
1020                 "punpckhbw %%mm7, %%mm3                         \n\t" // H5
1021                 "psubw %%mm2, %%mm4                             \n\t" // 2L2 - 5L3 + 5L4 - L5
1022                 "psubw %%mm3, %%mm5                             \n\t" // 2H2 - 5H3 + 5H4 - H5
1023                 "psubw %%mm2, %%mm4                             \n\t" // 2L2 - 5L3 + 5L4 - 2L5
1024                 "psubw %%mm3, %%mm5                             \n\t" // 2H2 - 5H3 + 5H4 - 2H5
1025
1026                 "movq (%%ebx, %1), %%mm6                        \n\t"
1027                 "punpcklbw %%mm7, %%mm6                         \n\t" // L6
1028                 "psubw %%mm6, %%mm2                             \n\t" // L5 - L6
1029                 "movq (%%ebx, %1), %%mm6                        \n\t"
1030                 "punpckhbw %%mm7, %%mm6                         \n\t" // H6
1031                 "psubw %%mm6, %%mm3                             \n\t" // H5 - H6
1032
1033                 "paddw %%mm0, %%mm0                             \n\t" // 2L4
1034                 "paddw %%mm1, %%mm1                             \n\t" // 2H4
1035                 "psubw %%mm2, %%mm0                             \n\t" // 2L4 - L5 + L6
1036                 "psubw %%mm3, %%mm1                             \n\t" // 2H4 - H5 + H6
1037
1038                 "psllw $2, %%mm2                                \n\t" // 4L5 - 4L6
1039                 "psllw $2, %%mm3                                \n\t" // 4H5 - 4H6
1040                 "psubw %%mm2, %%mm0                             \n\t" // 2L4 - 5L5 + 5L6
1041                 "psubw %%mm3, %%mm1                             \n\t" // 2H4 - 5H5 + 5H6
1042
1043                 "movq (%%ebx, %1, 2), %%mm2                     \n\t"
1044                 "movq %%mm2, %%mm3                              \n\t"
1045                 "punpcklbw %%mm7, %%mm2                         \n\t" // L7
1046                 "punpckhbw %%mm7, %%mm3                         \n\t" // H7
1047
1048                 "paddw %%mm2, %%mm2                             \n\t" // 2L7
1049                 "paddw %%mm3, %%mm3                             \n\t" // 2H7
1050                 "psubw %%mm2, %%mm0                             \n\t" // 2L4 - 5L5 + 5L6 - 2L7
1051                 "psubw %%mm3, %%mm1                             \n\t" // 2H4 - 5H5 + 5H6 - 2H7
1052
1053                 "movq "MANGLE(temp0)", %%mm2                    \n\t" // 2L0 - 5L1 + 5L2 - 2L3
1054                 "movq "MANGLE(temp1)", %%mm3                    \n\t" // 2H0 - 5H1 + 5H2 - 2H3
1055
1056 #ifdef HAVE_MMX2
1057                 "movq %%mm7, %%mm6                              \n\t" // 0
1058                 "psubw %%mm0, %%mm6                             \n\t"
1059                 "pmaxsw %%mm6, %%mm0                            \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
1060                 "movq %%mm7, %%mm6                              \n\t" // 0
1061                 "psubw %%mm1, %%mm6                             \n\t"
1062                 "pmaxsw %%mm6, %%mm1                            \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
1063                 "movq %%mm7, %%mm6                              \n\t" // 0
1064                 "psubw %%mm2, %%mm6                             \n\t"
1065                 "pmaxsw %%mm6, %%mm2                            \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
1066                 "movq %%mm7, %%mm6                              \n\t" // 0
1067                 "psubw %%mm3, %%mm6                             \n\t"
1068                 "pmaxsw %%mm6, %%mm3                            \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
1069 #else
1070                 "movq %%mm7, %%mm6                              \n\t" // 0
1071                 "pcmpgtw %%mm0, %%mm6                           \n\t"
1072                 "pxor %%mm6, %%mm0                              \n\t"
1073                 "psubw %%mm6, %%mm0                             \n\t" // |2L4 - 5L5 + 5L6 - 2L7|
1074                 "movq %%mm7, %%mm6                              \n\t" // 0
1075                 "pcmpgtw %%mm1, %%mm6                           \n\t"
1076                 "pxor %%mm6, %%mm1                              \n\t"
1077                 "psubw %%mm6, %%mm1                             \n\t" // |2H4 - 5H5 + 5H6 - 2H7|
1078                 "movq %%mm7, %%mm6                              \n\t" // 0
1079                 "pcmpgtw %%mm2, %%mm6                           \n\t"
1080                 "pxor %%mm6, %%mm2                              \n\t"
1081                 "psubw %%mm6, %%mm2                             \n\t" // |2L0 - 5L1 + 5L2 - 2L3|
1082                 "movq %%mm7, %%mm6                              \n\t" // 0
1083                 "pcmpgtw %%mm3, %%mm6                           \n\t"
1084                 "pxor %%mm6, %%mm3                              \n\t"
1085                 "psubw %%mm6, %%mm3                             \n\t" // |2H0 - 5H1 + 5H2 - 2H3|
1086 #endif
1087
1088 #ifdef HAVE_MMX2
1089                 "pminsw %%mm2, %%mm0                            \n\t"
1090                 "pminsw %%mm3, %%mm1                            \n\t"
1091 #else
1092                 "movq %%mm0, %%mm6                              \n\t"
1093                 "psubusw %%mm2, %%mm6                           \n\t"
1094                 "psubw %%mm6, %%mm0                             \n\t"
1095                 "movq %%mm1, %%mm6                              \n\t"
1096                 "psubusw %%mm3, %%mm6                           \n\t"
1097                 "psubw %%mm6, %%mm1                             \n\t"
1098 #endif
1099
1100                 "movq %%mm7, %%mm6                              \n\t" // 0
1101                 "pcmpgtw %%mm4, %%mm6                           \n\t" // sign(2L2 - 5L3 + 5L4 - 2L5)
1102                 "pxor %%mm6, %%mm4                              \n\t"
1103                 "psubw %%mm6, %%mm4                             \n\t" // |2L2 - 5L3 + 5L4 - 2L5|
1104                 "pcmpgtw %%mm5, %%mm7                           \n\t" // sign(2H2 - 5H3 + 5H4 - 2H5)
1105                 "pxor %%mm7, %%mm5                              \n\t"
1106                 "psubw %%mm7, %%mm5                             \n\t" // |2H2 - 5H3 + 5H4 - 2H5|
1107 // 100 opcodes
1108                 "movd %2, %%mm2                                 \n\t" // QP
1109                 "punpcklwd %%mm2, %%mm2                         \n\t"
1110                 "punpcklwd %%mm2, %%mm2                         \n\t"
1111                 "psllw $3, %%mm2                                \n\t" // 8QP
1112                 "movq %%mm2, %%mm3                              \n\t" // 8QP
1113                 "pcmpgtw %%mm4, %%mm2                           \n\t"
1114                 "pcmpgtw %%mm5, %%mm3                           \n\t"
1115                 "pand %%mm2, %%mm4                              \n\t"
1116                 "pand %%mm3, %%mm5                              \n\t"
1117
1118
1119                 "psubusw %%mm0, %%mm4                           \n\t" // hd
1120                 "psubusw %%mm1, %%mm5                           \n\t" // ld
1121
1122
1123                 "movq "MANGLE(w05)", %%mm2                      \n\t" // 5
1124                 "pmullw %%mm2, %%mm4                            \n\t"
1125                 "pmullw %%mm2, %%mm5                            \n\t"
1126                 "movq "MANGLE(w20)", %%mm2                      \n\t" // 32
1127                 "paddw %%mm2, %%mm4                             \n\t"
1128                 "paddw %%mm2, %%mm5                             \n\t"
1129                 "psrlw $6, %%mm4                                \n\t"
1130                 "psrlw $6, %%mm5                                \n\t"
1131
1132 /*
1133                 "movq w06, %%mm2                                \n\t" // 6
1134                 "paddw %%mm2, %%mm4                             \n\t"
1135                 "paddw %%mm2, %%mm5                             \n\t"
1136                 "movq w1400, %%mm2                              \n\t" // 1400h = 5120 = 5/64*2^16
1137 //FIXME if *5/64 is supposed to be /13 then we should use 5041 instead of 5120
1138                 "pmulhw %%mm2, %%mm4                            \n\t" // hd/13
1139                 "pmulhw %%mm2, %%mm5                            \n\t" // ld/13
1140 */
1141
1142                 "movq "MANGLE(temp2)", %%mm0                    \n\t" // L3 - L4
1143                 "movq "MANGLE(temp3)", %%mm1                    \n\t" // H3 - H4
1144
1145                 "pxor %%mm2, %%mm2                              \n\t"
1146                 "pxor %%mm3, %%mm3                              \n\t"
1147
1148                 "pcmpgtw %%mm0, %%mm2                           \n\t" // sign (L3-L4)
1149                 "pcmpgtw %%mm1, %%mm3                           \n\t" // sign (H3-H4)
1150                 "pxor %%mm2, %%mm0                              \n\t"
1151                 "pxor %%mm3, %%mm1                              \n\t"
1152                 "psubw %%mm2, %%mm0                             \n\t" // |L3-L4|
1153                 "psubw %%mm3, %%mm1                             \n\t" // |H3-H4|
1154                 "psrlw $1, %%mm0                                \n\t" // |L3 - L4|/2
1155                 "psrlw $1, %%mm1                                \n\t" // |H3 - H4|/2
1156
1157                 "pxor %%mm6, %%mm2                              \n\t"
1158                 "pxor %%mm7, %%mm3                              \n\t"
1159                 "pand %%mm2, %%mm4                              \n\t"
1160                 "pand %%mm3, %%mm5                              \n\t"
1161
1162 #ifdef HAVE_MMX2
1163                 "pminsw %%mm0, %%mm4                            \n\t"
1164                 "pminsw %%mm1, %%mm5                            \n\t"
1165 #else
1166                 "movq %%mm4, %%mm2                              \n\t"
1167                 "psubusw %%mm0, %%mm2                           \n\t"
1168                 "psubw %%mm2, %%mm4                             \n\t"
1169                 "movq %%mm5, %%mm2                              \n\t"
1170                 "psubusw %%mm1, %%mm2                           \n\t"
1171                 "psubw %%mm2, %%mm5                             \n\t"
1172 #endif
1173                 "pxor %%mm6, %%mm4                              \n\t"
1174                 "pxor %%mm7, %%mm5                              \n\t"
1175                 "psubw %%mm6, %%mm4                             \n\t"
1176                 "psubw %%mm7, %%mm5                             \n\t"
1177                 "packsswb %%mm5, %%mm4                          \n\t"
1178                 "movq (%%eax, %1, 2), %%mm0                     \n\t"
1179                 "paddb   %%mm4, %%mm0                           \n\t"
1180                 "movq %%mm0, (%%eax, %1, 2)                     \n\t"
1181                 "movq (%0, %1, 4), %%mm0                        \n\t"
1182                 "psubb %%mm4, %%mm0                             \n\t"
1183                 "movq %%mm0, (%0, %1, 4)                        \n\t"
1184
1185                 :
1186                 : "r" (src), "r" (stride), "r" (QP)
1187                 : "%eax", "%ebx"
1188         );
1189 #else
1190         const int l1= stride;
1191         const int l2= stride + l1;
1192         const int l3= stride + l2;
1193         const int l4= stride + l3;
1194         const int l5= stride + l4;
1195         const int l6= stride + l5;
1196         const int l7= stride + l6;
1197         const int l8= stride + l7;
1198 //      const int l9= stride + l8;
1199         int x;
1200         src+= stride*3;
1201         for(x=0; x<BLOCK_SIZE; x++)
1202         {
1203                 const int middleEnergy= 5*(src[l5] - src[l4]) + 2*(src[l3] - src[l6]);
1204                 if(ABS(middleEnergy) < 8*QP)
1205                 {
1206                         const int q=(src[l4] - src[l5])/2;
1207                         const int leftEnergy=  5*(src[l3] - src[l2]) + 2*(src[l1] - src[l4]);
1208                         const int rightEnergy= 5*(src[l7] - src[l6]) + 2*(src[l5] - src[l8]);
1209
1210                         int d= ABS(middleEnergy) - MIN( ABS(leftEnergy), ABS(rightEnergy) );
1211                         d= MAX(d, 0);
1212
1213                         d= (5*d + 32) >> 6;
1214                         d*= SIGN(-middleEnergy);
1215
1216                         if(q>0)
1217                         {
1218                                 d= d<0 ? 0 : d;
1219                                 d= d>q ? q : d;
1220                         }
1221                         else
1222                         {
1223                                 d= d>0 ? 0 : d;
1224                                 d= d<q ? q : d;
1225                         }
1226
1227                         src[l4]-= d;
1228                         src[l5]+= d;
1229                 }
1230                 src++;
1231         }
1232 #endif
1233 }
1234
1235 static inline void RENAME(dering)(uint8_t src[], int stride, int QP)
1236 {
1237 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1238         asm volatile(
1239                 "movq "MANGLE(pQPb)", %%mm0                     \n\t"
1240                 "paddusb %%mm0, %%mm0                           \n\t"
1241                 "movq %%mm0, "MANGLE(pQPb2)"                    \n\t"
1242
1243                 "leal (%0, %1), %%eax                           \n\t"
1244                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1245 //      0       1       2       3       4       5       6       7       8       9
1246 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1247
1248                 "pcmpeqb %%mm7, %%mm7                           \n\t"
1249                 "pxor %%mm6, %%mm6                              \n\t"
1250 #undef FIND_MIN_MAX
1251 #ifdef HAVE_MMX2
1252 #define FIND_MIN_MAX(addr)\
1253                 "movq " #addr ", %%mm0                          \n\t"\
1254                 "pminub %%mm0, %%mm7                            \n\t"\
1255                 "pmaxub %%mm0, %%mm6                            \n\t"
1256 #else
1257 #define FIND_MIN_MAX(addr)\
1258                 "movq " #addr ", %%mm0                          \n\t"\
1259                 "movq %%mm7, %%mm1                              \n\t"\
1260                 "psubusb %%mm0, %%mm6                           \n\t"\
1261                 "paddb %%mm0, %%mm6                             \n\t"\
1262                 "psubusb %%mm0, %%mm1                           \n\t"\
1263                 "psubb %%mm1, %%mm7                             \n\t"
1264 #endif
1265
1266 FIND_MIN_MAX((%%eax))
1267 FIND_MIN_MAX((%%eax, %1))
1268 FIND_MIN_MAX((%%eax, %1, 2))
1269 FIND_MIN_MAX((%0, %1, 4))
1270 FIND_MIN_MAX((%%ebx))
1271 FIND_MIN_MAX((%%ebx, %1))
1272 FIND_MIN_MAX((%%ebx, %1, 2))
1273 FIND_MIN_MAX((%0, %1, 8))
1274
1275                 "movq %%mm7, %%mm4                              \n\t"
1276                 "psrlq $8, %%mm7                                \n\t"
1277 #ifdef HAVE_MMX2
1278                 "pminub %%mm4, %%mm7                            \n\t" // min of pixels
1279                 "pshufw $0xF9, %%mm7, %%mm4                     \n\t"
1280                 "pminub %%mm4, %%mm7                            \n\t" // min of pixels
1281                 "pshufw $0xFE, %%mm7, %%mm4                     \n\t"
1282                 "pminub %%mm4, %%mm7                            \n\t"
1283 #else
1284                 "movq %%mm7, %%mm1                              \n\t"
1285                 "psubusb %%mm4, %%mm1                           \n\t"
1286                 "psubb %%mm1, %%mm7                             \n\t"
1287                 "movq %%mm7, %%mm4                              \n\t"
1288                 "psrlq $16, %%mm7                               \n\t"
1289                 "movq %%mm7, %%mm1                              \n\t"
1290                 "psubusb %%mm4, %%mm1                           \n\t"
1291                 "psubb %%mm1, %%mm7                             \n\t"
1292                 "movq %%mm7, %%mm4                              \n\t"
1293                 "psrlq $32, %%mm7                               \n\t"
1294                 "movq %%mm7, %%mm1                              \n\t"
1295                 "psubusb %%mm4, %%mm1                           \n\t"
1296                 "psubb %%mm1, %%mm7                             \n\t"
1297 #endif
1298
1299
1300                 "movq %%mm6, %%mm4                              \n\t"
1301                 "psrlq $8, %%mm6                                \n\t"
1302 #ifdef HAVE_MMX2
1303                 "pmaxub %%mm4, %%mm6                            \n\t" // max of pixels
1304                 "pshufw $0xF9, %%mm6, %%mm4                     \n\t"
1305                 "pmaxub %%mm4, %%mm6                            \n\t"
1306                 "pshufw $0xFE, %%mm6, %%mm4                     \n\t"
1307                 "pmaxub %%mm4, %%mm6                            \n\t"
1308 #else
1309                 "psubusb %%mm4, %%mm6                           \n\t"
1310                 "paddb %%mm4, %%mm6                             \n\t"
1311                 "movq %%mm6, %%mm4                              \n\t"
1312                 "psrlq $16, %%mm6                               \n\t"
1313                 "psubusb %%mm4, %%mm6                           \n\t"
1314                 "paddb %%mm4, %%mm6                             \n\t"
1315                 "movq %%mm6, %%mm4                              \n\t"
1316                 "psrlq $32, %%mm6                               \n\t"
1317                 "psubusb %%mm4, %%mm6                           \n\t"
1318                 "paddb %%mm4, %%mm6                             \n\t"
1319 #endif
1320                 "movq %%mm6, %%mm0                              \n\t" // max
1321                 "psubb %%mm7, %%mm6                             \n\t" // max - min
1322                 "movd %%mm6, %%ecx                              \n\t"
1323                 "cmpb "MANGLE(deringThreshold)", %%cl           \n\t"
1324                 " jb 1f                                         \n\t"
1325                 PAVGB(%%mm0, %%mm7)                                   // a=(max + min)/2
1326                 "punpcklbw %%mm7, %%mm7                         \n\t"
1327                 "punpcklbw %%mm7, %%mm7                         \n\t"
1328                 "punpcklbw %%mm7, %%mm7                         \n\t"
1329                 "movq %%mm7, "MANGLE(temp0)"                    \n\t"
1330
1331                 "movq (%0), %%mm0                               \n\t" // L10
1332                 "movq %%mm0, %%mm1                              \n\t" // L10
1333                 "movq %%mm0, %%mm2                              \n\t" // L10
1334                 "psllq $8, %%mm1                                \n\t"
1335                 "psrlq $8, %%mm2                                \n\t"
1336                 "movd -4(%0), %%mm3                             \n\t"
1337                 "movd 8(%0), %%mm4                              \n\t"
1338                 "psrlq $24, %%mm3                               \n\t"
1339                 "psllq $56, %%mm4                               \n\t"
1340                 "por %%mm3, %%mm1                               \n\t" // L00
1341                 "por %%mm4, %%mm2                               \n\t" // L20
1342                 "movq %%mm1, %%mm3                              \n\t" // L00
1343                 PAVGB(%%mm2, %%mm1)                                   // (L20 + L00)/2
1344                 PAVGB(%%mm0, %%mm1)                                   // (L20 + L00 + 2L10)/4
1345                 "psubusb %%mm7, %%mm0                           \n\t"
1346                 "psubusb %%mm7, %%mm2                           \n\t"
1347                 "psubusb %%mm7, %%mm3                           \n\t"
1348                 "pcmpeqb "MANGLE(b00)", %%mm0                   \n\t" // L10 > a ? 0 : -1
1349                 "pcmpeqb "MANGLE(b00)", %%mm2                   \n\t" // L20 > a ? 0 : -1
1350                 "pcmpeqb "MANGLE(b00)", %%mm3                   \n\t" // L00 > a ? 0 : -1
1351                 "paddb %%mm2, %%mm0                             \n\t"
1352                 "paddb %%mm3, %%mm0                             \n\t"
1353
1354                 "movq (%%eax), %%mm2                            \n\t" // L11
1355                 "movq %%mm2, %%mm3                              \n\t" // L11
1356                 "movq %%mm2, %%mm4                              \n\t" // L11
1357                 "psllq $8, %%mm3                                \n\t"
1358                 "psrlq $8, %%mm4                                \n\t"
1359                 "movd -4(%%eax), %%mm5                          \n\t"
1360                 "movd 8(%%eax), %%mm6                           \n\t"
1361                 "psrlq $24, %%mm5                               \n\t"
1362                 "psllq $56, %%mm6                               \n\t"
1363                 "por %%mm5, %%mm3                               \n\t" // L01
1364                 "por %%mm6, %%mm4                               \n\t" // L21
1365                 "movq %%mm3, %%mm5                              \n\t" // L01
1366                 PAVGB(%%mm4, %%mm3)                                   // (L21 + L01)/2
1367                 PAVGB(%%mm2, %%mm3)                                   // (L21 + L01 + 2L11)/4
1368                 "psubusb %%mm7, %%mm2                           \n\t"
1369                 "psubusb %%mm7, %%mm4                           \n\t"
1370                 "psubusb %%mm7, %%mm5                           \n\t"
1371                 "pcmpeqb "MANGLE(b00)", %%mm2                   \n\t" // L11 > a ? 0 : -1
1372                 "pcmpeqb "MANGLE(b00)", %%mm4                   \n\t" // L21 > a ? 0 : -1
1373                 "pcmpeqb "MANGLE(b00)", %%mm5                   \n\t" // L01 > a ? 0 : -1
1374                 "paddb %%mm4, %%mm2                             \n\t"
1375                 "paddb %%mm5, %%mm2                             \n\t"
1376 // 0, 2, 3, 1
1377 #define DERING_CORE(dst,src,ppsx,psx,sx,pplx,plx,lx,t0,t1) \
1378                 "movq " #src ", " #sx "                         \n\t" /* src[0] */\
1379                 "movq " #sx ", " #lx "                          \n\t" /* src[0] */\
1380                 "movq " #sx ", " #t0 "                          \n\t" /* src[0] */\
1381                 "psllq $8, " #lx "                              \n\t"\
1382                 "psrlq $8, " #t0 "                              \n\t"\
1383                 "movd -4" #src ", " #t1 "                       \n\t"\
1384                 "psrlq $24, " #t1 "                             \n\t"\
1385                 "por " #t1 ", " #lx "                           \n\t" /* src[-1] */\
1386                 "movd 8" #src ", " #t1 "                        \n\t"\
1387                 "psllq $56, " #t1 "                             \n\t"\
1388                 "por " #t1 ", " #t0 "                           \n\t" /* src[+1] */\
1389                 "movq " #lx ", " #t1 "                          \n\t" /* src[-1] */\
1390                 PAVGB(t0, lx)                                         /* (src[-1] + src[+1])/2 */\
1391                 PAVGB(sx, lx)                                 /* (src[-1] + 2src[0] + src[+1])/4 */\
1392                 PAVGB(lx, pplx)                                      \
1393                 "movq " #lx ", "MANGLE(temp1)"                  \n\t"\
1394                 "movq "MANGLE(temp0)", " #lx "                  \n\t"\
1395                 "psubusb " #lx ", " #t1 "                       \n\t"\
1396                 "psubusb " #lx ", " #t0 "                       \n\t"\
1397                 "psubusb " #lx ", " #sx "                       \n\t"\
1398                 "movq "MANGLE(b00)", " #lx "                    \n\t"\
1399                 "pcmpeqb " #lx ", " #t1 "                       \n\t" /* src[-1] > a ? 0 : -1*/\
1400                 "pcmpeqb " #lx ", " #t0 "                       \n\t" /* src[+1] > a ? 0 : -1*/\
1401                 "pcmpeqb " #lx ", " #sx "                       \n\t" /* src[0]  > a ? 0 : -1*/\
1402                 "paddb " #t1 ", " #t0 "                         \n\t"\
1403                 "paddb " #t0 ", " #sx "                         \n\t"\
1404 \
1405                 PAVGB(plx, pplx)                                      /* filtered */\
1406                 "movq " #dst ", " #t0 "                         \n\t" /* dst */\
1407                 "movq " #t0 ", " #t1 "                          \n\t" /* dst */\
1408                 "psubusb "MANGLE(pQPb2)", " #t0 "               \n\t"\
1409                 "paddusb "MANGLE(pQPb2)", " #t1 "               \n\t"\
1410                 PMAXUB(t0, pplx)\
1411                 PMINUB(t1, pplx, t0)\
1412                 "paddb " #sx ", " #ppsx "                       \n\t"\
1413                 "paddb " #psx ", " #ppsx "                      \n\t"\
1414                 "#paddb "MANGLE(b02)", " #ppsx "                \n\t"\
1415                 "pand "MANGLE(b08)", " #ppsx "                  \n\t"\
1416                 "pcmpeqb " #lx ", " #ppsx "                     \n\t"\
1417                 "pand " #ppsx ", " #pplx "                      \n\t"\
1418                 "pandn " #dst ", " #ppsx "                      \n\t"\
1419                 "por " #pplx ", " #ppsx "                       \n\t"\
1420                 "movq " #ppsx ", " #dst "                       \n\t"\
1421                 "movq "MANGLE(temp1)", " #lx "                  \n\t"
1422
1423 /*
1424 0000000
1425 1111111
1426
1427 1111110
1428 1111101
1429 1111100
1430 1111011
1431 1111010
1432 1111001
1433
1434 1111000
1435 1110111
1436
1437 */
1438 //DERING_CORE(dst,src                  ,ppsx ,psx  ,sx   ,pplx ,plx  ,lx   ,t0   ,t1)
1439 DERING_CORE((%%eax),(%%eax, %1)        ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1440 DERING_CORE((%%eax, %1),(%%eax, %1, 2) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1441 DERING_CORE((%%eax, %1, 2),(%0, %1, 4) ,%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1442 DERING_CORE((%0, %1, 4),(%%ebx)        ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1443 DERING_CORE((%%ebx),(%%ebx, %1)        ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1444 DERING_CORE((%%ebx, %1), (%%ebx, %1, 2),%%mm4,%%mm0,%%mm2,%%mm5,%%mm1,%%mm3,%%mm6,%%mm7)
1445 DERING_CORE((%%ebx, %1, 2),(%0, %1, 8) ,%%mm0,%%mm2,%%mm4,%%mm1,%%mm3,%%mm5,%%mm6,%%mm7)
1446 DERING_CORE((%0, %1, 8),(%%ebx, %1, 4) ,%%mm2,%%mm4,%%mm0,%%mm3,%%mm5,%%mm1,%%mm6,%%mm7)
1447
1448                 "1:                     \n\t"
1449                 : : "r" (src), "r" (stride), "r" (QP)
1450                 : "%eax", "%ebx", "%ecx"
1451         );
1452 #else
1453         int y;
1454         int min=255;
1455         int max=0;
1456         int avg;
1457         uint8_t *p;
1458         int s[10];
1459
1460         for(y=1; y<9; y++)
1461         {
1462                 int x;
1463                 p= src + stride*y;
1464                 for(x=1; x<9; x++)
1465                 {
1466                         p++;
1467                         if(*p > max) max= *p;
1468                         if(*p < min) min= *p;
1469                 }
1470         }
1471         avg= (min + max + 1)/2;
1472
1473         if(max - min <deringThreshold) return;
1474
1475         for(y=0; y<10; y++)
1476         {
1477                 int x;
1478                 int t = 0;
1479                 p= src + stride*y;
1480                 for(x=0; x<10; x++)
1481                 {
1482                         if(*p > avg) t |= (1<<x);
1483                         p++;
1484                 }
1485                 t |= (~t)<<16;
1486                 t &= (t<<1) & (t>>1);
1487                 s[y] = t;
1488         }
1489
1490         for(y=1; y<9; y++)
1491         {
1492                 int x;
1493                 int t = s[y-1] & s[y] & s[y+1];
1494                 t|= t>>16;
1495
1496                 p= src + stride*y;
1497                 for(x=1; x<9; x++)
1498                 {
1499                         p++;
1500                         if(t & (1<<x))
1501                         {
1502                                 int f= (*(p-stride-1)) + 2*(*(p-stride)) + (*(p-stride+1))
1503                                       +2*(*(p     -1)) + 4*(*p         ) + 2*(*(p     +1))
1504                                       +(*(p+stride-1)) + 2*(*(p+stride)) + (*(p+stride+1));
1505                                 f= (f + 8)>>4;
1506
1507 #ifdef DEBUG_DERING_THRESHOLD
1508                                 asm volatile("emms\n\t":);
1509                                 {
1510                                 static long long numPixels=0;
1511                                 if(x!=1 && x!=8 && y!=1 && y!=8) numPixels++;
1512 //                              if((max-min)<20 || (max-min)*QP<200)
1513 //                              if((max-min)*QP < 500)
1514 //                              if(max-min<QP/2)
1515                                 if(max-min < 20)
1516                                 {
1517                                         static int numSkiped=0;
1518                                         static int errorSum=0;
1519                                         static int worstQP=0;
1520                                         static int worstRange=0;
1521                                         static int worstDiff=0;
1522                                         int diff= (f - *p);
1523                                         int absDiff= ABS(diff);
1524                                         int error= diff*diff;
1525
1526                                         if(x==1 || x==8 || y==1 || y==8) continue;
1527
1528                                         numSkiped++;
1529                                         if(absDiff > worstDiff)
1530                                         {
1531                                                 worstDiff= absDiff;
1532                                                 worstQP= QP;
1533                                                 worstRange= max-min;
1534                                         }
1535                                         errorSum+= error;
1536
1537                                         if(1024LL*1024LL*1024LL % numSkiped == 0)
1538                                         {
1539                                                 printf( "sum:%1.3f, skip:%d, wQP:%d, "
1540                                                         "wRange:%d, wDiff:%d, relSkip:%1.3f\n",
1541                                                         (float)errorSum/numSkiped, numSkiped, worstQP, worstRange,
1542                                                         worstDiff, (float)numSkiped/numPixels);
1543                                         }
1544                                 }
1545                                 }
1546 #endif
1547                                 if     (*p + 2*QP < f) *p= *p + 2*QP;
1548                                 else if(*p - 2*QP > f) *p= *p - 2*QP;
1549                                 else *p=f;
1550                         }
1551                 }
1552         }
1553 #ifdef DEBUG_DERING_THRESHOLD
1554         if(max-min < 20)
1555         {
1556                 for(y=1; y<9; y++)
1557                 {
1558                         int x;
1559                         int t = 0;
1560                         p= src + stride*y;
1561                         for(x=1; x<9; x++)
1562                         {
1563                                 p++;
1564                                 *p = MIN(*p + 20, 255);
1565                         }
1566                 }
1567 //              src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255;
1568         }
1569 #endif
1570 #endif
1571 }
1572
1573 /**
1574  * Deinterlaces the given block
1575  * will be called for every 8x8 block and can read & write from line 4-15
1576  * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1577  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1578  */
1579 static inline void RENAME(deInterlaceInterpolateLinear)(uint8_t src[], int stride)
1580 {
1581 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1582         src+= 4*stride;
1583         asm volatile(
1584                 "leal (%0, %1), %%eax                           \n\t"
1585                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1586 //      0       1       2       3       4       5       6       7       8       9
1587 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1588
1589                 "movq (%0), %%mm0                               \n\t"
1590                 "movq (%%eax, %1), %%mm1                        \n\t"
1591                 PAVGB(%%mm1, %%mm0)
1592                 "movq %%mm0, (%%eax)                            \n\t"
1593                 "movq (%0, %1, 4), %%mm0                        \n\t"
1594                 PAVGB(%%mm0, %%mm1)
1595                 "movq %%mm1, (%%eax, %1, 2)                     \n\t"
1596                 "movq (%%ebx, %1), %%mm1                        \n\t"
1597                 PAVGB(%%mm1, %%mm0)
1598                 "movq %%mm0, (%%ebx)                            \n\t"
1599                 "movq (%0, %1, 8), %%mm0                        \n\t"
1600                 PAVGB(%%mm0, %%mm1)
1601                 "movq %%mm1, (%%ebx, %1, 2)                     \n\t"
1602
1603                 : : "r" (src), "r" (stride)
1604                 : "%eax", "%ebx"
1605         );
1606 #else
1607         int x;
1608         src+= 4*stride;
1609         for(x=0; x<8; x++)
1610         {
1611                 src[stride]   = (src[0]        + src[stride*2])>>1;
1612                 src[stride*3] = (src[stride*2] + src[stride*4])>>1;
1613                 src[stride*5] = (src[stride*4] + src[stride*6])>>1;
1614                 src[stride*7] = (src[stride*6] + src[stride*8])>>1;
1615                 src++;
1616         }
1617 #endif
1618 }
1619
1620 /**
1621  * Deinterlaces the given block
1622  * will be called for every 8x8 block and can read & write from line 4-15
1623  * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1624  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1625  * this filter will read lines 3-15 and write 7-13
1626  * no cliping in C version
1627  */
1628 static inline void RENAME(deInterlaceInterpolateCubic)(uint8_t src[], int stride)
1629 {
1630 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1631         src+= stride*3;
1632         asm volatile(
1633                 "leal (%0, %1), %%eax                           \n\t"
1634                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1635                 "leal (%%ebx, %1, 4), %%ecx                     \n\t"
1636                 "addl %1, %%ecx                                 \n\t"
1637                 "pxor %%mm7, %%mm7                              \n\t"
1638 //      0       1       2       3       4       5       6       7       8       9       10
1639 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1 ecx
1640
1641 #define DEINT_CUBIC(a,b,c,d,e)\
1642                 "movq " #a ", %%mm0                             \n\t"\
1643                 "movq " #b ", %%mm1                             \n\t"\
1644                 "movq " #d ", %%mm2                             \n\t"\
1645                 "movq " #e ", %%mm3                             \n\t"\
1646                 PAVGB(%%mm2, %%mm1)                                     /* (b+d) /2 */\
1647                 PAVGB(%%mm3, %%mm0)                                     /* a(a+e) /2 */\
1648                 "movq %%mm0, %%mm2                              \n\t"\
1649                 "punpcklbw %%mm7, %%mm0                         \n\t"\
1650                 "punpckhbw %%mm7, %%mm2                         \n\t"\
1651                 "movq %%mm1, %%mm3                              \n\t"\
1652                 "punpcklbw %%mm7, %%mm1                         \n\t"\
1653                 "punpckhbw %%mm7, %%mm3                         \n\t"\
1654                 "psubw %%mm1, %%mm0                             \n\t"   /* L(a+e - (b+d))/2 */\
1655                 "psubw %%mm3, %%mm2                             \n\t"   /* H(a+e - (b+d))/2 */\
1656                 "psraw $3, %%mm0                                \n\t"   /* L(a+e - (b+d))/16 */\
1657                 "psraw $3, %%mm2                                \n\t"   /* H(a+e - (b+d))/16 */\
1658                 "psubw %%mm0, %%mm1                             \n\t"   /* L(9b + 9d - a - e)/16 */\
1659                 "psubw %%mm2, %%mm3                             \n\t"   /* H(9b + 9d - a - e)/16 */\
1660                 "packuswb %%mm3, %%mm1                          \n\t"\
1661                 "movq %%mm1, " #c "                             \n\t"
1662
1663 DEINT_CUBIC((%0), (%%eax, %1), (%%eax, %1, 2), (%0, %1, 4), (%%ebx, %1))
1664 DEINT_CUBIC((%%eax, %1), (%0, %1, 4), (%%ebx), (%%ebx, %1), (%0, %1, 8))
1665 DEINT_CUBIC((%0, %1, 4), (%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8), (%%ecx))
1666 DEINT_CUBIC((%%ebx, %1), (%0, %1, 8), (%%ebx, %1, 4), (%%ecx), (%%ecx, %1, 2))
1667
1668                 : : "r" (src), "r" (stride)
1669                 : "%eax", "%ebx", "ecx"
1670         );
1671 #else
1672         int x;
1673         src+= stride*3;
1674         for(x=0; x<8; x++)
1675         {
1676                 src[stride*3] = (-src[0]        + 9*src[stride*2] + 9*src[stride*4] - src[stride*6])>>4;
1677                 src[stride*5] = (-src[stride*2] + 9*src[stride*4] + 9*src[stride*6] - src[stride*8])>>4;
1678                 src[stride*7] = (-src[stride*4] + 9*src[stride*6] + 9*src[stride*8] - src[stride*10])>>4;
1679                 src[stride*9] = (-src[stride*6] + 9*src[stride*8] + 9*src[stride*10] - src[stride*12])>>4;
1680                 src++;
1681         }
1682 #endif
1683 }
1684
1685 /**
1686  * Deinterlaces the given block
1687  * will be called for every 8x8 block and can read & write from line 4-15
1688  * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1689  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1690  * will shift the image up by 1 line (FIXME if this is a problem)
1691  * this filter will read lines 4-13 and write 4-11
1692  */
1693 static inline void RENAME(deInterlaceBlendLinear)(uint8_t src[], int stride)
1694 {
1695 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
1696         src+= 4*stride;
1697         asm volatile(
1698                 "leal (%0, %1), %%eax                           \n\t"
1699                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1700 //      0       1       2       3       4       5       6       7       8       9
1701 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1702
1703                 "movq (%0), %%mm0                               \n\t" // L0
1704                 "movq (%%eax, %1), %%mm1                        \n\t" // L2
1705                 PAVGB(%%mm1, %%mm0)                                   // L0+L2
1706                 "movq (%%eax), %%mm2                            \n\t" // L1
1707                 PAVGB(%%mm2, %%mm0)
1708                 "movq %%mm0, (%0)                               \n\t"
1709                 "movq (%%eax, %1, 2), %%mm0                     \n\t" // L3
1710                 PAVGB(%%mm0, %%mm2)                                   // L1+L3
1711                 PAVGB(%%mm1, %%mm2)                                   // 2L2 + L1 + L3
1712                 "movq %%mm2, (%%eax)                            \n\t"
1713                 "movq (%0, %1, 4), %%mm2                        \n\t" // L4
1714                 PAVGB(%%mm2, %%mm1)                                   // L2+L4
1715                 PAVGB(%%mm0, %%mm1)                                   // 2L3 + L2 + L4
1716                 "movq %%mm1, (%%eax, %1)                        \n\t"
1717                 "movq (%%ebx), %%mm1                            \n\t" // L5
1718                 PAVGB(%%mm1, %%mm0)                                   // L3+L5
1719                 PAVGB(%%mm2, %%mm0)                                   // 2L4 + L3 + L5
1720                 "movq %%mm0, (%%eax, %1, 2)                     \n\t"
1721                 "movq (%%ebx, %1), %%mm0                        \n\t" // L6
1722                 PAVGB(%%mm0, %%mm2)                                   // L4+L6
1723                 PAVGB(%%mm1, %%mm2)                                   // 2L5 + L4 + L6
1724                 "movq %%mm2, (%0, %1, 4)                        \n\t"
1725                 "movq (%%ebx, %1, 2), %%mm2                     \n\t" // L7
1726                 PAVGB(%%mm2, %%mm1)                                   // L5+L7
1727                 PAVGB(%%mm0, %%mm1)                                   // 2L6 + L5 + L7
1728                 "movq %%mm1, (%%ebx)                            \n\t"
1729                 "movq (%0, %1, 8), %%mm1                        \n\t" // L8
1730                 PAVGB(%%mm1, %%mm0)                                   // L6+L8
1731                 PAVGB(%%mm2, %%mm0)                                   // 2L7 + L6 + L8
1732                 "movq %%mm0, (%%ebx, %1)                        \n\t"
1733                 "movq (%%ebx, %1, 4), %%mm0                     \n\t" // L9
1734                 PAVGB(%%mm0, %%mm2)                                   // L7+L9
1735                 PAVGB(%%mm1, %%mm2)                                   // 2L8 + L7 + L9
1736                 "movq %%mm2, (%%ebx, %1, 2)                     \n\t"
1737
1738
1739                 : : "r" (src), "r" (stride)
1740                 : "%eax", "%ebx"
1741         );
1742 #else
1743         int x;
1744         src+= 4*stride;
1745         for(x=0; x<8; x++)
1746         {
1747                 src[0       ] = (src[0       ] + 2*src[stride  ] + src[stride*2])>>2;
1748                 src[stride  ] = (src[stride  ] + 2*src[stride*2] + src[stride*3])>>2;
1749                 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2;
1750                 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2;
1751                 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2;
1752                 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2;
1753                 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2;
1754                 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2;
1755                 src++;
1756         }
1757 #endif
1758 }
1759
1760 /**
1761  * Deinterlaces the given block
1762  * will be called for every 8x8 block and can read & write from line 4-15,
1763  * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too
1764  * lines 4-12 will be read into the deblocking filter and should be deinterlaced
1765  */
1766 static inline void RENAME(deInterlaceMedian)(uint8_t src[], int stride)
1767 {
1768 #ifdef HAVE_MMX
1769         src+= 4*stride;
1770 #ifdef HAVE_MMX2
1771         asm volatile(
1772                 "leal (%0, %1), %%eax                           \n\t"
1773                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1774 //      0       1       2       3       4       5       6       7       8       9
1775 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1776
1777                 "movq (%0), %%mm0                               \n\t" //
1778                 "movq (%%eax, %1), %%mm2                        \n\t" //
1779                 "movq (%%eax), %%mm1                            \n\t" //
1780                 "movq %%mm0, %%mm3                              \n\t"
1781                 "pmaxub %%mm1, %%mm0                            \n\t" //
1782                 "pminub %%mm3, %%mm1                            \n\t" //
1783                 "pmaxub %%mm2, %%mm1                            \n\t" //
1784                 "pminub %%mm1, %%mm0                            \n\t"
1785                 "movq %%mm0, (%%eax)                            \n\t"
1786
1787                 "movq (%0, %1, 4), %%mm0                        \n\t" //
1788                 "movq (%%eax, %1, 2), %%mm1                     \n\t" //
1789                 "movq %%mm2, %%mm3                              \n\t"
1790                 "pmaxub %%mm1, %%mm2                            \n\t" //
1791                 "pminub %%mm3, %%mm1                            \n\t" //
1792                 "pmaxub %%mm0, %%mm1                            \n\t" //
1793                 "pminub %%mm1, %%mm2                            \n\t"
1794                 "movq %%mm2, (%%eax, %1, 2)                     \n\t"
1795
1796                 "movq (%%ebx), %%mm2                            \n\t" //
1797                 "movq (%%ebx, %1), %%mm1                        \n\t" //
1798                 "movq %%mm2, %%mm3                              \n\t"
1799                 "pmaxub %%mm0, %%mm2                            \n\t" //
1800                 "pminub %%mm3, %%mm0                            \n\t" //
1801                 "pmaxub %%mm1, %%mm0                            \n\t" //
1802                 "pminub %%mm0, %%mm2                            \n\t"
1803                 "movq %%mm2, (%%ebx)                            \n\t"
1804
1805                 "movq (%%ebx, %1, 2), %%mm2                     \n\t" //
1806                 "movq (%0, %1, 8), %%mm0                        \n\t" //
1807                 "movq %%mm2, %%mm3                              \n\t"
1808                 "pmaxub %%mm0, %%mm2                            \n\t" //
1809                 "pminub %%mm3, %%mm0                            \n\t" //
1810                 "pmaxub %%mm1, %%mm0                            \n\t" //
1811                 "pminub %%mm0, %%mm2                            \n\t"
1812                 "movq %%mm2, (%%ebx, %1, 2)                     \n\t"
1813
1814
1815                 : : "r" (src), "r" (stride)
1816                 : "%eax", "%ebx"
1817         );
1818
1819 #else // MMX without MMX2
1820         asm volatile(
1821                 "leal (%0, %1), %%eax                           \n\t"
1822                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1823 //      0       1       2       3       4       5       6       7       8       9
1824 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1825                 "pxor %%mm7, %%mm7                              \n\t"
1826
1827 #define MEDIAN(a,b,c)\
1828                 "movq " #a ", %%mm0                             \n\t"\
1829                 "movq " #b ", %%mm2                             \n\t"\
1830                 "movq " #c ", %%mm1                             \n\t"\
1831                 "movq %%mm0, %%mm3                              \n\t"\
1832                 "movq %%mm1, %%mm4                              \n\t"\
1833                 "movq %%mm2, %%mm5                              \n\t"\
1834                 "psubusb %%mm1, %%mm3                           \n\t"\
1835                 "psubusb %%mm2, %%mm4                           \n\t"\
1836                 "psubusb %%mm0, %%mm5                           \n\t"\
1837                 "pcmpeqb %%mm7, %%mm3                           \n\t"\
1838                 "pcmpeqb %%mm7, %%mm4                           \n\t"\
1839                 "pcmpeqb %%mm7, %%mm5                           \n\t"\
1840                 "movq %%mm3, %%mm6                              \n\t"\
1841                 "pxor %%mm4, %%mm3                              \n\t"\
1842                 "pxor %%mm5, %%mm4                              \n\t"\
1843                 "pxor %%mm6, %%mm5                              \n\t"\
1844                 "por %%mm3, %%mm1                               \n\t"\
1845                 "por %%mm4, %%mm2                               \n\t"\
1846                 "por %%mm5, %%mm0                               \n\t"\
1847                 "pand %%mm2, %%mm0                              \n\t"\
1848                 "pand %%mm1, %%mm0                              \n\t"\
1849                 "movq %%mm0, " #b "                             \n\t"
1850
1851 MEDIAN((%0), (%%eax), (%%eax, %1))
1852 MEDIAN((%%eax, %1), (%%eax, %1, 2), (%0, %1, 4))
1853 MEDIAN((%0, %1, 4), (%%ebx), (%%ebx, %1))
1854 MEDIAN((%%ebx, %1), (%%ebx, %1, 2), (%0, %1, 8))
1855
1856                 : : "r" (src), "r" (stride)
1857                 : "%eax", "%ebx"
1858         );
1859 #endif // MMX
1860 #else
1861         //FIXME
1862         int x;
1863         src+= 4*stride;
1864         for(x=0; x<8; x++)
1865         {
1866                 src[0       ] = (src[0       ] + 2*src[stride  ] + src[stride*2])>>2;
1867                 src[stride  ] = (src[stride  ] + 2*src[stride*2] + src[stride*3])>>2;
1868                 src[stride*2] = (src[stride*2] + 2*src[stride*3] + src[stride*4])>>2;
1869                 src[stride*3] = (src[stride*3] + 2*src[stride*4] + src[stride*5])>>2;
1870                 src[stride*4] = (src[stride*4] + 2*src[stride*5] + src[stride*6])>>2;
1871                 src[stride*5] = (src[stride*5] + 2*src[stride*6] + src[stride*7])>>2;
1872                 src[stride*6] = (src[stride*6] + 2*src[stride*7] + src[stride*8])>>2;
1873                 src[stride*7] = (src[stride*7] + 2*src[stride*8] + src[stride*9])>>2;
1874                 src++;
1875         }
1876 #endif
1877 }
1878
1879 #ifdef HAVE_MMX
1880 /**
1881  * transposes and shift the given 8x8 Block into dst1 and dst2
1882  */
1883 static inline void RENAME(transpose1)(uint8_t *dst1, uint8_t *dst2, uint8_t *src, int srcStride)
1884 {
1885         asm(
1886                 "leal (%0, %1), %%eax                           \n\t"
1887                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1888 //      0       1       2       3       4       5       6       7       8       9
1889 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1890                 "movq (%0), %%mm0               \n\t" // 12345678
1891                 "movq (%%eax), %%mm1            \n\t" // abcdefgh
1892                 "movq %%mm0, %%mm2              \n\t" // 12345678
1893                 "punpcklbw %%mm1, %%mm0         \n\t" // 1a2b3c4d
1894                 "punpckhbw %%mm1, %%mm2         \n\t" // 5e6f7g8h
1895
1896                 "movq (%%eax, %1), %%mm1        \n\t"
1897                 "movq (%%eax, %1, 2), %%mm3     \n\t"
1898                 "movq %%mm1, %%mm4              \n\t"
1899                 "punpcklbw %%mm3, %%mm1         \n\t"
1900                 "punpckhbw %%mm3, %%mm4         \n\t"
1901
1902                 "movq %%mm0, %%mm3              \n\t"
1903                 "punpcklwd %%mm1, %%mm0         \n\t"
1904                 "punpckhwd %%mm1, %%mm3         \n\t"
1905                 "movq %%mm2, %%mm1              \n\t"
1906                 "punpcklwd %%mm4, %%mm2         \n\t"
1907                 "punpckhwd %%mm4, %%mm1         \n\t"
1908
1909                 "movd %%mm0, 128(%2)            \n\t"
1910                 "psrlq $32, %%mm0               \n\t"
1911                 "movd %%mm0, 144(%2)            \n\t"
1912                 "movd %%mm3, 160(%2)            \n\t"
1913                 "psrlq $32, %%mm3               \n\t"
1914                 "movd %%mm3, 176(%2)            \n\t"
1915                 "movd %%mm3, 48(%3)             \n\t"
1916                 "movd %%mm2, 192(%2)            \n\t"
1917                 "movd %%mm2, 64(%3)             \n\t"
1918                 "psrlq $32, %%mm2               \n\t"
1919                 "movd %%mm2, 80(%3)             \n\t"
1920                 "movd %%mm1, 96(%3)             \n\t"
1921                 "psrlq $32, %%mm1               \n\t"
1922                 "movd %%mm1, 112(%3)            \n\t"
1923
1924                 "movq (%0, %1, 4), %%mm0        \n\t" // 12345678
1925                 "movq (%%ebx), %%mm1            \n\t" // abcdefgh
1926                 "movq %%mm0, %%mm2              \n\t" // 12345678
1927                 "punpcklbw %%mm1, %%mm0         \n\t" // 1a2b3c4d
1928                 "punpckhbw %%mm1, %%mm2         \n\t" // 5e6f7g8h
1929
1930                 "movq (%%ebx, %1), %%mm1        \n\t"
1931                 "movq (%%ebx, %1, 2), %%mm3     \n\t"
1932                 "movq %%mm1, %%mm4              \n\t"
1933                 "punpcklbw %%mm3, %%mm1         \n\t"
1934                 "punpckhbw %%mm3, %%mm4         \n\t"
1935
1936                 "movq %%mm0, %%mm3              \n\t"
1937                 "punpcklwd %%mm1, %%mm0         \n\t"
1938                 "punpckhwd %%mm1, %%mm3         \n\t"
1939                 "movq %%mm2, %%mm1              \n\t"
1940                 "punpcklwd %%mm4, %%mm2         \n\t"
1941                 "punpckhwd %%mm4, %%mm1         \n\t"
1942
1943                 "movd %%mm0, 132(%2)            \n\t"
1944                 "psrlq $32, %%mm0               \n\t"
1945                 "movd %%mm0, 148(%2)            \n\t"
1946                 "movd %%mm3, 164(%2)            \n\t"
1947                 "psrlq $32, %%mm3               \n\t"
1948                 "movd %%mm3, 180(%2)            \n\t"
1949                 "movd %%mm3, 52(%3)             \n\t"
1950                 "movd %%mm2, 196(%2)            \n\t"
1951                 "movd %%mm2, 68(%3)             \n\t"
1952                 "psrlq $32, %%mm2               \n\t"
1953                 "movd %%mm2, 84(%3)             \n\t"
1954                 "movd %%mm1, 100(%3)            \n\t"
1955                 "psrlq $32, %%mm1               \n\t"
1956                 "movd %%mm1, 116(%3)            \n\t"
1957
1958
1959         :: "r" (src), "r" (srcStride), "r" (dst1), "r" (dst2)
1960         : "%eax", "%ebx"
1961         );
1962 }
1963
1964 /**
1965  * transposes the given 8x8 block
1966  */
1967 static inline void RENAME(transpose2)(uint8_t *dst, int dstStride, uint8_t *src)
1968 {
1969         asm(
1970                 "leal (%0, %1), %%eax                           \n\t"
1971                 "leal (%%eax, %1, 4), %%ebx                     \n\t"
1972 //      0       1       2       3       4       5       6       7       8       9
1973 //      %0      eax     eax+%1  eax+2%1 %0+4%1  ebx     ebx+%1  ebx+2%1 %0+8%1  ebx+4%1
1974                 "movq (%2), %%mm0               \n\t" // 12345678
1975                 "movq 16(%2), %%mm1             \n\t" // abcdefgh
1976                 "movq %%mm0, %%mm2              \n\t" // 12345678
1977                 "punpcklbw %%mm1, %%mm0         \n\t" // 1a2b3c4d
1978                 "punpckhbw %%mm1, %%mm2         \n\t" // 5e6f7g8h
1979
1980                 "movq 32(%2), %%mm1             \n\t"
1981                 "movq 48(%2), %%mm3             \n\t"
1982                 "movq %%mm1, %%mm4              \n\t"
1983                 "punpcklbw %%mm3, %%mm1         \n\t"
1984                 "punpckhbw %%mm3, %%mm4         \n\t"
1985
1986                 "movq %%mm0, %%mm3              \n\t"
1987                 "punpcklwd %%mm1, %%mm0         \n\t"
1988                 "punpckhwd %%mm1, %%mm3         \n\t"
1989                 "movq %%mm2, %%mm1              \n\t"
1990                 "punpcklwd %%mm4, %%mm2         \n\t"
1991                 "punpckhwd %%mm4, %%mm1         \n\t"
1992
1993                 "movd %%mm0, (%0)               \n\t"
1994                 "psrlq $32, %%mm0               \n\t"
1995                 "movd %%mm0, (%%eax)            \n\t"
1996                 "movd %%mm3, (%%eax, %1)        \n\t"
1997                 "psrlq $32, %%mm3               \n\t"
1998                 "movd %%mm3, (%%eax, %1, 2)     \n\t"
1999                 "movd %%mm2, (%0, %1, 4)        \n\t"
2000                 "psrlq $32, %%mm2               \n\t"
2001                 "movd %%mm2, (%%ebx)            \n\t"
2002                 "movd %%mm1, (%%ebx, %1)        \n\t"
2003                 "psrlq $32, %%mm1               \n\t"
2004                 "movd %%mm1, (%%ebx, %1, 2)     \n\t"
2005
2006
2007                 "movq 64(%2), %%mm0             \n\t" // 12345678
2008                 "movq 80(%2), %%mm1             \n\t" // abcdefgh
2009                 "movq %%mm0, %%mm2              \n\t" // 12345678
2010                 "punpcklbw %%mm1, %%mm0         \n\t" // 1a2b3c4d
2011                 "punpckhbw %%mm1, %%mm2         \n\t" // 5e6f7g8h
2012
2013                 "movq 96(%2), %%mm1             \n\t"
2014                 "movq 112(%2), %%mm3            \n\t"
2015                 "movq %%mm1, %%mm4              \n\t"
2016                 "punpcklbw %%mm3, %%mm1         \n\t"
2017                 "punpckhbw %%mm3, %%mm4         \n\t"
2018
2019                 "movq %%mm0, %%mm3              \n\t"
2020                 "punpcklwd %%mm1, %%mm0         \n\t"
2021                 "punpckhwd %%mm1, %%mm3         \n\t"
2022                 "movq %%mm2, %%mm1              \n\t"
2023                 "punpcklwd %%mm4, %%mm2         \n\t"
2024                 "punpckhwd %%mm4, %%mm1         \n\t"
2025
2026                 "movd %%mm0, 4(%0)              \n\t"
2027                 "psrlq $32, %%mm0               \n\t"
2028                 "movd %%mm0, 4(%%eax)           \n\t"
2029                 "movd %%mm3, 4(%%eax, %1)       \n\t"
2030                 "psrlq $32, %%mm3               \n\t"
2031                 "movd %%mm3, 4(%%eax, %1, 2)    \n\t"
2032                 "movd %%mm2, 4(%0, %1, 4)       \n\t"
2033                 "psrlq $32, %%mm2               \n\t"
2034                 "movd %%mm2, 4(%%ebx)           \n\t"
2035                 "movd %%mm1, 4(%%ebx, %1)       \n\t"
2036                 "psrlq $32, %%mm1               \n\t"
2037                 "movd %%mm1, 4(%%ebx, %1, 2)    \n\t"
2038
2039         :: "r" (dst), "r" (dstStride), "r" (src)
2040         : "%eax", "%ebx"
2041         );
2042 }
2043 #endif
2044 //static int test=0;
2045
2046 static void inline RENAME(tempNoiseReducer)(uint8_t *src, int stride,
2047                                     uint8_t *tempBlured, uint32_t *tempBluredPast, int *maxNoise)
2048 {
2049 #define FAST_L2_DIFF
2050 //#define L1_DIFF //u should change the thresholds too if u try that one
2051 #if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
2052         asm volatile(
2053                 "leal (%2, %2, 2), %%eax                        \n\t" // 3*stride
2054                 "leal (%2, %2, 4), %%ebx                        \n\t" // 5*stride
2055                 "leal (%%ebx, %2, 2), %%ecx                     \n\t" // 7*stride
2056 //      0       1       2       3       4       5       6       7       8       9
2057 //      %x      %x+%2   %x+2%2  %x+eax  %x+4%2  %x+ebx  %x+2eax %x+ecx  %x+8%2
2058 //FIXME reorder?
2059 #ifdef L1_DIFF //needs mmx2
2060                 "movq (%0), %%mm0                               \n\t" // L0
2061                 "psadbw (%1), %%mm0                             \n\t" // |L0-R0|
2062                 "movq (%0, %2), %%mm1                           \n\t" // L1
2063                 "psadbw (%1, %2), %%mm1                         \n\t" // |L1-R1|
2064                 "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2065                 "psadbw (%1, %2, 2), %%mm2                      \n\t" // |L2-R2|
2066                 "movq (%0, %%eax), %%mm3                        \n\t" // L3
2067                 "psadbw (%1, %%eax), %%mm3                      \n\t" // |L3-R3|
2068
2069                 "movq (%0, %2, 4), %%mm4                        \n\t" // L4
2070                 "paddw %%mm1, %%mm0                             \n\t"
2071                 "psadbw (%1, %2, 4), %%mm4                      \n\t" // |L4-R4|
2072                 "movq (%0, %%ebx), %%mm5                        \n\t" // L5
2073                 "paddw %%mm2, %%mm0                             \n\t"
2074                 "psadbw (%1, %%ebx), %%mm5                      \n\t" // |L5-R5|
2075                 "movq (%0, %%eax, 2), %%mm6                     \n\t" // L6
2076                 "paddw %%mm3, %%mm0                             \n\t"
2077                 "psadbw (%1, %%eax, 2), %%mm6                   \n\t" // |L6-R6|
2078                 "movq (%0, %%ecx), %%mm7                        \n\t" // L7
2079                 "paddw %%mm4, %%mm0                             \n\t"
2080                 "psadbw (%1, %%ecx), %%mm7                      \n\t" // |L7-R7|
2081                 "paddw %%mm5, %%mm6                             \n\t"
2082                 "paddw %%mm7, %%mm6                             \n\t"
2083                 "paddw %%mm6, %%mm0                             \n\t"
2084 #elif defined (FAST_L2_DIFF)
2085                 "pcmpeqb %%mm7, %%mm7                           \n\t"
2086                 "movq "MANGLE(b80)", %%mm6                      \n\t"
2087                 "pxor %%mm0, %%mm0                              \n\t"
2088 #define L2_DIFF_CORE(a, b)\
2089                 "movq " #a ", %%mm5                             \n\t"\
2090                 "movq " #b ", %%mm2                             \n\t"\
2091                 "pxor %%mm7, %%mm2                              \n\t"\
2092                 PAVGB(%%mm2, %%mm5)\
2093                 "paddb %%mm6, %%mm5                             \n\t"\
2094                 "movq %%mm5, %%mm2                              \n\t"\
2095                 "psllw $8, %%mm5                                \n\t"\
2096                 "pmaddwd %%mm5, %%mm5                           \n\t"\
2097                 "pmaddwd %%mm2, %%mm2                           \n\t"\
2098                 "paddd %%mm2, %%mm5                             \n\t"\
2099                 "psrld $14, %%mm5                               \n\t"\
2100                 "paddd %%mm5, %%mm0                             \n\t"
2101
2102 L2_DIFF_CORE((%0), (%1))
2103 L2_DIFF_CORE((%0, %2), (%1, %2))
2104 L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))
2105 L2_DIFF_CORE((%0, %%eax), (%1, %%eax))
2106 L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))
2107 L2_DIFF_CORE((%0, %%ebx), (%1, %%ebx))
2108 L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))
2109 L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))
2110
2111 #else
2112                 "pxor %%mm7, %%mm7                              \n\t"
2113                 "pxor %%mm0, %%mm0                              \n\t"
2114 #define L2_DIFF_CORE(a, b)\
2115                 "movq " #a ", %%mm5                             \n\t"\
2116                 "movq " #b ", %%mm2                             \n\t"\
2117                 "movq %%mm5, %%mm1                              \n\t"\
2118                 "movq %%mm2, %%mm3                              \n\t"\
2119                 "punpcklbw %%mm7, %%mm5                         \n\t"\
2120                 "punpckhbw %%mm7, %%mm1                         \n\t"\
2121                 "punpcklbw %%mm7, %%mm2                         \n\t"\
2122                 "punpckhbw %%mm7, %%mm3                         \n\t"\
2123                 "psubw %%mm2, %%mm5                             \n\t"\
2124                 "psubw %%mm3, %%mm1                             \n\t"\
2125                 "pmaddwd %%mm5, %%mm5                           \n\t"\
2126                 "pmaddwd %%mm1, %%mm1                           \n\t"\
2127                 "paddd %%mm1, %%mm5                             \n\t"\
2128                 "paddd %%mm5, %%mm0                             \n\t"
2129
2130 L2_DIFF_CORE((%0), (%1))
2131 L2_DIFF_CORE((%0, %2), (%1, %2))
2132 L2_DIFF_CORE((%0, %2, 2), (%1, %2, 2))
2133 L2_DIFF_CORE((%0, %%eax), (%1, %%eax))
2134 L2_DIFF_CORE((%0, %2, 4), (%1, %2, 4))
2135 L2_DIFF_CORE((%0, %%ebx), (%1, %%ebx))
2136 L2_DIFF_CORE((%0, %%eax,2), (%1, %%eax,2))
2137 L2_DIFF_CORE((%0, %%ecx), (%1, %%ecx))
2138
2139 #endif
2140
2141                 "movq %%mm0, %%mm4                              \n\t"
2142                 "psrlq $32, %%mm0                               \n\t"
2143                 "paddd %%mm0, %%mm4                             \n\t"
2144                 "movd %%mm4, %%ecx                              \n\t"
2145                 "shll $2, %%ecx                                 \n\t"
2146                 "movl %3, %%ebx                                 \n\t"
2147                 "addl -4(%%ebx), %%ecx                          \n\t"
2148                 "addl 4(%%ebx), %%ecx                           \n\t"
2149                 "addl -1024(%%ebx), %%ecx                       \n\t"
2150                 "addl $4, %%ecx                                 \n\t"
2151                 "addl 1024(%%ebx), %%ecx                        \n\t"
2152                 "shrl $3, %%ecx                                 \n\t"
2153                 "movl %%ecx, (%%ebx)                            \n\t"
2154                 "leal (%%eax, %2, 2), %%ebx                     \n\t" // 5*stride
2155
2156 //              "movl %3, %%ecx                                 \n\t"
2157 //              "movl %%ecx, test                               \n\t"
2158 //              "jmp 4f \n\t"
2159                 "cmpl 4+"MANGLE(maxTmpNoise)", %%ecx            \n\t"
2160                 " jb 2f                                         \n\t"
2161                 "cmpl 8+"MANGLE(maxTmpNoise)", %%ecx            \n\t"
2162                 " jb 1f                                         \n\t"
2163
2164                 "leal (%%ebx, %2, 2), %%ecx                     \n\t" // 7*stride
2165                 "movq (%0), %%mm0                               \n\t" // L0
2166                 "movq (%0, %2), %%mm1                           \n\t" // L1
2167                 "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2168                 "movq (%0, %%eax), %%mm3                        \n\t" // L3
2169                 "movq (%0, %2, 4), %%mm4                        \n\t" // L4
2170                 "movq (%0, %%ebx), %%mm5                        \n\t" // L5
2171                 "movq (%0, %%eax, 2), %%mm6                     \n\t" // L6
2172                 "movq (%0, %%ecx), %%mm7                        \n\t" // L7
2173                 "movq %%mm0, (%1)                               \n\t" // L0
2174                 "movq %%mm1, (%1, %2)                           \n\t" // L1
2175                 "movq %%mm2, (%1, %2, 2)                        \n\t" // L2
2176                 "movq %%mm3, (%1, %%eax)                        \n\t" // L3
2177                 "movq %%mm4, (%1, %2, 4)                        \n\t" // L4
2178                 "movq %%mm5, (%1, %%ebx)                        \n\t" // L5
2179                 "movq %%mm6, (%1, %%eax, 2)                     \n\t" // L6
2180                 "movq %%mm7, (%1, %%ecx)                        \n\t" // L7
2181                 "jmp 4f                                         \n\t"
2182
2183                 "1:                                             \n\t"
2184                 "leal (%%ebx, %2, 2), %%ecx                     \n\t" // 7*stride
2185                 "movq (%0), %%mm0                               \n\t" // L0
2186                 "pavgb (%1), %%mm0                              \n\t" // L0
2187                 "movq (%0, %2), %%mm1                           \n\t" // L1
2188                 "pavgb (%1, %2), %%mm1                          \n\t" // L1
2189                 "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2190                 "pavgb (%1, %2, 2), %%mm2                       \n\t" // L2
2191                 "movq (%0, %%eax), %%mm3                        \n\t" // L3
2192                 "pavgb (%1, %%eax), %%mm3                       \n\t" // L3
2193                 "movq (%0, %2, 4), %%mm4                        \n\t" // L4
2194                 "pavgb (%1, %2, 4), %%mm4                       \n\t" // L4
2195                 "movq (%0, %%ebx), %%mm5                        \n\t" // L5
2196                 "pavgb (%1, %%ebx), %%mm5                       \n\t" // L5
2197                 "movq (%0, %%eax, 2), %%mm6                     \n\t" // L6
2198                 "pavgb (%1, %%eax, 2), %%mm6                    \n\t" // L6
2199                 "movq (%0, %%ecx), %%mm7                        \n\t" // L7
2200                 "pavgb (%1, %%ecx), %%mm7                       \n\t" // L7
2201                 "movq %%mm0, (%1)                               \n\t" // R0
2202                 "movq %%mm1, (%1, %2)                           \n\t" // R1
2203                 "movq %%mm2, (%1, %2, 2)                        \n\t" // R2
2204                 "movq %%mm3, (%1, %%eax)                        \n\t" // R3
2205                 "movq %%mm4, (%1, %2, 4)                        \n\t" // R4
2206                 "movq %%mm5, (%1, %%ebx)                        \n\t" // R5
2207                 "movq %%mm6, (%1, %%eax, 2)                     \n\t" // R6
2208                 "movq %%mm7, (%1, %%ecx)                        \n\t" // R7
2209                 "movq %%mm0, (%0)                               \n\t" // L0
2210                 "movq %%mm1, (%0, %2)                           \n\t" // L1
2211                 "movq %%mm2, (%0, %2, 2)                        \n\t" // L2
2212                 "movq %%mm3, (%0, %%eax)                        \n\t" // L3
2213                 "movq %%mm4, (%0, %2, 4)                        \n\t" // L4
2214                 "movq %%mm5, (%0, %%ebx)                        \n\t" // L5
2215                 "movq %%mm6, (%0, %%eax, 2)                     \n\t" // L6
2216                 "movq %%mm7, (%0, %%ecx)                        \n\t" // L7
2217                 "jmp 4f                                         \n\t"
2218
2219                 "2:                                             \n\t"
2220                 "cmpl "MANGLE(maxTmpNoise)", %%ecx              \n\t"
2221                 " jb 3f                                         \n\t"
2222
2223                 "leal (%%ebx, %2, 2), %%ecx                     \n\t" // 7*stride
2224                 "movq (%0), %%mm0                               \n\t" // L0
2225                 "movq (%0, %2), %%mm1                           \n\t" // L1
2226                 "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2227                 "movq (%0, %%eax), %%mm3                        \n\t" // L3
2228                 "movq (%1), %%mm4                               \n\t" // R0
2229                 "movq (%1, %2), %%mm5                           \n\t" // R1
2230                 "movq (%1, %2, 2), %%mm6                        \n\t" // R2
2231                 "movq (%1, %%eax), %%mm7                        \n\t" // R3
2232                 PAVGB(%%mm4, %%mm0)
2233                 PAVGB(%%mm5, %%mm1)
2234                 PAVGB(%%mm6, %%mm2)
2235                 PAVGB(%%mm7, %%mm3)
2236                 PAVGB(%%mm4, %%mm0)
2237                 PAVGB(%%mm5, %%mm1)
2238                 PAVGB(%%mm6, %%mm2)
2239                 PAVGB(%%mm7, %%mm3)
2240                 "movq %%mm0, (%1)                               \n\t" // R0
2241                 "movq %%mm1, (%1, %2)                           \n\t" // R1
2242                 "movq %%mm2, (%1, %2, 2)                        \n\t" // R2
2243                 "movq %%mm3, (%1, %%eax)                        \n\t" // R3
2244                 "movq %%mm0, (%0)                               \n\t" // L0
2245                 "movq %%mm1, (%0, %2)                           \n\t" // L1
2246                 "movq %%mm2, (%0, %2, 2)                        \n\t" // L2
2247                 "movq %%mm3, (%0, %%eax)                        \n\t" // L3
2248
2249                 "movq (%0, %2, 4), %%mm0                        \n\t" // L4
2250                 "movq (%0, %%ebx), %%mm1                        \n\t" // L5
2251                 "movq (%0, %%eax, 2), %%mm2                     \n\t" // L6
2252                 "movq (%0, %%ecx), %%mm3                        \n\t" // L7
2253                 "movq (%1, %2, 4), %%mm4                        \n\t" // R4
2254                 "movq (%1, %%ebx), %%mm5                        \n\t" // R5
2255                 "movq (%1, %%eax, 2), %%mm6                     \n\t" // R6
2256                 "movq (%1, %%ecx), %%mm7                        \n\t" // R7
2257                 PAVGB(%%mm4, %%mm0)
2258                 PAVGB(%%mm5, %%mm1)
2259                 PAVGB(%%mm6, %%mm2)
2260                 PAVGB(%%mm7, %%mm3)
2261                 PAVGB(%%mm4, %%mm0)
2262                 PAVGB(%%mm5, %%mm1)
2263                 PAVGB(%%mm6, %%mm2)
2264                 PAVGB(%%mm7, %%mm3)
2265                 "movq %%mm0, (%1, %2, 4)                        \n\t" // R4
2266                 "movq %%mm1, (%1, %%ebx)                        \n\t" // R5
2267                 "movq %%mm2, (%1, %%eax, 2)                     \n\t" // R6
2268                 "movq %%mm3, (%1, %%ecx)                        \n\t" // R7
2269                 "movq %%mm0, (%0, %2, 4)                        \n\t" // L4
2270                 "movq %%mm1, (%0, %%ebx)                        \n\t" // L5
2271                 "movq %%mm2, (%0, %%eax, 2)                     \n\t" // L6
2272                 "movq %%mm3, (%0, %%ecx)                        \n\t" // L7
2273                 "jmp 4f                                         \n\t"
2274
2275                 "3:                                             \n\t"
2276                 "leal (%%ebx, %2, 2), %%ecx                     \n\t" // 7*stride
2277                 "movq (%0), %%mm0                               \n\t" // L0
2278                 "movq (%0, %2), %%mm1                           \n\t" // L1
2279                 "movq (%0, %2, 2), %%mm2                        \n\t" // L2
2280                 "movq (%0, %%eax), %%mm3                        \n\t" // L3
2281                 "movq (%1), %%mm4                               \n\t" // R0
2282                 "movq (%1, %2), %%mm5                           \n\t" // R1
2283                 "movq (%1, %2, 2), %%mm6                        \n\t" // R2
2284                 "movq (%1, %%eax), %%mm7                        \n\t" // R3
2285                 PAVGB(%%mm4, %%mm0)
2286                 PAVGB(%%mm5, %%mm1)
2287                 PAVGB(%%mm6, %%mm2)
2288                 PAVGB(%%mm7, %%mm3)
2289                 PAVGB(%%mm4, %%mm0)
2290                 PAVGB(%%mm5, %%mm1)
2291                 PAVGB(%%mm6, %%mm2)
2292                 PAVGB(%%mm7, %%mm3)
2293                 PAVGB(%%mm4, %%mm0)
2294                 PAVGB(%%mm5, %%mm1)
2295                 PAVGB(%%mm6, %%mm2)
2296                 PAVGB(%%mm7, %%mm3)
2297                 "movq %%mm0, (%1)                               \n\t" // R0
2298                 "movq %%mm1, (%1, %2)                           \n\t" // R1
2299                 "movq %%mm2, (%1, %2, 2)                        \n\t" // R2
2300                 "movq %%mm3, (%1, %%eax)                        \n\t" // R3
2301                 "movq %%mm0, (%0)                               \n\t" // L0
2302                 "movq %%mm1, (%0, %2)                           \n\t" // L1
2303                 "movq %%mm2, (%0, %2, 2)                        \n\t" // L2
2304                 "movq %%mm3, (%0, %%eax)                        \n\t" // L3
2305
2306                 "movq (%0, %2, 4), %%mm0                        \n\t" // L4
2307                 "movq (%0, %%ebx), %%mm1                        \n\t" // L5
2308                 "movq (%0, %%eax, 2), %%mm2                     \n\t" // L6
2309                 "movq (%0, %%ecx), %%mm3                        \n\t" // L7
2310                 "movq (%1, %2, 4), %%mm4                        \n\t" // R4
2311                 "movq (%1, %%ebx), %%mm5                        \n\t" // R5
2312                 "movq (%1, %%eax, 2), %%mm6                     \n\t" // R6
2313                 "movq (%1, %%ecx), %%mm7                        \n\t" // R7
2314                 PAVGB(%%mm4, %%mm0)
2315                 PAVGB(%%mm5, %%mm1)
2316                 PAVGB(%%mm6, %%mm2)
2317                 PAVGB(%%mm7, %%mm3)
2318                 PAVGB(%%mm4, %%mm0)
2319                 PAVGB(%%mm5, %%mm1)
2320                 PAVGB(%%mm6, %%mm2)
2321                 PAVGB(%%mm7, %%mm3)
2322                 PAVGB(%%mm4, %%mm0)
2323                 PAVGB(%%mm5, %%mm1)
2324                 PAVGB(%%mm6, %%mm2)
2325                 PAVGB(%%mm7, %%mm3)
2326                 "movq %%mm0, (%1, %2, 4)                        \n\t" // R4
2327                 "movq %%mm1, (%1, %%ebx)                        \n\t" // R5
2328                 "movq %%mm2, (%1, %%eax, 2)                     \n\t" // R6
2329                 "movq %%mm3, (%1, %%ecx)                        \n\t" // R7
2330                 "movq %%mm0, (%0, %2, 4)                        \n\t" // L4
2331                 "movq %%mm1, (%0, %%ebx)                        \n\t" // L5
2332                 "movq %%mm2, (%0, %%eax, 2)                     \n\t" // L6
2333                 "movq %%mm3, (%0, %%ecx)                        \n\t" // L7
2334
2335                 "4:                                             \n\t"
2336
2337                 :: "r" (src), "r" (tempBlured), "r"(stride), "m" (tempBluredPast)
2338                 : "%eax", "%ebx", "%ecx", "memory"
2339                 );
2340 //printf("%d\n", test);
2341 #else
2342         int y;
2343         int d=0;
2344         int sysd=0;
2345         int i;
2346
2347         for(y=0; y<8; y++)
2348         {
2349                 int x;
2350                 for(x=0; x<8; x++)
2351                 {
2352                         int ref= tempBlured[ x + y*stride ];
2353                         int cur= src[ x + y*stride ];
2354                         int d1=ref - cur;
2355 //                      if(x==0 || x==7) d1+= d1>>1;
2356 //                      if(y==0 || y==7) d1+= d1>>1;
2357 //                      d+= ABS(d1);
2358                         d+= d1*d1;
2359                         sysd+= d1;
2360                 }
2361         }
2362         i=d;
2363         d=      (
2364                 4*d
2365                 +(*(tempBluredPast-256))
2366                 +(*(tempBluredPast-1))+ (*(tempBluredPast+1))
2367                 +(*(tempBluredPast+256))
2368                 +4)>>3;
2369         *tempBluredPast=i;
2370 //      ((*tempBluredPast)*3 + d + 2)>>2;
2371
2372 //printf("%d %d %d\n", maxNoise[0], maxNoise[1], maxNoise[2]);
2373 /*
2374 Switch between
2375  1  0  0  0  0  0  0  (0)
2376 64 32 16  8  4  2  1  (1)
2377 64 48 36 27 20 15 11 (33) (approx)
2378 64 56 49 43 37 33 29 (200) (approx)
2379 */
2380         if(d > maxNoise[1])
2381         {
2382                 if(d < maxNoise[2])
2383                 {
2384                         for(y=0; y<8; y++)
2385                         {
2386                                 int x;
2387                                 for(x=0; x<8; x++)
2388                                 {
2389                                         int ref= tempBlured[ x + y*stride ];
2390                                         int cur= src[ x + y*stride ];
2391                                         tempBlured[ x + y*stride ]=
2392                                         src[ x + y*stride ]=
2393                                                 (ref + cur + 1)>>1;
2394                                 }
2395                         }
2396                 }
2397                 else
2398                 {
2399                         for(y=0; y<8; y++)
2400                         {
2401                                 int x;
2402                                 for(x=0; x<8; x++)
2403                                 {
2404                                         tempBlured[ x + y*stride ]= src[ x + y*stride ];
2405                                 }
2406                         }
2407                 }
2408         }
2409         else
2410         {
2411                 if(d < maxNoise[0])
2412                 {
2413                         for(y=0; y<8; y++)
2414                         {
2415                                 int x;
2416                                 for(x=0; x<8; x++)
2417                                 {
2418                                         int ref= tempBlured[ x + y*stride ];
2419                                         int cur= src[ x + y*stride ];
2420                                         tempBlured[ x + y*stride ]=
2421                                         src[ x + y*stride ]=
2422                                                 (ref*7 + cur + 4)>>3;
2423                                 }
2424                         }
2425                 }
2426                 else
2427                 {
2428                         for(y=0; y<8; y++)
2429                         {
2430                                 int x;
2431                                 for(x=0; x<8; x++)
2432                                 {
2433                                         int ref= tempBlured[ x + y*stride ];
2434                                         int cur= src[ x + y*stride ];
2435                                         tempBlured[ x + y*stride ]=
2436                                         src[ x + y*stride ]=
2437                                                 (ref*3 + cur + 2)>>2;
2438                                 }
2439                         }
2440                 }
2441         }
2442 #endif
2443 }
2444
2445 static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
2446         QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode);
2447
2448 /**
2449  * Copies a block from src to dst and fixes the blacklevel
2450  * levelFix == 0 -> dont touch the brighness & contrast
2451  */
2452 static inline void RENAME(blockCopy)(uint8_t dst[], int dstStride, uint8_t src[], int srcStride,
2453         int levelFix)
2454 {
2455 #ifndef HAVE_MMX
2456         int i;
2457 #endif
2458         if(levelFix)
2459         {
2460 #ifdef HAVE_MMX
2461                                         asm volatile(
2462                                                 "leal (%0,%2), %%eax    \n\t"
2463                                                 "leal (%1,%3), %%ebx    \n\t"
2464                                                 "movq "MANGLE(packedYOffset)", %%mm2\n\t"
2465                                                 "movq "MANGLE(packedYScale)", %%mm3\n\t"
2466                                                 "pxor %%mm4, %%mm4      \n\t"
2467 #ifdef HAVE_MMX2
2468 #define SCALED_CPY(src1, src2, dst1, dst2)                                      \
2469                                                 "movq " #src1 ", %%mm0  \n\t"\
2470                                                 "movq " #src1 ", %%mm5  \n\t"\
2471                                                 "movq " #src2 ", %%mm1  \n\t"\
2472                                                 "movq " #src2 ", %%mm6  \n\t"\
2473                                                 "punpcklbw %%mm0, %%mm0 \n\t"\
2474                                                 "punpckhbw %%mm5, %%mm5 \n\t"\
2475                                                 "punpcklbw %%mm1, %%mm1 \n\t"\
2476                                                 "punpckhbw %%mm6, %%mm6 \n\t"\
2477                                                 "pmulhuw %%mm3, %%mm0   \n\t"\
2478                                                 "pmulhuw %%mm3, %%mm5   \n\t"\
2479                                                 "pmulhuw %%mm3, %%mm1   \n\t"\
2480                                                 "pmulhuw %%mm3, %%mm6   \n\t"\
2481                                                 "psubw %%mm2, %%mm0     \n\t"\
2482                                                 "psubw %%mm2, %%mm5     \n\t"\
2483                                                 "psubw %%mm2, %%mm1     \n\t"\
2484                                                 "psubw %%mm2, %%mm6     \n\t"\
2485                                                 "packuswb %%mm5, %%mm0  \n\t"\
2486                                                 "packuswb %%mm6, %%mm1  \n\t"\
2487                                                 "movq %%mm0, " #dst1 "  \n\t"\
2488                                                 "movq %%mm1, " #dst2 "  \n\t"\
2489
2490 #else //HAVE_MMX2
2491 #define SCALED_CPY(src1, src2, dst1, dst2)                                      \
2492                                                 "movq " #src1 ", %%mm0  \n\t"\
2493                                                 "movq " #src1 ", %%mm5  \n\t"\
2494                                                 "punpcklbw %%mm4, %%mm0 \n\t"\
2495                                                 "punpckhbw %%mm4, %%mm5 \n\t"\
2496                                                 "psubw %%mm2, %%mm0     \n\t"\
2497                                                 "psubw %%mm2, %%mm5     \n\t"\
2498                                                 "movq " #src2 ", %%mm1  \n\t"\
2499                                                 "psllw $6, %%mm0        \n\t"\
2500                                                 "psllw $6, %%mm5        \n\t"\
2501                                                 "pmulhw %%mm3, %%mm0    \n\t"\
2502                                                 "movq " #src2 ", %%mm6  \n\t"\
2503                                                 "pmulhw %%mm3, %%mm5    \n\t"\
2504                                                 "punpcklbw %%mm4, %%mm1 \n\t"\
2505                                                 "punpckhbw %%mm4, %%mm6 \n\t"\
2506                                                 "psubw %%mm2, %%mm1     \n\t"\
2507                                                 "psubw %%mm2, %%mm6     \n\t"\
2508                                                 "psllw $6, %%mm1        \n\t"\
2509                                                 "psllw $6, %%mm6        \n\t"\
2510                                                 "pmulhw %%mm3, %%mm1    \n\t"\
2511                                                 "pmulhw %%mm3, %%mm6    \n\t"\
2512                                                 "packuswb %%mm5, %%mm0  \n\t"\
2513                                                 "packuswb %%mm6, %%mm1  \n\t"\
2514                                                 "movq %%mm0, " #dst1 "  \n\t"\
2515                                                 "movq %%mm1, " #dst2 "  \n\t"\
2516
2517 #endif //!HAVE_MMX2
2518
2519 SCALED_CPY((%0)       , (%0, %2)      , (%1)       , (%1, %3))
2520 SCALED_CPY((%0, %2, 2), (%%eax, %2, 2), (%1, %3, 2), (%%ebx, %3, 2))
2521 SCALED_CPY((%0, %2, 4), (%%eax, %2, 4), (%1, %3, 4), (%%ebx, %3, 4))
2522                                                 "leal (%%eax,%2,4), %%eax       \n\t"
2523                                                 "leal (%%ebx,%3,4), %%ebx       \n\t"
2524 SCALED_CPY((%%eax, %2), (%%eax, %2, 2), (%%ebx, %3), (%%ebx, %3, 2))
2525
2526
2527                                                 : : "r"(src),
2528                                                 "r"(dst),
2529                                                 "r" (srcStride),
2530                                                 "r" (dstStride)
2531                                                 : "%eax", "%ebx"
2532                                         );
2533 #else
2534                                 for(i=0; i<8; i++)
2535                                         memcpy( &(dst[dstStride*i]),
2536                                                 &(src[srcStride*i]), BLOCK_SIZE);
2537 #endif
2538         }
2539         else
2540         {
2541 #ifdef HAVE_MMX
2542                                         asm volatile(
2543                                                 "leal (%0,%2), %%eax    \n\t"
2544                                                 "leal (%1,%3), %%ebx    \n\t"
2545
2546 #define SIMPLE_CPY(src1, src2, dst1, dst2)                              \
2547                                                 "movq " #src1 ", %%mm0  \n\t"\
2548                                                 "movq " #src2 ", %%mm1  \n\t"\
2549                                                 "movq %%mm0, " #dst1 "  \n\t"\
2550                                                 "movq %%mm1, " #dst2 "  \n\t"\
2551
2552 SIMPLE_CPY((%0)       , (%0, %2)      , (%1)       , (%1, %3))
2553 SIMPLE_CPY((%0, %2, 2), (%%eax, %2, 2), (%1, %3, 2), (%%ebx, %3, 2))
2554 SIMPLE_CPY((%0, %2, 4), (%%eax, %2, 4), (%1, %3, 4), (%%ebx, %3, 4))
2555                                                 "leal (%%eax,%2,4), %%eax       \n\t"
2556                                                 "leal (%%ebx,%3,4), %%ebx       \n\t"
2557 SIMPLE_CPY((%%eax, %2), (%%eax, %2, 2), (%%ebx, %3), (%%ebx, %3, 2))
2558
2559                                                 : : "r" (src),
2560                                                 "r" (dst),
2561                                                 "r" (srcStride),
2562                                                 "r" (dstStride)
2563                                                 : "%eax", "%ebx"
2564                                         );
2565 #else
2566                                 for(i=0; i<8; i++)
2567                                         memcpy( &(dst[dstStride*i]),
2568                                                 &(src[srcStride*i]), BLOCK_SIZE);
2569 #endif
2570         }
2571 }
2572
2573 /**
2574  * Duplicates the given 8 src pixels ? times upward
2575  */
2576 static inline void RENAME(duplicate)(uint8_t src[], int stride)
2577 {
2578 #ifdef HAVE_MMX
2579         asm volatile(
2580                 "movq (%0), %%mm0               \n\t"
2581                 "addl %1, %0                    \n\t"
2582                 "movq %%mm0, (%0)               \n\t"
2583                 "movq %%mm0, (%0, %1)           \n\t"
2584                 "movq %%mm0, (%0, %1, 2)        \n\t"
2585                 : "+r" (src)
2586                 : "r" (-stride)
2587         );
2588 #else
2589         int i;
2590         uint8_t *p=src;
2591         for(i=0; i<3; i++)
2592         {
2593                 p-= stride;
2594                 memcpy(p, src, 8);
2595         }
2596 #endif
2597 }
2598
2599 /**
2600  * Filters array of bytes (Y or U or V values)
2601  */
2602 static void RENAME(postProcess)(uint8_t src[], int srcStride, uint8_t dst[], int dstStride, int width, int height,
2603         QP_STORE_T QPs[], int QPStride, int isColor, struct PPMode *ppMode)
2604 {
2605         int x,y;
2606 #ifdef COMPILE_TIME_MODE
2607         const int mode= COMPILE_TIME_MODE;
2608 #else
2609         const int mode= isColor ? ppMode->chromMode : ppMode->lumMode;
2610 #endif
2611         /* we need 64bit here otherwise we´ll going to have a problem
2612            after watching a black picture for 5 hours*/
2613         static uint64_t *yHistogram= NULL;
2614         int black=0, white=255; // blackest black and whitest white in the picture
2615         int QPCorrecture= 256*256;
2616
2617         /* Temporary buffers for handling the last row(s) */
2618         static uint8_t *tempDst= NULL;
2619         static uint8_t *tempSrc= NULL;
2620
2621         /* Temporary buffers for handling the last block */
2622         static uint8_t *tempDstBlock= NULL;
2623         static uint8_t *tempSrcBlock= NULL;
2624
2625         /* Temporal noise reducing buffers */
2626         static uint8_t *tempBlured[3]= {NULL,NULL,NULL};
2627         static uint32_t *tempBluredPast[3]= {NULL,NULL,NULL};
2628
2629         int copyAhead;
2630
2631 #ifdef PP_FUNNY_STRIDE
2632         uint8_t *dstBlockPtrBackup;
2633         uint8_t *srcBlockPtrBackup;
2634 #endif
2635
2636 #ifdef MORE_TIMING
2637         long long T0, T1, diffTime=0;
2638 #endif
2639 #ifdef TIMING
2640         long long memcpyTime=0, vertTime=0, horizTime=0, sumTime;
2641         sumTime= rdtsc();
2642 #endif
2643         dcOffset= ppMode->maxDcDiff;
2644         dcThreshold= ppMode->maxDcDiff*2 + 1;
2645
2646 #ifdef HAVE_MMX
2647         maxTmpNoise[0]= ppMode->maxTmpNoise[0];
2648         maxTmpNoise[1]= ppMode->maxTmpNoise[1];
2649         maxTmpNoise[2]= ppMode->maxTmpNoise[2];
2650
2651         mmxDCOffset= 0x7F - dcOffset;
2652         mmxDCThreshold= 0x7F - dcThreshold;
2653
2654         mmxDCOffset*= 0x0101010101010101LL;
2655         mmxDCThreshold*= 0x0101010101010101LL;
2656 #endif
2657
2658         if(mode & CUBIC_IPOL_DEINT_FILTER) copyAhead=16;
2659         else if(mode & LINEAR_BLEND_DEINT_FILTER) copyAhead=14;
2660         else if(   (mode & V_DEBLOCK)
2661                 || (mode & LINEAR_IPOL_DEINT_FILTER)
2662                 || (mode & MEDIAN_DEINT_FILTER)) copyAhead=13;
2663         else if(mode & V_X1_FILTER) copyAhead=11;
2664         else if(mode & V_RK1_FILTER) copyAhead=10;
2665         else if(mode & DERING) copyAhead=9;
2666         else copyAhead=8;
2667
2668         copyAhead-= 8;
2669
2670         if(tempDst==NULL)
2671         {
2672                 tempDst= (uint8_t*)memalign(8, 1024*24);
2673                 tempSrc= (uint8_t*)memalign(8, 1024*24);
2674                 tempDstBlock= (uint8_t*)memalign(8, 1024*24);
2675                 tempSrcBlock= (uint8_t*)memalign(8, 1024*24);
2676         }
2677
2678         if(tempBlured[isColor]==NULL && (mode & TEMP_NOISE_FILTER))
2679         {
2680 //              printf("%d %d %d\n", isColor, dstStride, height);
2681                 //FIXME works only as long as the size doesnt increase
2682                 //Note:the +17*1024 is just there so i dont have to worry about r/w over te end
2683                 tempBlured[isColor]= (uint8_t*)memalign(8, dstStride*((height+7)&(~7)) + 17*1024);
2684                 tempBluredPast[isColor]= (uint32_t*)memalign(8, 256*((height+7)&(~7))/2 + 17*1024);
2685
2686                 memset(tempBlured[isColor], 0, dstStride*((height+7)&(~7)) + 17*1024);
2687                 memset(tempBluredPast[isColor], 0, 256*((height+7)&(~7))/2 + 17*1024);
2688         }
2689
2690         if(!yHistogram)
2691         {
2692                 int i;
2693                 yHistogram= (uint64_t*)malloc(8*256);
2694                 for(i=0; i<256; i++) yHistogram[i]= width*height/64*15/256;
2695
2696                 if(mode & FULL_Y_RANGE)
2697                 {
2698                         ppMode->maxAllowedY=255;
2699                         ppMode->minAllowedY=0;
2700                 }
2701         }
2702
2703         if(!isColor)
2704         {
2705                 uint64_t sum= 0;
2706                 int i;
2707                 static int framenum= -1;
2708                 uint64_t maxClipped;
2709                 uint64_t clipped;
2710                 double scale;
2711
2712                 framenum++;
2713                 if(framenum == 1) yHistogram[0]= width*height/64*15/256;
2714
2715                 for(i=0; i<256; i++)
2716                 {
2717                         sum+= yHistogram[i];
2718 //                      printf("%d ", yHistogram[i]);
2719                 }
2720 //              printf("\n\n");
2721
2722                 /* we allways get a completly black picture first */
2723                 maxClipped= (uint64_t)(sum * maxClippedThreshold);
2724
2725                 clipped= sum;
2726                 for(black=255; black>0; black--)
2727                 {
2728                         if(clipped < maxClipped) break;
2729                         clipped-= yHistogram[black];
2730                 }
2731
2732                 clipped= sum;
2733                 for(white=0; white<256; white++)
2734                 {
2735                         if(clipped < maxClipped) break;
2736                         clipped-= yHistogram[white];
2737                 }
2738
2739                 scale= (double)(ppMode->maxAllowedY - ppMode->minAllowedY) / (double)(white-black);
2740
2741 #ifdef HAVE_MMX2
2742                 packedYScale= (uint16_t)(scale*256.0 + 0.5);
2743                 packedYOffset= (((black*packedYScale)>>8) - ppMode->minAllowedY) & 0xFFFF;
2744 #else
2745                 packedYScale= (uint16_t)(scale*1024.0 + 0.5);
2746                 packedYOffset= (black - ppMode->minAllowedY) & 0xFFFF;
2747 #endif
2748
2749                 packedYOffset|= packedYOffset<<32;
2750                 packedYOffset|= packedYOffset<<16;
2751
2752                 packedYScale|= packedYScale<<32;
2753                 packedYScale|= packedYScale<<16;
2754
2755                 if(mode & LEVEL_FIX)    QPCorrecture= (int)(scale*256*256 + 0.5);
2756                 else                    QPCorrecture= 256*256;
2757         }
2758         else
2759         {
2760                 packedYScale= 0x0100010001000100LL;
2761                 packedYOffset= 0;
2762                 QPCorrecture= 256*256;
2763         }
2764
2765         /* copy & deinterlace first row of blocks */
2766         y=-BLOCK_SIZE;
2767         {
2768                 uint8_t *srcBlock= &(src[y*srcStride]);
2769                 uint8_t *dstBlock= tempDst + dstStride;
2770
2771                 // From this point on it is guranteed that we can read and write 16 lines downward
2772                 // finish 1 block before the next otherwise we´ll might have a problem
2773                 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
2774                 for(x=0; x<width; x+=BLOCK_SIZE)
2775                 {
2776
2777 #ifdef HAVE_MMX2
2778 /*
2779                         prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
2780                         prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
2781                         prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
2782                         prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
2783 */
2784
2785                         asm(
2786                                 "movl %4, %%eax                 \n\t"
2787                                 "shrl $2, %%eax                 \n\t"
2788                                 "andl $6, %%eax                 \n\t"
2789                                 "addl %5, %%eax                 \n\t"
2790                                 "movl %%eax, %%ebx              \n\t"
2791                                 "imul %1, %%eax                 \n\t"
2792                                 "imul %3, %%ebx                 \n\t"
2793                                 "prefetchnta 32(%%eax, %0)      \n\t"
2794                                 "prefetcht0 32(%%ebx, %2)       \n\t"
2795                                 "addl %1, %%eax                 \n\t"
2796                                 "addl %3, %%ebx                 \n\t"
2797                                 "prefetchnta 32(%%eax, %0)      \n\t"
2798                                 "prefetcht0 32(%%ebx, %2)       \n\t"
2799                         :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),
2800                         "m" (x), "m" (copyAhead)
2801                         : "%eax", "%ebx"
2802                         );
2803
2804 #elif defined(HAVE_3DNOW)
2805 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...
2806 /*                      prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
2807                         prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
2808                         prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
2809                         prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
2810 */
2811 #endif
2812
2813                         RENAME(blockCopy)(dstBlock + dstStride*8, dstStride,
2814                                 srcBlock + srcStride*8, srcStride, mode & LEVEL_FIX);
2815
2816                         RENAME(duplicate)(dstBlock + dstStride*8, dstStride);
2817
2818                         if(mode & LINEAR_IPOL_DEINT_FILTER)
2819                                 RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
2820                         else if(mode & LINEAR_BLEND_DEINT_FILTER)
2821                                 RENAME(deInterlaceBlendLinear)(dstBlock, dstStride);
2822                         else if(mode & MEDIAN_DEINT_FILTER)
2823                                 RENAME(deInterlaceMedian)(dstBlock, dstStride);
2824                         else if(mode & CUBIC_IPOL_DEINT_FILTER)
2825                                 RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
2826 /*                      else if(mode & CUBIC_BLEND_DEINT_FILTER)
2827                                 RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
2828 */
2829                         dstBlock+=8;
2830                         srcBlock+=8;
2831                 }
2832                 memcpy(dst, tempDst + 9*dstStride, copyAhead*dstStride );
2833         }
2834
2835         for(y=0; y<height; y+=BLOCK_SIZE)
2836         {
2837                 //1% speedup if these are here instead of the inner loop
2838                 uint8_t *srcBlock= &(src[y*srcStride]);
2839                 uint8_t *dstBlock= &(dst[y*dstStride]);
2840 #ifdef HAVE_MMX
2841                 uint8_t *tempBlock1= tempBlocks;
2842                 uint8_t *tempBlock2= tempBlocks + 8;
2843 #endif
2844 #ifdef ARCH_X86
2845                 int *QPptr= isColor ? &QPs[(y>>3)*QPStride] :&QPs[(y>>4)*QPStride];
2846                 int QPDelta= isColor ? (-1) : 1<<31;
2847                 int QPFrac= 1<<30;
2848 #endif
2849                 int QP=0;
2850                 /* can we mess with a 8x16 block from srcBlock/dstBlock downwards and 1 line upwards
2851                    if not than use a temporary buffer */
2852                 if(y+15 >= height)
2853                 {
2854                         int i;
2855                         /* copy from line (copyAhead) to (copyAhead+7) of src, these will be copied with
2856                            blockcopy to dst later */
2857                         memcpy(tempSrc + srcStride*copyAhead, srcBlock + srcStride*copyAhead,
2858                                 srcStride*MAX(height-y-copyAhead, 0) );
2859
2860                         /* duplicate last line of src to fill the void upto line (copyAhead+7) */
2861                         for(i=MAX(height-y, 8); i<copyAhead+8; i++)
2862                                 memcpy(tempSrc + srcStride*i, src + srcStride*(height-1), srcStride);
2863
2864                         /* copy up to (copyAhead+1) lines of dst (line -1 to (copyAhead-1))*/
2865                         memcpy(tempDst, dstBlock - dstStride, dstStride*MIN(height-y+1, copyAhead+1) );
2866
2867                         /* duplicate last line of dst to fill the void upto line (copyAhead) */
2868                         for(i=height-y+1; i<=copyAhead; i++)
2869                                 memcpy(tempDst + dstStride*i, dst + dstStride*(height-1), dstStride);
2870
2871                         dstBlock= tempDst + dstStride;
2872                         srcBlock= tempSrc;
2873                 }
2874
2875                 // From this point on it is guranteed that we can read and write 16 lines downward
2876                 // finish 1 block before the next otherwise we´ll might have a problem
2877                 // with the L1 Cache of the P4 ... or only a few blocks at a time or soemthing
2878                 for(x=0; x<width; x+=BLOCK_SIZE)
2879                 {
2880                         const int stride= dstStride;
2881 #ifdef HAVE_MMX
2882                         uint8_t *tmpXchg;
2883 #endif
2884 #ifdef ARCH_X86
2885                         QP= *QPptr;
2886                         asm volatile(
2887                                 "addl %2, %1            \n\t"
2888                                 "sbbl %%eax, %%eax      \n\t"
2889                                 "shll $2, %%eax         \n\t"
2890                                 "subl %%eax, %0         \n\t"
2891                                 : "+r" (QPptr), "+m" (QPFrac)
2892                                 : "r" (QPDelta)
2893                                 : "%eax"
2894                         );
2895 #else
2896                         QP= isColor ?
2897                                 QPs[(y>>3)*QPStride + (x>>3)]:
2898                                 QPs[(y>>4)*QPStride + (x>>4)];
2899 #endif
2900                         if(!isColor)
2901                         {
2902                                 QP= (QP* QPCorrecture + 256*128)>>16;
2903                                 yHistogram[ srcBlock[srcStride*12 + 4] ]++;
2904                         }
2905 #ifdef HAVE_MMX
2906                         asm volatile(
2907                                 "movd %0, %%mm7                                 \n\t"
2908                                 "packuswb %%mm7, %%mm7                          \n\t" // 0, 0, 0, QP, 0, 0, 0, QP
2909                                 "packuswb %%mm7, %%mm7                          \n\t" // 0,QP, 0, QP, 0,QP, 0, QP
2910                                 "packuswb %%mm7, %%mm7                          \n\t" // QP,..., QP
2911                                 "movq %%mm7, "MANGLE(pQPb)"                     \n\t"
2912                                 : : "r" (QP)
2913                         );
2914 #endif
2915
2916 #ifdef MORE_TIMING
2917                         T0= rdtsc();
2918 #endif
2919
2920 #ifdef HAVE_MMX2
2921 /*
2922                         prefetchnta(srcBlock + (((x>>2)&6) + 5)*srcStride + 32);
2923                         prefetchnta(srcBlock + (((x>>2)&6) + 6)*srcStride + 32);
2924                         prefetcht0(dstBlock + (((x>>2)&6) + 5)*dstStride + 32);
2925                         prefetcht0(dstBlock + (((x>>2)&6) + 6)*dstStride + 32);
2926 */
2927
2928                         asm(
2929                                 "movl %4, %%eax                 \n\t"
2930                                 "shrl $2, %%eax                 \n\t"
2931                                 "andl $6, %%eax                 \n\t"
2932                                 "addl %5, %%eax                 \n\t"
2933                                 "movl %%eax, %%ebx              \n\t"
2934                                 "imul %1, %%eax                 \n\t"
2935                                 "imul %3, %%ebx                 \n\t"
2936                                 "prefetchnta 32(%%eax, %0)      \n\t"
2937                                 "prefetcht0 32(%%ebx, %2)       \n\t"
2938                                 "addl %1, %%eax                 \n\t"
2939                                 "addl %3, %%ebx                 \n\t"
2940                                 "prefetchnta 32(%%eax, %0)      \n\t"
2941                                 "prefetcht0 32(%%ebx, %2)       \n\t"
2942                         :: "r" (srcBlock), "r" (srcStride), "r" (dstBlock), "r" (dstStride),
2943                         "m" (x), "m" (copyAhead)
2944                         : "%eax", "%ebx"
2945                         );
2946
2947 #elif defined(HAVE_3DNOW)
2948 //FIXME check if this is faster on an 3dnow chip or if its faster without the prefetch or ...
2949 /*                      prefetch(srcBlock + (((x>>3)&3) + 5)*srcStride + 32);
2950                         prefetch(srcBlock + (((x>>3)&3) + 9)*srcStride + 32);
2951                         prefetchw(dstBlock + (((x>>3)&3) + 5)*dstStride + 32);
2952                         prefetchw(dstBlock + (((x>>3)&3) + 9)*dstStride + 32);
2953 */
2954 #endif
2955
2956 #ifdef PP_FUNNY_STRIDE
2957                         //can we mess with a 8x16 block, if not use a temp buffer, yes again
2958                         if(x+7 >= width)
2959                         {
2960                                 int i;
2961                                 dstBlockPtrBackup= dstBlock;
2962                                 srcBlockPtrBackup= srcBlock;
2963
2964                                 for(i=0;i<BLOCK_SIZE*2; i++)
2965                                 {
2966                                         memcpy(tempSrcBlock+i*srcStride, srcBlock+i*srcStride, width-x);
2967                                         memcpy(tempDstBlock+i*dstStride, dstBlock+i*dstStride, width-x);
2968                                 }
2969
2970                                 dstBlock= tempDstBlock;
2971                                 srcBlock= tempSrcBlock;
2972                         }
2973 #endif
2974
2975                         RENAME(blockCopy)(dstBlock + dstStride*copyAhead, dstStride,
2976                                 srcBlock + srcStride*copyAhead, srcStride, mode & LEVEL_FIX);
2977
2978                         if(mode & LINEAR_IPOL_DEINT_FILTER)
2979                                 RENAME(deInterlaceInterpolateLinear)(dstBlock, dstStride);
2980                         else if(mode & LINEAR_BLEND_DEINT_FILTER)
2981                                 RENAME(deInterlaceBlendLinear)(dstBlock, dstStride);
2982                         else if(mode & MEDIAN_DEINT_FILTER)
2983                                 RENAME(deInterlaceMedian)(dstBlock, dstStride);
2984                         else if(mode & CUBIC_IPOL_DEINT_FILTER)
2985                                 RENAME(deInterlaceInterpolateCubic)(dstBlock, dstStride);
2986 /*                      else if(mode & CUBIC_BLEND_DEINT_FILTER)
2987                                 RENAME(deInterlaceBlendCubic)(dstBlock, dstStride);
2988 */
2989
2990                         /* only deblock if we have 2 blocks */
2991                         if(y + 8 < height)
2992                         {
2993 #ifdef MORE_TIMING
2994                                 T1= rdtsc();
2995                                 memcpyTime+= T1-T0;
2996                                 T0=T1;
2997 #endif
2998                                 if(mode & V_RK1_FILTER)
2999                                         RENAME(vertRK1Filter)(dstBlock, stride, QP);
3000                                 else if(mode & V_X1_FILTER)
3001                                         RENAME(vertX1Filter)(dstBlock, stride, QP);
3002                                 else if(mode & V_DEBLOCK)
3003                                 {
3004                                         if( RENAME(isVertDC)(dstBlock, stride))
3005                                         {
3006                                                 if(RENAME(isVertMinMaxOk)(dstBlock, stride, QP))
3007                                                         RENAME(doVertLowPass)(dstBlock, stride, QP);
3008                                         }
3009                                         else
3010                                                 RENAME(doVertDefFilter)(dstBlock, stride, QP);
3011                                 }
3012 #ifdef MORE_TIMING
3013                                 T1= rdtsc();
3014                                 vertTime+= T1-T0;
3015                                 T0=T1;
3016 #endif
3017                         }
3018
3019 #ifdef HAVE_MMX
3020                         RENAME(transpose1)(tempBlock1, tempBlock2, dstBlock, dstStride);
3021 #endif
3022                         /* check if we have a previous block to deblock it with dstBlock */
3023                         if(x - 8 >= 0)
3024                         {
3025 #ifdef MORE_TIMING
3026                                 T0= rdtsc();
3027 #endif
3028 #ifdef HAVE_MMX
3029                                 if(mode & H_RK1_FILTER)
3030                                         RENAME(vertRK1Filter)(tempBlock1, 16, QP);
3031                                 else if(mode & H_X1_FILTER)
3032                                         RENAME(vertX1Filter)(tempBlock1, 16, QP);
3033                                 else if(mode & H_DEBLOCK)
3034                                 {
3035                                         if( RENAME(isVertDC)(tempBlock1, 16) )
3036                                         {
3037                                                 if(RENAME(isVertMinMaxOk)(tempBlock1, 16, QP))
3038                                                         RENAME(doVertLowPass)(tempBlock1, 16, QP);
3039                                         }
3040                                         else
3041                                                 RENAME(doVertDefFilter)(tempBlock1, 16, QP);
3042                                 }
3043
3044                                 RENAME(transpose2)(dstBlock-4, dstStride, tempBlock1 + 4*16);
3045
3046 #else
3047                                 if(mode & H_X1_FILTER)
3048                                         horizX1Filter(dstBlock-4, stride, QP);
3049                                 else if(mode & H_DEBLOCK)
3050                                 {
3051                                         if( isHorizDC(dstBlock-4, stride))
3052                                         {
3053                                                 if(isHorizMinMaxOk(dstBlock-4, stride, QP))
3054                                                         doHorizLowPass(dstBlock-4, stride, QP);
3055                                         }
3056                                         else
3057                                                 doHorizDefFilter(dstBlock-4, stride, QP);
3058                                 }
3059 #endif
3060 #ifdef MORE_TIMING
3061                                 T1= rdtsc();
3062                                 horizTime+= T1-T0;
3063                                 T0=T1;
3064 #endif
3065                                 if(mode & DERING)
3066                                 {
3067                                 //FIXME filter first line
3068                                         if(y>0) RENAME(dering)(dstBlock - stride - 8, stride, QP);
3069                                 }
3070
3071                                 if(mode & TEMP_NOISE_FILTER)
3072                                 {
3073                                         RENAME(tempNoiseReducer)(dstBlock-8, stride,
3074                                                 tempBlured[isColor] + y*dstStride + x,
3075                                                 tempBluredPast[isColor] + (y>>3)*256 + (x>>3),
3076                                                 ppMode->maxTmpNoise);
3077                                 }
3078                         }
3079
3080 #ifdef PP_FUNNY_STRIDE
3081                         /* did we use a tmp-block buffer */
3082                         if(x+7 >= width)
3083                         {
3084                                 int i;
3085                                 dstBlock= dstBlockPtrBackup;
3086                                 srcBlock= srcBlockPtrBackup;
3087
3088                                 for(i=0;i<BLOCK_SIZE*2; i++)
3089                                 {
3090                                         memcpy(dstBlock+i*dstStride, tempDstBlock+i*dstStride, width-x);
3091                                 }
3092                         }
3093 #endif
3094
3095                         dstBlock+=8;
3096                         srcBlock+=8;
3097
3098 #ifdef HAVE_MMX
3099                         tmpXchg= tempBlock1;
3100                         tempBlock1= tempBlock2;
3101                         tempBlock2 = tmpXchg;
3102 #endif
3103                 }
3104
3105                 if(mode & DERING)
3106                 {
3107                                 if(y > 0) RENAME(dering)(dstBlock - dstStride - 8, dstStride, QP);
3108                 }
3109
3110                 if((mode & TEMP_NOISE_FILTER))
3111                 {
3112                         RENAME(tempNoiseReducer)(dstBlock-8, dstStride,
3113                                 tempBlured[isColor] + y*dstStride + x,
3114                                 tempBluredPast[isColor] + (y>>3)*256 + (x>>3),
3115                                 ppMode->maxTmpNoise);
3116                 }
3117
3118                 /* did we use a tmp buffer for the last lines*/
3119                 if(y+15 >= height)
3120                 {
3121                         uint8_t *dstBlock= &(dst[y*dstStride]);
3122                         memcpy(dstBlock, tempDst + dstStride, dstStride*(height-y) );
3123                 }
3124 /*
3125                 for(x=0; x<width; x+=32)
3126                 {
3127                         volatile int i;
3128                         i+=     + dstBlock[x + 7*dstStride] + dstBlock[x + 8*dstStride]
3129                                 + dstBlock[x + 9*dstStride] + dstBlock[x +10*dstStride]
3130                                 + dstBlock[x +11*dstStride] + dstBlock[x +12*dstStride];
3131 //                              + dstBlock[x +13*dstStride]
3132 //                              + dstBlock[x +14*dstStride] + dstBlock[x +15*dstStride];
3133                 }*/
3134         }
3135 #ifdef HAVE_3DNOW
3136         asm volatile("femms");
3137 #elif defined (HAVE_MMX)
3138         asm volatile("emms");
3139 #endif
3140
3141 #ifdef TIMING
3142         // FIXME diff is mostly the time spent for rdtsc (should subtract that but ...)
3143         sumTime= rdtsc() - sumTime;
3144         if(!isColor)
3145                 printf("cpy:%4dk, vert:%4dk, horiz:%4dk, sum:%4dk, diff:%4dk, color: %d/%d    \r",
3146                         (int)(memcpyTime/1000), (int)(vertTime/1000), (int)(horizTime/1000),
3147                         (int)(sumTime/1000), (int)((sumTime-memcpyTime-vertTime-horizTime)/1000)
3148                         , black, white);
3149 #endif
3150 #ifdef DEBUG_BRIGHTNESS
3151         if(!isColor)
3152         {
3153                 int max=1;
3154                 int i;
3155                 for(i=0; i<256; i++)
3156                         if(yHistogram[i] > max) max=yHistogram[i];
3157
3158                 for(i=1; i<256; i++)
3159                 {
3160                         int x;
3161                         int start=yHistogram[i-1]/(max/256+1);
3162                         int end=yHistogram[i]/(max/256+1);
3163                         int inc= end > start ? 1 : -1;
3164                         for(x=start; x!=end+inc; x+=inc)
3165                                 dst[ i*dstStride + x]+=128;
3166                 }
3167
3168                 for(i=0; i<100; i+=2)
3169                 {
3170                         dst[ (white)*dstStride + i]+=128;
3171                         dst[ (black)*dstStride + i]+=128;
3172                 }
3173
3174         }
3175 #endif
3176
3177 }