git.sesse.net Git - ffmpeg/blob - postproc/swscale.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 /*
  20   supported Input formats: YV12 (grayscale soon too)
  21   supported output formats: YV12, BGR15, BGR16, BGR24, BGR32 (grayscale soon too)
  22 */
  23
  24 #include <inttypes.h>
  25 #include <string.h>
  26 #include <math.h>
  27 #include <stdio.h>
  28 #include "../config.h"
  29 #include "../mangle.h"
  30 #ifdef HAVE_MALLOC_H
  31 #include <malloc.h>
  32 #endif
  33 #include "swscale.h"
  34 #include "../cpudetect.h"
  35 #include "../libvo/img_format.h"
  36 #undef MOVNTQ
  37 #undef PAVGB
  38
  39 //#undef HAVE_MMX2
  40 //#define HAVE_3DNOW
  41 //#undef HAVE_MMX
  42 //#undef ARCH_X86
  43 #define DITHER1XBPP
  44
  45 #define RET 0xC3 //near return opcode
  46
  47 #ifdef MP_DEBUG
  48 #define ASSERT(x) if(!(x)) { printf("ASSERT " #x " failed\n"); *((int*)0)=0; }
  49 #else
  50 #define ASSERT(x) ;
  51 #endif
  52
  53 #ifdef M_PI
  54 #define PI M_PI
  55 #else
  56 #define PI 3.14159265358979323846
  57 #endif
  58
  59 extern int verbose; // defined in mplayer.c
  60 /*
  61 NOTES
  62
  63 known BUGS with known cause (no bugreports please!, but patches are welcome :) )
  64 horizontal fast_bilinear MMX2 scaler reads 1-7 samples too much (might cause a sig11)
  65
  66 Supported output formats BGR15 BGR16 BGR24 BGR32 YV12
  67 BGR15 & BGR16 MMX verions support dithering
  68 Special versions: fast Y 1:1 scaling (no interpolation in y direction)
  69
  70 TODO
  71 more intelligent missalignment avoidance for the horizontal scaler
  72 change the distance of the u & v buffer
  73 write special vertical cubic upscale version
  74 Optimize C code (yv12 / minmax)
  75 add support for packed pixel yuv input & output
  76 add support for Y8 input & output
  77 add BGR4 output support
  78 add BGR32 / BGR24 input support
  79 */
  80
  81 #define ABS(a) ((a) > 0 ? (a) : (-(a)))
  82 #define MIN(a,b) ((a) > (b) ? (b) : (a))
  83 #define MAX(a,b) ((a) < (b) ? (b) : (a))
  84
  85 #ifdef ARCH_X86
  86 #define CAN_COMPILE_X86_ASM
  87 #endif
  88
  89 #ifdef CAN_COMPILE_X86_ASM
  90 static uint64_t __attribute__((aligned(8))) yCoeff=    0x2568256825682568LL;
  91 static uint64_t __attribute__((aligned(8))) vrCoeff=   0x3343334333433343LL;
  92 static uint64_t __attribute__((aligned(8))) ubCoeff=   0x40cf40cf40cf40cfLL;
  93 static uint64_t __attribute__((aligned(8))) vgCoeff=   0xE5E2E5E2E5E2E5E2LL;
  94 static uint64_t __attribute__((aligned(8))) ugCoeff=   0xF36EF36EF36EF36ELL;
  95 static uint64_t __attribute__((aligned(8))) bF8=       0xF8F8F8F8F8F8F8F8LL;
  96 static uint64_t __attribute__((aligned(8))) bFC=       0xFCFCFCFCFCFCFCFCLL;
  97 static uint64_t __attribute__((aligned(8))) w400=      0x0400040004000400LL;
  98 static uint64_t __attribute__((aligned(8))) w80=       0x0080008000800080LL;
  99 static uint64_t __attribute__((aligned(8))) w10=       0x0010001000100010LL;
 100 static uint64_t __attribute__((aligned(8))) w02=       0x0002000200020002LL;
 101 static uint64_t __attribute__((aligned(8))) bm00001111=0x00000000FFFFFFFFLL;
 102 static uint64_t __attribute__((aligned(8))) bm00000111=0x0000000000FFFFFFLL;
 103 static uint64_t __attribute__((aligned(8))) bm11111000=0xFFFFFFFFFF000000LL;
 104
 105 static volatile uint64_t __attribute__((aligned(8))) b5Dither;
 106 static volatile uint64_t __attribute__((aligned(8))) g5Dither;
 107 static volatile uint64_t __attribute__((aligned(8))) g6Dither;
 108 static volatile uint64_t __attribute__((aligned(8))) r5Dither;
 109
 110 static uint64_t __attribute__((aligned(8))) dither4[2]={
 111         0x0103010301030103LL,
 112         0x0200020002000200LL,};
 113
 114 static uint64_t __attribute__((aligned(8))) dither8[2]={
 115         0x0602060206020602LL,
 116         0x0004000400040004LL,};
 117
 118 static uint64_t __attribute__((aligned(8))) b16Mask=   0x001F001F001F001FLL;
 119 static uint64_t __attribute__((aligned(8))) g16Mask=   0x07E007E007E007E0LL;
 120 static uint64_t __attribute__((aligned(8))) r16Mask=   0xF800F800F800F800LL;
 121 static uint64_t __attribute__((aligned(8))) b15Mask=   0x001F001F001F001FLL;
 122 static uint64_t __attribute__((aligned(8))) g15Mask=   0x03E003E003E003E0LL;
 123 static uint64_t __attribute__((aligned(8))) r15Mask=   0x7C007C007C007C00LL;
 124
 125 static uint64_t __attribute__((aligned(8))) M24A=   0x00FF0000FF0000FFLL;
 126 static uint64_t __attribute__((aligned(8))) M24B=   0xFF0000FF0000FF00LL;
 127 static uint64_t __attribute__((aligned(8))) M24C=   0x0000FF0000FF0000LL;
 128
 129 // FIXME remove
 130 static uint64_t __attribute__((aligned(8))) asm_yalpha1;
 131 static uint64_t __attribute__((aligned(8))) asm_uvalpha1;
 132 #endif
 133
 134 // clipping helper table for C implementations:
 135 static unsigned char clip_table[768];
 136
 137 static unsigned short clip_table16b[768];
 138 static unsigned short clip_table16g[768];
 139 static unsigned short clip_table16r[768];
 140 static unsigned short clip_table15b[768];
 141 static unsigned short clip_table15g[768];
 142 static unsigned short clip_table15r[768];
 143
 144 // yuv->rgb conversion tables:
 145 static    int yuvtab_2568[256];
 146 static    int yuvtab_3343[256];
 147 static    int yuvtab_0c92[256];
 148 static    int yuvtab_1a1e[256];
 149 static    int yuvtab_40cf[256];
 150 // Needed for cubic scaler to catch overflows
 151 static    int clip_yuvtab_2568[768];
 152 static    int clip_yuvtab_3343[768];
 153 static    int clip_yuvtab_0c92[768];
 154 static    int clip_yuvtab_1a1e[768];
 155 static    int clip_yuvtab_40cf[768];
 156
 157 //global sws_flags from the command line
 158 int sws_flags=0;
 159
 160 //global srcFilter
 161 SwsFilter src_filter= {NULL, NULL, NULL, NULL};
 162
 163 float sws_lum_gblur= 0.0;
 164 float sws_chr_gblur= 0.0;
 165 int sws_chr_vshift= 0;
 166 int sws_chr_hshift= 0;
 167 float sws_chr_sharpen= 0.0;
 168 float sws_lum_sharpen= 0.0;
 169
 170 /* cpuCaps combined from cpudetect and whats actually compiled in
 171    (if there is no support for something compiled in it wont appear here) */
 172 static CpuCaps cpuCaps;
 173
 174 void (*swScale)(SwsContext *context, uint8_t* src[], int srcStride[], int srcSliceY,
 175              int srcSliceH, uint8_t* dst[], int dstStride[])=NULL;
 176
 177 static SwsVector *getConvVec(SwsVector *a, SwsVector *b);
 178
 179 #ifdef CAN_COMPILE_X86_ASM
 180 void in_asm_used_var_warning_killer()
 181 {
 182  volatile int i= yCoeff+vrCoeff+ubCoeff+vgCoeff+ugCoeff+bF8+bFC+w400+w80+w10+
 183  bm00001111+bm00000111+bm11111000+b16Mask+g16Mask+r16Mask+b15Mask+g15Mask+r15Mask+asm_yalpha1+ asm_uvalpha1+
 184  M24A+M24B+M24C+w02 + b5Dither+g5Dither+r5Dither+g6Dither+dither4[0]+dither8[0];
 185  if(i) i=0;
 186 }
 187 #endif
 188
 189 static inline void yuv2yuvXinC(int16_t *lumFilter, int16_t **lumSrc, int lumFilterSize,
 190                                     int16_t *chrFilter, int16_t **chrSrc, int chrFilterSize,
 191                                     uint8_t *dest, uint8_t *uDest, uint8_t *vDest, int dstW)
 192 {
 193         //FIXME Optimize (just quickly writen not opti..)
 194         int i;
 195         for(i=0; i<dstW; i++)
 196         {
 197                 int val=0;
 198                 int j;
 199                 for(j=0; j<lumFilterSize; j++)
 200                         val += lumSrc[j][i] * lumFilter[j];
 201
 202                 dest[i]= MIN(MAX(val>>19, 0), 255);
 203         }
 204
 205         if(uDest != NULL)
 206                 for(i=0; i<(dstW>>1); i++)
 207                 {
 208                         int u=0;
 209                         int v=0;
 210                         int j;
 211                         for(j=0; j<chrFilterSize; j++)
 212                         {
 213                                 u += chrSrc[j][i] * chrFilter[j];
 214                                 v += chrSrc[j][i + 2048] * chrFilter[j];
 215                         }
 216
 217                         uDest[i]= MIN(MAX(u>>19, 0), 255);
 218                         vDest[i]= MIN(MAX(v>>19, 0), 255);
 219                 }
 220 }
 221
 222 static inline void yuv2rgbXinC(int16_t *lumFilter, int16_t **lumSrc, int lumFilterSize,
 223                                     int16_t *chrFilter, int16_t **chrSrc, int chrFilterSize,
 224                                     uint8_t *dest, int dstW, int dstFormat)
 225 {
 226         if(dstFormat==IMGFMT_BGR32)
 227         {
 228                 int i;
 229                 for(i=0; i<(dstW>>1); i++){
 230                         int j;
 231                         int Y1=0;
 232                         int Y2=0;
 233                         int U=0;
 234                         int V=0;
 235                         int Cb, Cr, Cg;
 236                         for(j=0; j<lumFilterSize; j++)
 237                         {
 238                                 Y1 += lumSrc[j][2*i] * lumFilter[j];
 239                                 Y2 += lumSrc[j][2*i+1] * lumFilter[j];
 240                         }
 241                         for(j=0; j<chrFilterSize; j++)
 242                         {
 243                                 U += chrSrc[j][i] * chrFilter[j];
 244                                 V += chrSrc[j][i+2048] * chrFilter[j];
 245                         }
 246                         Y1= clip_yuvtab_2568[ (Y1>>19) + 256 ];
 247                         Y2= clip_yuvtab_2568[ (Y2>>19) + 256 ];
 248                         U >>= 19;
 249                         V >>= 19;
 250
 251                         Cb= clip_yuvtab_40cf[U+ 256];
 252                         Cg= clip_yuvtab_1a1e[V+ 256] + yuvtab_0c92[U+ 256];
 253                         Cr= clip_yuvtab_3343[V+ 256];
 254
 255                         dest[8*i+0]=clip_table[((Y1 + Cb) >>13)];
 256                         dest[8*i+1]=clip_table[((Y1 + Cg) >>13)];
 257                         dest[8*i+2]=clip_table[((Y1 + Cr) >>13)];
 258
 259                         dest[8*i+4]=clip_table[((Y2 + Cb) >>13)];
 260                         dest[8*i+5]=clip_table[((Y2 + Cg) >>13)];
 261                         dest[8*i+6]=clip_table[((Y2 + Cr) >>13)];
 262                 }
 263         }
 264         else if(dstFormat==IMGFMT_BGR24)
 265         {
 266                 int i;
 267                 for(i=0; i<(dstW>>1); i++){
 268                         int j;
 269                         int Y1=0;
 270                         int Y2=0;
 271                         int U=0;
 272                         int V=0;
 273                         int Cb, Cr, Cg;
 274                         for(j=0; j<lumFilterSize; j++)
 275                         {
 276                                 Y1 += lumSrc[j][2*i] * lumFilter[j];
 277                                 Y2 += lumSrc[j][2*i+1] * lumFilter[j];
 278                         }
 279                         for(j=0; j<chrFilterSize; j++)
 280                         {
 281                                 U += chrSrc[j][i] * chrFilter[j];
 282                                 V += chrSrc[j][i+2048] * chrFilter[j];
 283                         }
 284                         Y1= clip_yuvtab_2568[ (Y1>>19) + 256 ];
 285                         Y2= clip_yuvtab_2568[ (Y2>>19) + 256 ];
 286                         U >>= 19;
 287                         V >>= 19;
 288
 289                         Cb= clip_yuvtab_40cf[U+ 256];
 290                         Cg= clip_yuvtab_1a1e[V+ 256] + yuvtab_0c92[U+ 256];
 291                         Cr= clip_yuvtab_3343[V+ 256];
 292
 293                         dest[0]=clip_table[((Y1 + Cb) >>13)];
 294                         dest[1]=clip_table[((Y1 + Cg) >>13)];
 295                         dest[2]=clip_table[((Y1 + Cr) >>13)];
 296
 297                         dest[3]=clip_table[((Y2 + Cb) >>13)];
 298                         dest[4]=clip_table[((Y2 + Cg) >>13)];
 299                         dest[5]=clip_table[((Y2 + Cr) >>13)];
 300                         dest+=6;
 301                 }
 302         }
 303         else if(dstFormat==IMGFMT_BGR16)
 304         {
 305                 int i;
 306 #ifdef DITHER1XBPP
 307                 static int ditherb1=1<<14;
 308                 static int ditherg1=1<<13;
 309                 static int ditherr1=2<<14;
 310                 static int ditherb2=3<<14;
 311                 static int ditherg2=3<<13;
 312                 static int ditherr2=0<<14;
 313
 314                 ditherb1 ^= (1^2)<<14;
 315                 ditherg1 ^= (1^2)<<13;
 316                 ditherr1 ^= (1^2)<<14;
 317                 ditherb2 ^= (3^0)<<14;
 318                 ditherg2 ^= (3^0)<<13;
 319                 ditherr2 ^= (3^0)<<14;
 320 #else
 321                 const int ditherb1=0;
 322                 const int ditherg1=0;
 323                 const int ditherr1=0;
 324                 const int ditherb2=0;
 325                 const int ditherg2=0;
 326                 const int ditherr2=0;
 327 #endif
 328                 for(i=0; i<(dstW>>1); i++){
 329                         int j;
 330                         int Y1=0;
 331                         int Y2=0;
 332                         int U=0;
 333                         int V=0;
 334                         int Cb, Cr, Cg;
 335                         for(j=0; j<lumFilterSize; j++)
 336                         {
 337                                 Y1 += lumSrc[j][2*i] * lumFilter[j];
 338                                 Y2 += lumSrc[j][2*i+1] * lumFilter[j];
 339                         }
 340                         for(j=0; j<chrFilterSize; j++)
 341                         {
 342                                 U += chrSrc[j][i] * chrFilter[j];
 343                                 V += chrSrc[j][i+2048] * chrFilter[j];
 344                         }
 345                         Y1= clip_yuvtab_2568[ (Y1>>19) + 256 ];
 346                         Y2= clip_yuvtab_2568[ (Y2>>19) + 256 ];
 347                         U >>= 19;
 348                         V >>= 19;
 349
 350                         Cb= clip_yuvtab_40cf[U+ 256];
 351                         Cg= clip_yuvtab_1a1e[V+ 256] + yuvtab_0c92[U+ 256];
 352                         Cr= clip_yuvtab_3343[V+ 256];
 353
 354                         ((uint16_t*)dest)[2*i] =
 355                                 clip_table16b[(Y1 + Cb + ditherb1) >>13] |
 356                                 clip_table16g[(Y1 + Cg + ditherg1) >>13] |
 357                                 clip_table16r[(Y1 + Cr + ditherr1) >>13];
 358
 359                         ((uint16_t*)dest)[2*i+1] =
 360                                 clip_table16b[(Y2 + Cb + ditherb2) >>13] |
 361                                 clip_table16g[(Y2 + Cg + ditherg2) >>13] |
 362                                 clip_table16r[(Y2 + Cr + ditherr2) >>13];
 363                 }
 364         }
 365         else if(dstFormat==IMGFMT_BGR15)
 366         {
 367                 int i;
 368 #ifdef DITHER1XBPP
 369                 static int ditherb1=1<<14;
 370                 static int ditherg1=1<<14;
 371                 static int ditherr1=2<<14;
 372                 static int ditherb2=3<<14;
 373                 static int ditherg2=3<<14;
 374                 static int ditherr2=0<<14;
 375
 376                 ditherb1 ^= (1^2)<<14;
 377                 ditherg1 ^= (1^2)<<14;
 378                 ditherr1 ^= (1^2)<<14;
 379                 ditherb2 ^= (3^0)<<14;
 380                 ditherg2 ^= (3^0)<<14;
 381                 ditherr2 ^= (3^0)<<14;
 382 #else
 383                 const int ditherb1=0;
 384                 const int ditherg1=0;
 385                 const int ditherr1=0;
 386                 const int ditherb2=0;
 387                 const int ditherg2=0;
 388                 const int ditherr2=0;
 389 #endif
 390                 for(i=0; i<(dstW>>1); i++){
 391                         int j;
 392                         int Y1=0;
 393                         int Y2=0;
 394                         int U=0;
 395                         int V=0;
 396                         int Cb, Cr, Cg;
 397                         for(j=0; j<lumFilterSize; j++)
 398                         {
 399                                 Y1 += lumSrc[j][2*i] * lumFilter[j];
 400                                 Y2 += lumSrc[j][2*i+1] * lumFilter[j];
 401                         }
 402                         for(j=0; j<chrFilterSize; j++)
 403                         {
 404                                 U += chrSrc[j][i] * chrFilter[j];
 405                                 V += chrSrc[j][i+2048] * chrFilter[j];
 406                         }
 407                         Y1= clip_yuvtab_2568[ (Y1>>19) + 256 ];
 408                         Y2= clip_yuvtab_2568[ (Y2>>19) + 256 ];
 409                         U >>= 19;
 410                         V >>= 19;
 411
 412                         Cb= clip_yuvtab_40cf[U+ 256];
 413                         Cg= clip_yuvtab_1a1e[V+ 256] + yuvtab_0c92[U+ 256];
 414                         Cr= clip_yuvtab_3343[V+ 256];
 415
 416                         ((uint16_t*)dest)[2*i] =
 417                                 clip_table15b[(Y1 + Cb + ditherb1) >>13] |
 418                                 clip_table15g[(Y1 + Cg + ditherg1) >>13] |
 419                                 clip_table15r[(Y1 + Cr + ditherr1) >>13];
 420
 421                         ((uint16_t*)dest)[2*i+1] =
 422                                 clip_table15b[(Y2 + Cb + ditherb2) >>13] |
 423                                 clip_table15g[(Y2 + Cg + ditherg2) >>13] |
 424                                 clip_table15r[(Y2 + Cr + ditherr2) >>13];
 425                 }
 426         }
 427 }
 428
 429
 430 //Note: we have C, X86, MMX, MMX2, 3DNOW version therse no 3DNOW+MMX2 one
 431 //Plain C versions
 432 #if !defined (HAVE_MMX) || defined (RUNTIME_CPUDETECT)
 433 #define COMPILE_C
 434 #endif
 435
 436 #ifdef CAN_COMPILE_X86_ASM
 437
 438 #if (defined (HAVE_MMX) && !defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
 439 #define COMPILE_MMX
 440 #endif
 441
 442 #if defined (HAVE_MMX2) || defined (RUNTIME_CPUDETECT)
 443 #define COMPILE_MMX2
 444 #endif
 445
 446 #if (defined (HAVE_3DNOW) && !defined (HAVE_MMX2)) || defined (RUNTIME_CPUDETECT)
 447 #define COMPILE_3DNOW
 448 #endif
 449 #endif //CAN_COMPILE_X86_ASM
 450
 451 #undef HAVE_MMX
 452 #undef HAVE_MMX2
 453 #undef HAVE_3DNOW
 454
 455 #ifdef COMPILE_C
 456 #undef HAVE_MMX
 457 #undef HAVE_MMX2
 458 #undef HAVE_3DNOW
 459 #define RENAME(a) a ## _C
 460 #include "swscale_template.c"
 461 #endif
 462
 463 #ifdef CAN_COMPILE_X86_ASM
 464
 465 //X86 versions
 466 /*
 467 #undef RENAME
 468 #undef HAVE_MMX
 469 #undef HAVE_MMX2
 470 #undef HAVE_3DNOW
 471 #define ARCH_X86
 472 #define RENAME(a) a ## _X86
 473 #include "swscale_template.c"
 474 */
 475 //MMX versions
 476 #ifdef COMPILE_MMX
 477 #undef RENAME
 478 #define HAVE_MMX
 479 #undef HAVE_MMX2
 480 #undef HAVE_3DNOW
 481 #define RENAME(a) a ## _MMX
 482 #include "swscale_template.c"
 483 #endif
 484
 485 //MMX2 versions
 486 #ifdef COMPILE_MMX2
 487 #undef RENAME
 488 #define HAVE_MMX
 489 #define HAVE_MMX2
 490 #undef HAVE_3DNOW
 491 #define RENAME(a) a ## _MMX2
 492 #include "swscale_template.c"
 493 #endif
 494
 495 //3DNOW versions
 496 #ifdef COMPILE_3DNOW
 497 #undef RENAME
 498 #define HAVE_MMX
 499 #undef HAVE_MMX2
 500 #define HAVE_3DNOW
 501 #define RENAME(a) a ## _3DNow
 502 #include "swscale_template.c"
 503 #endif
 504
 505 #endif //CAN_COMPILE_X86_ASM
 506
 507 // minor note: the HAVE_xyz is messed up after that line so dont use it
 508
 509
 510 // old global scaler, dont use for new code, unless it uses only the stuff from the command line
 511 // will use sws_flags from the command line
 512 void SwScale_YV12slice(unsigned char* src[], int srcStride[], int srcSliceY ,
 513                              int srcSliceH, uint8_t* dst[], int dstStride, int dstbpp,
 514                              int srcW, int srcH, int dstW, int dstH){
 515
 516         static SwsContext *context=NULL;
 517         int dstFormat;
 518         int flags=0;
 519         static int firstTime=1;
 520         int dstStride3[3]= {dstStride, dstStride>>1, dstStride>>1};
 521
 522         if(firstTime)
 523         {
 524 #ifdef ARCH_X86
 525                 if(gCpuCaps.hasMMX)
 526                         asm volatile("emms\n\t"::: "memory"); //FIXME this shouldnt be required but it IS (even for non mmx versions)
 527 #endif
 528                 flags= SWS_PRINT_INFO;
 529                 firstTime=0;
 530
 531                 if(src_filter.lumH) freeVec(src_filter.lumH);
 532                 if(src_filter.lumV) freeVec(src_filter.lumV);
 533                 if(src_filter.chrH) freeVec(src_filter.chrH);
 534                 if(src_filter.chrV) freeVec(src_filter.chrV);
 535
 536                 if(sws_lum_gblur!=0.0){
 537                         src_filter.lumH= getGaussianVec(sws_lum_gblur, 3.0);
 538                         src_filter.lumV= getGaussianVec(sws_lum_gblur, 3.0);
 539                 }else{
 540                         src_filter.lumH= getIdentityVec();
 541                         src_filter.lumV= getIdentityVec();
 542                 }
 543
 544                 if(sws_chr_gblur!=0.0){
 545                         src_filter.chrH= getGaussianVec(sws_chr_gblur, 3.0);
 546                         src_filter.chrV= getGaussianVec(sws_chr_gblur, 3.0);
 547                 }else{
 548                         src_filter.chrH= getIdentityVec();
 549                         src_filter.chrV= getIdentityVec();
 550                 }
 551
 552                 if(sws_chr_sharpen!=0.0){
 553                         SwsVector *g= getConstVec(-1.0, 3);
 554                         SwsVector *id= getConstVec(10.0/sws_chr_sharpen, 1);
 555                         g->coeff[1]=2.0;
 556                         addVec(id, g);
 557                         convVec(src_filter.chrH, id);
 558                         convVec(src_filter.chrV, id);
 559                         freeVec(g);
 560                         freeVec(id);
 561                 }
 562
 563                 if(sws_lum_sharpen!=0.0){
 564                         SwsVector *g= getConstVec(-1.0, 3);
 565                         SwsVector *id= getConstVec(10.0/sws_lum_sharpen, 1);
 566                         g->coeff[1]=2.0;
 567                         addVec(id, g);
 568                         convVec(src_filter.lumH, id);
 569                         convVec(src_filter.lumV, id);
 570                         freeVec(g);
 571                         freeVec(id);
 572                 }
 573
 574                 if(sws_chr_hshift)
 575                         shiftVec(src_filter.chrH, sws_chr_hshift);
 576
 577                 if(sws_chr_vshift)
 578                         shiftVec(src_filter.chrV, sws_chr_vshift);
 579
 580                 normalizeVec(src_filter.chrH, 1.0);
 581                 normalizeVec(src_filter.chrV, 1.0);
 582                 normalizeVec(src_filter.lumH, 1.0);
 583                 normalizeVec(src_filter.lumV, 1.0);
 584
 585                 if(verbose > 1) printVec(src_filter.chrH);
 586                 if(verbose > 1) printVec(src_filter.lumH);
 587         }
 588
 589         switch(dstbpp)
 590         {
 591                 case 8 : dstFormat= IMGFMT_Y8;          break;
 592                 case 12: dstFormat= IMGFMT_YV12;        break;
 593                 case 15: dstFormat= IMGFMT_BGR15;       break;
 594                 case 16: dstFormat= IMGFMT_BGR16;       break;
 595                 case 24: dstFormat= IMGFMT_BGR24;       break;
 596                 case 32: dstFormat= IMGFMT_BGR32;       break;
 597                 default: return;
 598         }
 599
 600         switch(sws_flags)
 601         {
 602                 case 0: flags|= SWS_FAST_BILINEAR; break;
 603                 case 1: flags|= SWS_BILINEAR; break;
 604                 case 2: flags|= SWS_BICUBIC; break;
 605                 case 3: flags|= SWS_X; break;
 606                 case 4: flags|= SWS_POINT; break;
 607                 default:flags|= SWS_BILINEAR; break;
 608         }
 609
 610         if(!context) context=getSwsContext(srcW, srcH, IMGFMT_YV12, dstW, dstH, dstFormat, flags, &src_filter, NULL);
 611
 612
 613         swScale(context, src, srcStride, srcSliceY, srcSliceH, dst, dstStride3);
 614 }
 615
 616 static inline void initFilter(int16_t **outFilter, int16_t **filterPos, int *outFilterSize, int xInc,
 617                               int srcW, int dstW, int filterAlign, int one, int flags,
 618                               SwsVector *srcFilter, SwsVector *dstFilter)
 619 {
 620         int i;
 621         int filterSize;
 622         int filter2Size;
 623         int minFilterSize;
 624         double *filter=NULL;
 625         double *filter2=NULL;
 626 #ifdef ARCH_X86
 627         if(gCpuCaps.hasMMX)
 628                 asm volatile("emms\n\t"::: "memory"); //FIXME this shouldnt be required but it IS (even for non mmx versions)
 629 #endif
 630
 631         *filterPos = (int16_t*)memalign(8, dstW*sizeof(int16_t));
 632         if(ABS(xInc - 0x10000) <10) // unscaled
 633         {
 634                 int i;
 635                 filterSize= 1;
 636                 filter= (double*)memalign(8, dstW*sizeof(double)*filterSize);
 637                 for(i=0; i<dstW*filterSize; i++) filter[i]=0;
 638
 639                 for(i=0; i<dstW; i++)
 640                 {
 641                         filter[i*filterSize]=1;
 642                         (*filterPos)[i]=i;
 643                 }
 644
 645         }
 646         else if(flags&SWS_POINT) // lame looking point sampling mode
 647         {
 648                 int i;
 649                 int xDstInSrc;
 650                 filterSize= 1;
 651                 filter= (double*)memalign(8, dstW*sizeof(double)*filterSize);
 652
 653                 xDstInSrc= xInc/2 - 0x8000;
 654                 for(i=0; i<dstW; i++)
 655                 {
 656                         int xx= (xDstInSrc>>16) - (filterSize>>1) + 1;
 657
 658                         (*filterPos)[i]= xx;
 659                         filter[i]= 1.0;
 660                         xDstInSrc+= xInc;
 661                 }
 662         }
 663         else if(xInc <= (1<<16) || (flags&SWS_FAST_BILINEAR)) // upscale
 664         {
 665                 int i;
 666                 int xDstInSrc;
 667                 if     (flags&SWS_BICUBIC) filterSize= 4;
 668                 else if(flags&SWS_X      ) filterSize= 4;
 669                 else                       filterSize= 2;
 670 //              printf("%d %d %d\n", filterSize, srcW, dstW);
 671                 filter= (double*)memalign(8, dstW*sizeof(double)*filterSize);
 672
 673                 xDstInSrc= xInc/2 - 0x8000;
 674                 for(i=0; i<dstW; i++)
 675                 {
 676                         int xx= (xDstInSrc>>16) - (filterSize>>1) + 1;
 677                         int j;
 678
 679                         (*filterPos)[i]= xx;
 680                         if((flags & SWS_BICUBIC) || (flags & SWS_X))
 681                         {
 682                                 double d= ABS(((xx+1)<<16) - xDstInSrc)/(double)(1<<16);
 683                                 double y1,y2,y3,y4;
 684                                 double A= -0.6;
 685                                 if(flags & SWS_BICUBIC){
 686                                                 // Equation is from VirtualDub
 687                                         y1 = (        +     A*d -       2.0*A*d*d +       A*d*d*d);
 688                                         y2 = (+ 1.0             -     (A+3.0)*d*d + (A+2.0)*d*d*d);
 689                                         y3 = (        -     A*d + (2.0*A+3.0)*d*d - (A+2.0)*d*d*d);
 690                                         y4 = (                  +           A*d*d -       A*d*d*d);
 691                                 }else{
 692                                                 // cubic interpolation (derived it myself)
 693                                         y1 = (    -2.0*d + 3.0*d*d - 1.0*d*d*d)/6.0;
 694                                         y2 = (6.0 -3.0*d - 6.0*d*d + 3.0*d*d*d)/6.0;
 695                                         y3 = (    +6.0*d + 3.0*d*d - 3.0*d*d*d)/6.0;
 696                                         y4 = (    -1.0*d           + 1.0*d*d*d)/6.0;
 697                                 }
 698
 699 //                              printf("%d %d %d \n", coeff, (int)d, xDstInSrc);
 700                                 filter[i*filterSize + 0]= y1;
 701                                 filter[i*filterSize + 1]= y2;
 702                                 filter[i*filterSize + 2]= y3;
 703                                 filter[i*filterSize + 3]= y4;
 704 //                              printf("%1.3f %1.3f %1.3f %1.3f %1.3f\n",d , y1, y2, y3, y4);
 705                         }
 706                         else
 707                         {
 708                                 for(j=0; j<filterSize; j++)
 709                                 {
 710                                         double d= ABS((xx<<16) - xDstInSrc)/(double)(1<<16);
 711                                         double coeff= 1.0 - d;
 712                                         if(coeff<0) coeff=0;
 713         //                              printf("%d %d %d \n", coeff, (int)d, xDstInSrc);
 714                                         filter[i*filterSize + j]= coeff;
 715                                         xx++;
 716                                 }
 717                         }
 718                         xDstInSrc+= xInc;
 719                 }
 720         }
 721         else // downscale
 722         {
 723                 int xDstInSrc;
 724                 if(flags&SWS_BICUBIC) filterSize= (int)ceil(1 + 4.0*srcW / (double)dstW);
 725                 else if(flags&SWS_X)  filterSize= (int)ceil(1 + 4.0*srcW / (double)dstW);
 726                 else                  filterSize= (int)ceil(1 + 2.0*srcW / (double)dstW);
 727 //              printf("%d %d %d\n", *filterSize, srcW, dstW);
 728                 filter= (double*)memalign(8, dstW*sizeof(double)*filterSize);
 729
 730                 xDstInSrc= xInc/2 - 0x8000;
 731                 for(i=0; i<dstW; i++)
 732                 {
 733                         int xx= (int)((double)xDstInSrc/(double)(1<<16) - (filterSize-1)*0.5 + 0.5);
 734                         int j;
 735                         (*filterPos)[i]= xx;
 736                         for(j=0; j<filterSize; j++)
 737                         {
 738                                 double d= ABS((xx<<16) - xDstInSrc)/(double)xInc;
 739                                 double coeff;
 740                                 if((flags & SWS_BICUBIC) || (flags & SWS_X))
 741                                 {
 742                                         double A= -0.75;
 743 //                                      d*=2;
 744                                         // Equation is from VirtualDub
 745                                         if(d<1.0)
 746                                                 coeff = (1.0 - (A+3.0)*d*d + (A+2.0)*d*d*d);
 747                                         else if(d<2.0)
 748                                                 coeff = (-4.0*A + 8.0*A*d - 5.0*A*d*d + A*d*d*d);
 749                                         else
 750                                                 coeff=0.0;
 751                                 }
 752 /*                              else if(flags & SWS_X)
 753                                 {
 754                                 }*/
 755                                 else
 756                                 {
 757                                         coeff= 1.0 - d;
 758                                         if(coeff<0) coeff=0;
 759                                 }
 760 //                              printf("%1.3f %d %d \n", coeff, (int)d, xDstInSrc);
 761                                 filter[i*filterSize + j]= coeff;
 762                                 xx++;
 763                         }
 764                         xDstInSrc+= xInc;
 765                 }
 766         }
 767
 768         /* apply src & dst Filter to filter -> filter2
 769            free(filter);
 770         */
 771         filter2Size= filterSize;
 772         if(srcFilter) filter2Size+= srcFilter->length - 1;
 773         if(dstFilter) filter2Size+= dstFilter->length - 1;
 774         filter2= (double*)memalign(8, filter2Size*dstW*sizeof(double));
 775
 776         for(i=0; i<dstW; i++)
 777         {
 778                 int j;
 779                 SwsVector scaleFilter;
 780                 SwsVector *outVec;
 781
 782                 scaleFilter.coeff= filter + i*filterSize;
 783                 scaleFilter.length= filterSize;
 784
 785                 if(srcFilter) outVec= getConvVec(srcFilter, &scaleFilter);
 786                 else          outVec= &scaleFilter;
 787
 788                 ASSERT(outVec->length == filter2Size)
 789                 //FIXME dstFilter
 790
 791                 for(j=0; j<outVec->length; j++)
 792                 {
 793                         filter2[i*filter2Size + j]= outVec->coeff[j];
 794                 }
 795
 796                 (*filterPos)[i]+= (filterSize-1)/2 - (filter2Size-1)/2;
 797
 798                 if(outVec != &scaleFilter) freeVec(outVec);
 799         }
 800         free(filter); filter=NULL;
 801
 802         /* try to reduce the filter-size (step1 find size and shift left) */
 803         // Assume its near normalized (*0.5 or *2.0 is ok but * 0.001 is not)
 804         minFilterSize= 0;
 805         for(i=dstW-1; i>=0; i--)
 806         {
 807                 int min= filter2Size;
 808                 int j;
 809                 double cutOff=0.0;
 810
 811                 /* get rid off near zero elements on the left by shifting left */
 812                 for(j=0; j<filter2Size; j++)
 813                 {
 814                         int k;
 815                         cutOff += ABS(filter2[i*filter2Size]);
 816
 817                         if(cutOff > SWS_MAX_REDUCE_CUTOFF) break;
 818
 819                         /* preserve Monotonicity because the core cant handle the filter otherwise */
 820                         if(i<dstW-1 && (*filterPos)[i] >= (*filterPos)[i+1]) break;
 821
 822                         // Move filter coeffs left
 823                         for(k=1; k<filter2Size; k++)
 824                                 filter2[i*filter2Size + k - 1]= filter2[i*filter2Size + k];
 825                         filter2[i*filter2Size + k - 1]= 0.0;
 826                         (*filterPos)[i]++;
 827                 }
 828
 829                 cutOff=0.0;
 830                 /* count near zeros on the right */
 831                 for(j=filter2Size-1; j>0; j--)
 832                 {
 833                         cutOff += ABS(filter2[i*filter2Size + j]);
 834
 835                         if(cutOff > SWS_MAX_REDUCE_CUTOFF) break;
 836                         min--;
 837                 }
 838
 839                 if(min>minFilterSize) minFilterSize= min;
 840         }
 841
 842         /* try to reduce the filter-size (step2 reduce it) */
 843         for(i=0; i<dstW; i++)
 844         {
 845                 int j;
 846
 847                 for(j=0; j<minFilterSize; j++)
 848                         filter2[i*minFilterSize + j]= filter2[i*filter2Size + j];
 849         }
 850         if((flags&SWS_PRINT_INFO) && verbose)
 851                 printf("SwScaler: reducing filtersize %d -> %d\n", filter2Size, minFilterSize);
 852         filter2Size= minFilterSize;
 853         ASSERT(filter2Size > 0)
 854
 855         //FIXME try to align filterpos if possible
 856
 857         //fix borders
 858         for(i=0; i<dstW; i++)
 859         {
 860                 int j;
 861                 if((*filterPos)[i] < 0)
 862                 {
 863                         // Move filter coeffs left to compensate for filterPos
 864                         for(j=1; j<filter2Size; j++)
 865                         {
 866                                 int left= MAX(j + (*filterPos)[i], 0);
 867                                 filter2[i*filter2Size + left] += filter2[i*filter2Size + j];
 868                                 filter2[i*filter2Size + j]=0;
 869                         }
 870                         (*filterPos)[i]= 0;
 871                 }
 872
 873                 if((*filterPos)[i] + filter2Size > srcW)
 874                 {
 875                         int shift= (*filterPos)[i] + filter2Size - srcW;
 876                         // Move filter coeffs right to compensate for filterPos
 877                         for(j=filter2Size-2; j>=0; j--)
 878                         {
 879                                 int right= MIN(j + shift, filter2Size-1);
 880                                 filter2[i*filter2Size +right] += filter2[i*filter2Size +j];
 881                                 filter2[i*filter2Size +j]=0;
 882                         }
 883                         (*filterPos)[i]= srcW - filter2Size;
 884                 }
 885         }
 886
 887
 888         *outFilterSize= (filter2Size +(filterAlign-1)) & (~(filterAlign-1));
 889         *outFilter= (int16_t*)memalign(8, *outFilterSize*dstW*sizeof(int16_t));
 890         memset(*outFilter, 0, *outFilterSize*dstW*sizeof(int16_t));
 891
 892         /* Normalize & Store in outFilter */
 893         for(i=0; i<dstW; i++)
 894         {
 895                 int j;
 896                 double sum=0;
 897                 double scale= one;
 898                 for(j=0; j<filter2Size; j++)
 899                 {
 900                         sum+= filter2[i*filter2Size + j];
 901                 }
 902                 scale/= sum;
 903                 for(j=0; j<filter2Size; j++)
 904                 {
 905                         (*outFilter)[i*(*outFilterSize) + j]= (int)(filter2[i*filter2Size + j]*scale);
 906                 }
 907         }
 908
 909         free(filter2);
 910 }
 911
 912 #ifdef ARCH_X86
 913 static void initMMX2HScaler(int dstW, int xInc, uint8_t *funnyCode)
 914 {
 915         uint8_t *fragment;
 916         int imm8OfPShufW1;
 917         int imm8OfPShufW2;
 918         int fragmentLength;
 919
 920         int xpos, i;
 921
 922         // create an optimized horizontal scaling routine
 923
 924         //code fragment
 925
 926         asm volatile(
 927                 "jmp 9f                         \n\t"
 928         // Begin
 929                 "0:                             \n\t"
 930                 "movq (%%esi), %%mm0            \n\t" //FIXME Alignment
 931                 "movq %%mm0, %%mm1              \n\t"
 932                 "psrlq $8, %%mm0                \n\t"
 933                 "punpcklbw %%mm7, %%mm1 \n\t"
 934                 "movq %%mm2, %%mm3              \n\t"
 935                 "punpcklbw %%mm7, %%mm0 \n\t"
 936                 "addw %%bx, %%cx                \n\t" //2*xalpha += (4*lumXInc)&0xFFFF
 937                 "pshufw $0xFF, %%mm1, %%mm1     \n\t"
 938                 "1:                             \n\t"
 939                 "adcl %%edx, %%esi              \n\t" //xx+= (4*lumXInc)>>16 + carry
 940                 "pshufw $0xFF, %%mm0, %%mm0     \n\t"
 941                 "2:                             \n\t"
 942                 "psrlw $9, %%mm3                \n\t"
 943                 "psubw %%mm1, %%mm0             \n\t"
 944                 "pmullw %%mm3, %%mm0            \n\t"
 945                 "paddw %%mm6, %%mm2             \n\t" // 2*alpha += xpos&0xFFFF
 946                 "psllw $7, %%mm1                \n\t"
 947                 "paddw %%mm1, %%mm0             \n\t"
 948
 949                 "movq %%mm0, (%%edi, %%eax)     \n\t"
 950
 951                 "addl $8, %%eax                 \n\t"
 952         // End
 953                 "9:                             \n\t"
 954 //              "int $3\n\t"
 955                 "leal 0b, %0                    \n\t"
 956                 "leal 1b, %1                    \n\t"
 957                 "leal 2b, %2                    \n\t"
 958                 "decl %1                        \n\t"
 959                 "decl %2                        \n\t"
 960                 "subl %0, %1                    \n\t"
 961                 "subl %0, %2                    \n\t"
 962                 "leal 9b, %3                    \n\t"
 963                 "subl %0, %3                    \n\t"
 964                 :"=r" (fragment), "=r" (imm8OfPShufW1), "=r" (imm8OfPShufW2),
 965                 "=r" (fragmentLength)
 966         );
 967
 968         xpos= 0; //lumXInc/2 - 0x8000; // difference between pixel centers
 969
 970         for(i=0; i<dstW/8; i++)
 971         {
 972                 int xx=xpos>>16;
 973
 974                 if((i&3) == 0)
 975                 {
 976                         int a=0;
 977                         int b=((xpos+xInc)>>16) - xx;
 978                         int c=((xpos+xInc*2)>>16) - xx;
 979                         int d=((xpos+xInc*3)>>16) - xx;
 980
 981                         memcpy(funnyCode + fragmentLength*i/4, fragment, fragmentLength);
 982
 983                         funnyCode[fragmentLength*i/4 + imm8OfPShufW1]=
 984                         funnyCode[fragmentLength*i/4 + imm8OfPShufW2]=
 985                                 a | (b<<2) | (c<<4) | (d<<6);
 986
 987                         // if we dont need to read 8 bytes than dont :), reduces the chance of
 988                         // crossing a cache line
 989                         if(d<3) funnyCode[fragmentLength*i/4 + 1]= 0x6E;
 990
 991                         funnyCode[fragmentLength*(i+4)/4]= RET;
 992                 }
 993                 xpos+=xInc;
 994         }
 995 }
 996 #endif // ARCH_X86
 997
 998 //FIXME remove
 999 void SwScale_Init(){
1000 }
1001
1002 static void globalInit(){
1003     // generating tables:
1004     int i;
1005     for(i=0; i<768; i++){
1006         int c= MIN(MAX(i-256, 0), 255);
1007         clip_table[i]=c;
1008         yuvtab_2568[c]= clip_yuvtab_2568[i]=(0x2568*(c-16))+(256<<13);
1009         yuvtab_3343[c]= clip_yuvtab_3343[i]=0x3343*(c-128);
1010         yuvtab_0c92[c]= clip_yuvtab_0c92[i]=-0x0c92*(c-128);
1011         yuvtab_1a1e[c]= clip_yuvtab_1a1e[i]=-0x1a1e*(c-128);
1012         yuvtab_40cf[c]= clip_yuvtab_40cf[i]=0x40cf*(c-128);
1013     }
1014
1015     for(i=0; i<768; i++)
1016     {
1017         int v= clip_table[i];
1018         clip_table16b[i]= v>>3;
1019         clip_table16g[i]= (v<<3)&0x07E0;
1020         clip_table16r[i]= (v<<8)&0xF800;
1021         clip_table15b[i]= v>>3;
1022         clip_table15g[i]= (v<<2)&0x03E0;
1023         clip_table15r[i]= (v<<7)&0x7C00;
1024     }
1025
1026 cpuCaps= gCpuCaps;
1027
1028 #ifdef RUNTIME_CPUDETECT
1029 #ifdef CAN_COMPILE_X86_ASM
1030         // ordered per speed fasterst first
1031         if(gCpuCaps.hasMMX2)
1032                 swScale= swScale_MMX2;
1033         else if(gCpuCaps.has3DNow)
1034                 swScale= swScale_3DNow;
1035         else if(gCpuCaps.hasMMX)
1036                 swScale= swScale_MMX;
1037         else
1038                 swScale= swScale_C;
1039
1040 #else
1041         swScale= swScale_C;
1042         cpuCaps.hasMMX2 = cpuCaps.hasMMX = cpuCaps.has3DNow = 0;
1043 #endif
1044 #else //RUNTIME_CPUDETECT
1045 #ifdef HAVE_MMX2
1046         swScale= swScale_MMX2;
1047         cpuCaps.has3DNow = 0;
1048 #elif defined (HAVE_3DNOW)
1049         swScale= swScale_3DNow;
1050         cpuCaps.hasMMX2 = 0;
1051 #elif defined (HAVE_MMX)
1052         swScale= swScale_MMX;
1053         cpuCaps.hasMMX2 = cpuCaps.has3DNow = 0;
1054 #else
1055         swScale= swScale_C;
1056         cpuCaps.hasMMX2 = cpuCaps.hasMMX = cpuCaps.has3DNow = 0;
1057 #endif
1058 #endif //!RUNTIME_CPUDETECT
1059 }
1060
1061
1062 SwsContext *getSwsContext(int srcW, int srcH, int srcFormat, int dstW, int dstH, int dstFormat, int flags,
1063                          SwsFilter *srcFilter, SwsFilter *dstFilter){
1064
1065         const int widthAlign= dstFormat==IMGFMT_YV12 ? 16 : 8;
1066         SwsContext *c;
1067         int i;
1068         SwsFilter dummyFilter= {NULL, NULL, NULL, NULL};
1069
1070 #ifdef ARCH_X86
1071         if(gCpuCaps.hasMMX)
1072                 asm volatile("emms\n\t"::: "memory");
1073 #endif
1074
1075         if(swScale==NULL) globalInit();
1076
1077         /* sanity check */
1078         if(srcW<1 || srcH<1 || dstW<1 || dstH<1) return NULL;
1079
1080 /* FIXME
1081         if(dstStride[0]%widthAlign !=0 )
1082         {
1083                 if(flags & SWS_PRINT_INFO)
1084                         fprintf(stderr, "SwScaler: Warning: dstStride is not a multiple of %d!\n"
1085                                         "SwScaler:          ->cannot do aligned memory acesses anymore\n",
1086                                         widthAlign);
1087         }
1088 */
1089         if(!dstFilter) dstFilter= &dummyFilter;
1090         if(!srcFilter) srcFilter= &dummyFilter;
1091
1092         c= memalign(64, sizeof(SwsContext));
1093         memset(c, 0, sizeof(SwsContext));
1094
1095         c->srcW= srcW;
1096         c->srcH= srcH;
1097         c->dstW= dstW;
1098         c->dstH= dstH;
1099         c->lumXInc= ((srcW<<16) + (dstW>>1))/dstW;
1100         c->lumYInc= ((srcH<<16) + (dstH>>1))/dstH;
1101         c->flags= flags;
1102         c->dstFormat= dstFormat;
1103         c->srcFormat= srcFormat;
1104
1105         if(cpuCaps.hasMMX2)
1106         {
1107                 c->canMMX2BeUsed= (dstW >=srcW && (dstW&31)==0 && (srcW&15)==0) ? 1 : 0;
1108                 if(!c->canMMX2BeUsed && dstW >=srcW && (srcW&15)==0 && (flags&SWS_FAST_BILINEAR))
1109                 {
1110                         if(flags&SWS_PRINT_INFO)
1111                                 fprintf(stderr, "SwScaler: output Width is not a multiple of 32 -> no MMX2 scaler\n");
1112                 }
1113         }
1114         else
1115                 c->canMMX2BeUsed=0;
1116
1117         // match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src to pixel n-2 of dst
1118         // but only for the FAST_BILINEAR mode otherwise do correct scaling
1119         // n-2 is the last chrominance sample available
1120         // this is not perfect, but noone shuld notice the difference, the more correct variant
1121         // would be like the vertical one, but that would require some special code for the
1122         // first and last pixel
1123         if(flags&SWS_FAST_BILINEAR)
1124         {
1125                 if(c->canMMX2BeUsed)    c->lumXInc+= 20;
1126                 //we dont use the x86asm scaler if mmx is available
1127                 else if(cpuCaps.hasMMX) c->lumXInc = ((srcW-2)<<16)/(dstW-2) - 20;
1128         }
1129
1130         /* set chrXInc & chrDstW */
1131         if((flags&SWS_FULL_UV_IPOL) && dstFormat!=IMGFMT_YV12)
1132                 c->chrXInc= c->lumXInc>>1, c->chrDstW= dstW;
1133         else
1134                 c->chrXInc= c->lumXInc,    c->chrDstW= (dstW+1)>>1;
1135
1136         /* set chrYInc & chrDstH */
1137         if(dstFormat==IMGFMT_YV12)      c->chrYInc= c->lumYInc,    c->chrDstH= (dstH+1)>>1;
1138         else                            c->chrYInc= c->lumYInc>>1, c->chrDstH= dstH;
1139
1140         /* precalculate horizontal scaler filter coefficients */
1141         {
1142                 const int filterAlign= cpuCaps.hasMMX ? 4 : 1;
1143
1144                 initFilter(&c->hLumFilter, &c->hLumFilterPos, &c->hLumFilterSize, c->lumXInc,
1145                                  srcW      ,       dstW, filterAlign, 1<<14, flags,
1146                                  srcFilter->lumH, dstFilter->lumH);
1147                 initFilter(&c->hChrFilter, &c->hChrFilterPos, &c->hChrFilterSize, c->chrXInc,
1148                                 (srcW+1)>>1, c->chrDstW, filterAlign, 1<<14, flags,
1149                                  srcFilter->chrH, dstFilter->chrH);
1150
1151 #ifdef ARCH_X86
1152 // cant downscale !!!
1153                 if(c->canMMX2BeUsed && (flags & SWS_FAST_BILINEAR))
1154                 {
1155                         initMMX2HScaler(      dstW, c->lumXInc, c->funnyYCode);
1156                         initMMX2HScaler(c->chrDstW, c->chrXInc, c->funnyUVCode);
1157                 }
1158 #endif
1159         } // Init Horizontal stuff
1160
1161
1162
1163         /* precalculate vertical scaler filter coefficients */
1164         initFilter(&c->vLumFilter, &c->vLumFilterPos, &c->vLumFilterSize, c->lumYInc,
1165                         srcH      ,        dstH, 1, (1<<12)-4, flags,
1166                         srcFilter->lumV, dstFilter->lumV);
1167         initFilter(&c->vChrFilter, &c->vChrFilterPos, &c->vChrFilterSize, c->chrYInc,
1168                         (srcH+1)>>1, c->chrDstH, 1, (1<<12)-4, flags,
1169                          srcFilter->chrV, dstFilter->chrV);
1170
1171         // Calculate Buffer Sizes so that they wont run out while handling these damn slices
1172         c->vLumBufSize= c->vLumFilterSize;
1173         c->vChrBufSize= c->vChrFilterSize;
1174         for(i=0; i<dstH; i++)
1175         {
1176                 int chrI= i*c->chrDstH / dstH;
1177                 int nextSlice= MAX(c->vLumFilterPos[i   ] + c->vLumFilterSize - 1,
1178                                  ((c->vChrFilterPos[chrI] + c->vChrFilterSize - 1)<<1));
1179                 nextSlice&= ~1; // Slices start at even boundaries
1180                 if(c->vLumFilterPos[i   ] + c->vLumBufSize < nextSlice)
1181                         c->vLumBufSize= nextSlice - c->vLumFilterPos[i   ];
1182                 if(c->vChrFilterPos[chrI] + c->vChrBufSize < (nextSlice>>1))
1183                         c->vChrBufSize= (nextSlice>>1) - c->vChrFilterPos[chrI];
1184         }
1185
1186         // allocate pixbufs (we use dynamic allocation because otherwise we would need to
1187         // allocate several megabytes to handle all possible cases)
1188         c->lumPixBuf= (int16_t**)memalign(4, c->vLumBufSize*2*sizeof(int16_t*));
1189         c->chrPixBuf= (int16_t**)memalign(4, c->vChrBufSize*2*sizeof(int16_t*));
1190         for(i=0; i<c->vLumBufSize; i++)
1191                 c->lumPixBuf[i]= c->lumPixBuf[i+c->vLumBufSize]= (uint16_t*)memalign(8, 4000);
1192         for(i=0; i<c->vChrBufSize; i++)
1193                 c->chrPixBuf[i]= c->chrPixBuf[i+c->vChrBufSize]= (uint16_t*)memalign(8, 8000);
1194
1195         //try to avoid drawing green stuff between the right end and the stride end
1196         for(i=0; i<c->vLumBufSize; i++) memset(c->lumPixBuf[i], 0, 4000);
1197         for(i=0; i<c->vChrBufSize; i++) memset(c->chrPixBuf[i], 64, 8000);
1198
1199         ASSERT(c->chrDstH <= dstH)
1200
1201         // pack filter data for mmx code
1202         if(cpuCaps.hasMMX)
1203         {
1204                 c->lumMmxFilter= (int16_t*)memalign(8, c->vLumFilterSize*      dstH*4*sizeof(int16_t));
1205                 c->chrMmxFilter= (int16_t*)memalign(8, c->vChrFilterSize*c->chrDstH*4*sizeof(int16_t));
1206                 for(i=0; i<c->vLumFilterSize*dstH; i++)
1207                         c->lumMmxFilter[4*i]=c->lumMmxFilter[4*i+1]=c->lumMmxFilter[4*i+2]=c->lumMmxFilter[4*i+3]=
1208                                 c->vLumFilter[i];
1209                 for(i=0; i<c->vChrFilterSize*c->chrDstH; i++)
1210                         c->chrMmxFilter[4*i]=c->chrMmxFilter[4*i+1]=c->chrMmxFilter[4*i+2]=c->chrMmxFilter[4*i+3]=
1211                                 c->vChrFilter[i];
1212         }
1213
1214         if(flags&SWS_PRINT_INFO)
1215         {
1216 #ifdef DITHER1XBPP
1217                 char *dither= " dithered";
1218 #else
1219                 char *dither= "";
1220 #endif
1221                 if(flags&SWS_FAST_BILINEAR)
1222                         fprintf(stderr, "\nSwScaler: FAST_BILINEAR scaler ");
1223                 else if(flags&SWS_BILINEAR)
1224                         fprintf(stderr, "\nSwScaler: BILINEAR scaler ");
1225                 else if(flags&SWS_BICUBIC)
1226                         fprintf(stderr, "\nSwScaler: BICUBIC scaler ");
1227                 else if(flags&SWS_POINT)
1228                         fprintf(stderr, "\nSwScaler: POINT scaler ");
1229                 else
1230                         fprintf(stderr, "\nSwScaler: ehh flags invalid?! ");
1231
1232                 if(dstFormat==IMGFMT_BGR15)
1233                         fprintf(stderr, "with%s BGR15 output ", dither);
1234                 else if(dstFormat==IMGFMT_BGR16)
1235                         fprintf(stderr, "with%s BGR16 output ", dither);
1236                 else if(dstFormat==IMGFMT_BGR24)
1237                         fprintf(stderr, "with BGR24 output ");
1238                 else if(dstFormat==IMGFMT_BGR32)
1239                         fprintf(stderr, "with BGR32 output ");
1240                 else if(dstFormat==IMGFMT_YV12)
1241                         fprintf(stderr, "with YV12 output ");
1242                 else
1243                         fprintf(stderr, "without output ");
1244
1245                 if(cpuCaps.hasMMX2)
1246                         fprintf(stderr, "using MMX2\n");
1247                 else if(cpuCaps.has3DNow)
1248                         fprintf(stderr, "using 3DNOW\n");
1249                 else if(cpuCaps.hasMMX)
1250                         fprintf(stderr, "using MMX\n");
1251                 else
1252                         fprintf(stderr, "using C\n");
1253         }
1254
1255         if((flags & SWS_PRINT_INFO) && verbose)
1256         {
1257                 if(cpuCaps.hasMMX)
1258                 {
1259                         if(c->canMMX2BeUsed && (flags&SWS_FAST_BILINEAR))
1260                                 printf("SwScaler: using FAST_BILINEAR MMX2 scaler for horizontal scaling\n");
1261                         else
1262                         {
1263                                 if(c->hLumFilterSize==4)
1264                                         printf("SwScaler: using 4-tap MMX scaler for horizontal luminance scaling\n");
1265                                 else if(c->hLumFilterSize==8)
1266                                         printf("SwScaler: using 8-tap MMX scaler for horizontal luminance scaling\n");
1267                                 else
1268                                         printf("SwScaler: using n-tap MMX scaler for horizontal luminance scaling\n");
1269
1270                                 if(c->hChrFilterSize==4)
1271                                         printf("SwScaler: using 4-tap MMX scaler for horizontal chrominance scaling\n");
1272                                 else if(c->hChrFilterSize==8)
1273                                         printf("SwScaler: using 8-tap MMX scaler for horizontal chrominance scaling\n");
1274                                 else
1275                                         printf("SwScaler: using n-tap MMX scaler for horizontal chrominance scaling\n");
1276                         }
1277                 }
1278                 else
1279                 {
1280 #ifdef ARCH_X86
1281                         printf("SwScaler: using X86-Asm scaler for horizontal scaling\n");
1282 #else
1283                         if(flags & SWS_FAST_BILINEAR)
1284                                 printf("SwScaler: using FAST_BILINEAR C scaler for horizontal scaling\n");
1285                         else
1286                                 printf("SwScaler: using C scaler for horizontal scaling\n");
1287 #endif
1288                 }
1289                 if(dstFormat==IMGFMT_YV12)
1290                 {
1291                         if(c->vLumFilterSize==1)
1292                                 printf("SwScaler: using 1-tap %s \"scaler\" for vertical scaling (YV12)\n", cpuCaps.hasMMX ? "MMX" : "C");
1293                         else
1294                                 printf("SwScaler: using n-tap %s scaler for vertical scaling (YV12)\n", cpuCaps.hasMMX ? "MMX" : "C");
1295                 }
1296                 else
1297                 {
1298                         if(c->vLumFilterSize==1 && c->vChrFilterSize==2)
1299                                 printf("SwScaler: using 1-tap %s \"scaler\" for vertical luminance scaling (BGR)\n"
1300                                        "SwScaler:       2-tap scaler for vertical chrominance scaling (BGR)\n",cpuCaps.hasMMX ? "MMX" : "C");
1301                         else if(c->vLumFilterSize==2 && c->vChrFilterSize==2)
1302                                 printf("SwScaler: using 2-tap linear %s scaler for vertical scaling (BGR)\n", cpuCaps.hasMMX ? "MMX" : "C");
1303                         else
1304                                 printf("SwScaler: using n-tap %s scaler for vertical scaling (BGR)\n", cpuCaps.hasMMX ? "MMX" : "C");
1305                 }
1306
1307                 if(dstFormat==IMGFMT_BGR24)
1308                         printf("SwScaler: using %s YV12->BGR24 Converter\n",
1309                                 cpuCaps.hasMMX2 ? "MMX2" : (cpuCaps.hasMMX ? "MMX" : "C"));
1310                 else if(dstFormat==IMGFMT_BGR32)
1311                         printf("SwScaler: using %s YV12->BGR32 Converter\n", cpuCaps.hasMMX ? "MMX" : "C");
1312                 else if(dstFormat==IMGFMT_BGR16)
1313                         printf("SwScaler: using %s YV12->BGR16 Converter\n", cpuCaps.hasMMX ? "MMX" : "C");
1314                 else if(dstFormat==IMGFMT_BGR15)
1315                         printf("SwScaler: using %s YV12->BGR15 Converter\n", cpuCaps.hasMMX ? "MMX" : "C");
1316
1317                 printf("SwScaler: %dx%d -> %dx%d\n", srcW, srcH, dstW, dstH);
1318         }
1319
1320         return c;
1321 }
1322
1323 /**
1324  * returns a normalized gaussian curve used to filter stuff
1325  * quality=3 is high quality, lowwer is lowwer quality
1326  */
1327
1328 SwsVector *getGaussianVec(double variance, double quality){
1329         const int length= (int)(variance*quality + 0.5) | 1;
1330         int i;
1331         double *coeff= memalign(sizeof(double), length*sizeof(double));
1332         double middle= (length-1)*0.5;
1333         SwsVector *vec= malloc(sizeof(SwsVector));
1334
1335         vec->coeff= coeff;
1336         vec->length= length;
1337
1338         for(i=0; i<length; i++)
1339         {
1340                 double dist= i-middle;
1341                 coeff[i]= exp( -dist*dist/(2*variance*variance) ) / sqrt(2*variance*PI);
1342         }
1343
1344         normalizeVec(vec, 1.0);
1345
1346         return vec;
1347 }
1348
1349 SwsVector *getConstVec(double c, int length){
1350         int i;
1351         double *coeff= memalign(sizeof(double), length*sizeof(double));
1352         SwsVector *vec= malloc(sizeof(SwsVector));
1353
1354         vec->coeff= coeff;
1355         vec->length= length;
1356
1357         for(i=0; i<length; i++)
1358                 coeff[i]= c;
1359
1360         return vec;
1361 }
1362
1363
1364 SwsVector *getIdentityVec(void){
1365         double *coeff= memalign(sizeof(double), sizeof(double));
1366         SwsVector *vec= malloc(sizeof(SwsVector));
1367         coeff[0]= 1.0;
1368
1369         vec->coeff= coeff;
1370         vec->length= 1;
1371
1372         return vec;
1373 }
1374
1375 void normalizeVec(SwsVector *a, double height){
1376         int i;
1377         double sum=0;
1378         double inv;
1379
1380         for(i=0; i<a->length; i++)
1381                 sum+= a->coeff[i];
1382
1383         inv= height/sum;
1384
1385         for(i=0; i<a->length; i++)
1386                 a->coeff[i]*= height;
1387 }
1388
1389 void scaleVec(SwsVector *a, double scalar){
1390         int i;
1391
1392         for(i=0; i<a->length; i++)
1393                 a->coeff[i]*= scalar;
1394 }
1395
1396 static SwsVector *getConvVec(SwsVector *a, SwsVector *b){
1397         int length= a->length + b->length - 1;
1398         double *coeff= memalign(sizeof(double), length*sizeof(double));
1399         int i, j;
1400         SwsVector *vec= malloc(sizeof(SwsVector));
1401
1402         vec->coeff= coeff;
1403         vec->length= length;
1404
1405         for(i=0; i<length; i++) coeff[i]= 0.0;
1406
1407         for(i=0; i<a->length; i++)
1408         {
1409                 for(j=0; j<b->length; j++)
1410                 {
1411                         coeff[i+j]+= a->coeff[i]*b->coeff[j];
1412                 }
1413         }
1414
1415         return vec;
1416 }
1417
1418 static SwsVector *sumVec(SwsVector *a, SwsVector *b){
1419         int length= MAX(a->length, b->length);
1420         double *coeff= memalign(sizeof(double), length*sizeof(double));
1421         int i;
1422         SwsVector *vec= malloc(sizeof(SwsVector));
1423
1424         vec->coeff= coeff;
1425         vec->length= length;
1426
1427         for(i=0; i<length; i++) coeff[i]= 0.0;
1428
1429         for(i=0; i<a->length; i++) coeff[i + (length-1)/2 - (a->length-1)/2]+= a->coeff[i];
1430         for(i=0; i<b->length; i++) coeff[i + (length-1)/2 - (b->length-1)/2]+= b->coeff[i];
1431
1432         return vec;
1433 }
1434
1435 static SwsVector *diffVec(SwsVector *a, SwsVector *b){
1436         int length= MAX(a->length, b->length);
1437         double *coeff= memalign(sizeof(double), length*sizeof(double));
1438         int i;
1439         SwsVector *vec= malloc(sizeof(SwsVector));
1440
1441         vec->coeff= coeff;
1442         vec->length= length;
1443
1444         for(i=0; i<length; i++) coeff[i]= 0.0;
1445
1446         for(i=0; i<a->length; i++) coeff[i + (length-1)/2 - (a->length-1)/2]+= a->coeff[i];
1447         for(i=0; i<b->length; i++) coeff[i + (length-1)/2 - (b->length-1)/2]-= b->coeff[i];
1448
1449         return vec;
1450 }
1451
1452 /* shift left / or right if "shift" is negative */
1453 static SwsVector *getShiftedVec(SwsVector *a, int shift){
1454         int length= a->length + ABS(shift)*2;
1455         double *coeff= memalign(sizeof(double), length*sizeof(double));
1456         int i;
1457         SwsVector *vec= malloc(sizeof(SwsVector));
1458
1459         vec->coeff= coeff;
1460         vec->length= length;
1461
1462         for(i=0; i<length; i++) coeff[i]= 0.0;
1463
1464         for(i=0; i<a->length; i++)
1465         {
1466                 coeff[i + (length-1)/2 - (a->length-1)/2 - shift]= a->coeff[i];
1467         }
1468
1469         return vec;
1470 }
1471
1472 void shiftVec(SwsVector *a, int shift){
1473         SwsVector *shifted= getShiftedVec(a, shift);
1474         free(a->coeff);
1475         a->coeff= shifted->coeff;
1476         a->length= shifted->length;
1477         free(shifted);
1478 }
1479
1480 void addVec(SwsVector *a, SwsVector *b){
1481         SwsVector *sum= sumVec(a, b);
1482         free(a->coeff);
1483         a->coeff= sum->coeff;
1484         a->length= sum->length;
1485         free(sum);
1486 }
1487
1488 void subVec(SwsVector *a, SwsVector *b){
1489         SwsVector *diff= diffVec(a, b);
1490         free(a->coeff);
1491         a->coeff= diff->coeff;
1492         a->length= diff->length;
1493         free(diff);
1494 }
1495
1496 void convVec(SwsVector *a, SwsVector *b){
1497         SwsVector *conv= getConvVec(a, b);
1498         free(a->coeff);
1499         a->coeff= conv->coeff;
1500         a->length= conv->length;
1501         free(conv);
1502 }
1503
1504 SwsVector *cloneVec(SwsVector *a){
1505         double *coeff= memalign(sizeof(double), a->length*sizeof(double));
1506         int i;
1507         SwsVector *vec= malloc(sizeof(SwsVector));
1508
1509         vec->coeff= coeff;
1510         vec->length= a->length;
1511
1512         for(i=0; i<a->length; i++) coeff[i]= a->coeff[i];
1513
1514         return vec;
1515 }
1516
1517 void printVec(SwsVector *a){
1518         int i;
1519         double max=0;
1520         double min=0;
1521         double range;
1522
1523         for(i=0; i<a->length; i++)
1524                 if(a->coeff[i]>max) max= a->coeff[i];
1525
1526         for(i=0; i<a->length; i++)
1527                 if(a->coeff[i]<min) min= a->coeff[i];
1528
1529         range= max - min;
1530
1531         for(i=0; i<a->length; i++)
1532         {
1533                 int x= (int)((a->coeff[i]-min)*60.0/range +0.5);
1534                 printf("%1.3f ", a->coeff[i]);
1535                 for(;x>0; x--) printf(" ");
1536                 printf("|\n");
1537         }
1538 }
1539
1540 void freeVec(SwsVector *a){
1541         if(!a) return;
1542         if(a->coeff) free(a->coeff);
1543         a->coeff=NULL;
1544         a->length=0;
1545         free(a);
1546 }
1547
1548 void freeSwsContext(SwsContext *c){
1549         int i;
1550
1551         if(!c) return;
1552
1553         if(c->lumPixBuf)
1554         {
1555                 for(i=0; i<c->vLumBufSize*2; i++)
1556                 {
1557                         if(c->lumPixBuf[i]) free(c->lumPixBuf[i]);
1558                         c->lumPixBuf[i]=NULL;
1559                 }
1560                 free(c->lumPixBuf);
1561                 c->lumPixBuf=NULL;
1562         }
1563
1564         if(c->chrPixBuf)
1565         {
1566                 for(i=0; i<c->vChrBufSize*2; i++)
1567                 {
1568                         if(c->chrPixBuf[i]) free(c->chrPixBuf[i]);
1569                         c->chrPixBuf[i]=NULL;
1570                 }
1571                 free(c->chrPixBuf);
1572                 c->chrPixBuf=NULL;
1573         }
1574
1575         if(c->vLumFilter) free(c->vLumFilter);
1576         c->vLumFilter = NULL;
1577         if(c->vChrFilter) free(c->vChrFilter);
1578         c->vChrFilter = NULL;
1579         if(c->hLumFilter) free(c->hLumFilter);
1580         c->hLumFilter = NULL;
1581         if(c->hChrFilter) free(c->hChrFilter);
1582         c->hChrFilter = NULL;
1583
1584         if(c->vLumFilterPos) free(c->vLumFilterPos);
1585         c->vLumFilterPos = NULL;
1586         if(c->vChrFilterPos) free(c->vChrFilterPos);
1587         c->vChrFilterPos = NULL;
1588         if(c->hLumFilterPos) free(c->hLumFilterPos);
1589         c->hLumFilterPos = NULL;
1590         if(c->hChrFilterPos) free(c->hChrFilterPos);
1591         c->hChrFilterPos = NULL;
1592
1593         if(c->lumMmxFilter) free(c->lumMmxFilter);
1594         c->lumMmxFilter = NULL;
1595         if(c->chrMmxFilter) free(c->chrMmxFilter);
1596         c->chrMmxFilter = NULL;
1597
1598         free(c);
1599 }
1600
1601