scale2x/scale3x.c

   1 /*
   2  * This file is part of the Scale2x project.
   3  *
   4  * Copyright (C) 2001, 2002, 2003, 2004 Andrea Mazzoleni
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License as published by
   8  * the Free Software Foundation; either version 2 of the License, or
   9  * (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  * GNU General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 /*
  22  * This file contains a C and MMX implementation of the Scale2x effect.
  23  *
  24  * You can find an high level description of the effect at :
  25  *
  26  * http://scale2x.sourceforge.net/
  27  *
  28  * Alternatively at the previous license terms, you are allowed to use this
  29  * code in your program with these conditions:
  30  * - the program is not used in commercial activities.
  31  * - the whole source code of the program is released with the binary.
  32  * - derivative works of the program are allowed.
  33  */
  34
  35 #if HAVE_CONFIG_H
  36 #include <config.h>
  37 #endif
  38
  39 #include "scale3x.h"
  40
  41 #include <assert.h>
  42
  43 /***************************************************************************/
  44 /* Scale3x C implementation */
  45
  46 /**
  47  * Define the macro USE_SCALE_RANDOMWRITE to enable
  48  * an optimized version which writes memory in random order.
  49  * This version is a little faster if you write in system memory.
  50  * But it's a lot slower if you write in video memory.
  51  * So, enable it only if you are sure to never write directly in video memory.
  52  */
  53 /* #define USE_SCALE_RANDOMWRITE */
  54
  55 #ifdef USE_SCALE_RANDOMWRITE
  56
  57 static inline void scale3x_8_def_whole(scale3x_uint8* restrict dst0, scale3x_uint8* restrict dst1, scale3x_uint8* restrict dst2, const scale3x_uint8* restrict src0, const scale3x_uint8* restrict src1, const scale3x_uint8* restrict src2, unsigned count)
  58 {
  59         assert(count >= 2);
  60
  61         /* first pixel */
  62         if (src0[0] != src2[0] && src1[0] != src1[1]) {
  63                 dst0[0] = src1[0];
  64                 dst0[1] = (src1[0] == src0[0] && src1[0] != src0[1]) || (src1[1] == src0[0] && src1[0] != src0[0]) ? src0[0] : src1[0];
  65                 dst0[2] = src1[1] == src0[0] ? src1[1] : src1[0];
  66                 dst1[0] = (src1[0] == src0[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src0[0]) ? src1[0] : src1[0];
  67                 dst1[1] = src1[0];
  68                 dst1[2] = (src1[1] == src0[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src0[1]) ? src1[1] : src1[0];
  69                 dst2[0] = src1[0];
  70                 dst2[1] = (src1[0] == src2[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src2[0]) ? src2[0] : src1[0];
  71                 dst2[2] = src1[1] == src2[0] ? src1[1] : src1[0];
  72         } else {
  73                 dst0[0] = src1[0];
  74                 dst0[1] = src1[0];
  75                 dst0[2] = src1[0];
  76                 dst1[0] = src1[0];
  77                 dst1[1] = src1[0];
  78                 dst1[2] = src1[0];
  79                 dst2[0] = src1[0];
  80                 dst2[1] = src1[0];
  81                 dst2[2] = src1[0];
  82         }
  83         ++src0;
  84         ++src1;
  85         ++src2;
  86         dst0 += 3;
  87         dst1 += 3;
  88         dst2 += 3;
  89
  90         /* central pixels */
  91         count -= 2;
  92         while (count) {
  93                 if (src0[0] != src2[0] && src1[-1] != src1[1]) {
  94                         dst0[0] = src1[-1] == src0[0] ? src1[-1] : src1[0];
  95                         dst0[1] = (src1[-1] == src0[0] && src1[0] != src0[1]) || (src1[1] == src0[0] && src1[0] != src0[-1]) ? src0[0] : src1[0];
  96                         dst0[2] = src1[1] == src0[0] ? src1[1] : src1[0];
  97                         dst1[0] = (src1[-1] == src0[0] && src1[0] != src2[-1]) || (src1[-1] == src2[0] && src1[0] != src0[-1]) ? src1[-1] : src1[0];
  98                         dst1[1] = src1[0];
  99                         dst1[2] = (src1[1] == src0[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src0[1]) ? src1[1] : src1[0];
 100                         dst2[0] = src1[-1] == src2[0] ? src1[-1] : src1[0];
 101                         dst2[1] = (src1[-1] == src2[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src2[-1]) ? src2[0] : src1[0];
 102                         dst2[2] = src1[1] == src2[0] ? src1[1] : src1[0];
 103                 } else {
 104                         dst0[0] = src1[0];
 105                         dst0[1] = src1[0];
 106                         dst0[2] = src1[0];
 107                         dst1[0] = src1[0];
 108                         dst1[1] = src1[0];
 109                         dst1[2] = src1[0];
 110                         dst2[0] = src1[0];
 111                         dst2[1] = src1[0];
 112                         dst2[2] = src1[0];
 113                 }
 114
 115                 ++src0;
 116                 ++src1;
 117                 ++src2;
 118                 dst0 += 3;
 119                 dst1 += 3;
 120                 dst2 += 3;
 121                 --count;
 122         }
 123
 124         /* last pixel */
 125         if (src0[0] != src2[0] && src1[-1] != src1[0]) {
 126                 dst0[0] = src1[-1] == src0[0] ? src1[-1] : src1[0];
 127                 dst0[1] = (src1[-1] == src0[0] && src1[0] != src0[0]) || (src1[0] == src0[0] && src1[0] != src0[-1]) ? src0[0] : src1[0];
 128                 dst0[2] = src1[0];
 129                 dst1[0] = (src1[-1] == src0[0] && src1[0] != src2[-1]) || (src1[-1] == src2[0] && src1[0] != src0[-1]) ? src1[-1] : src1[0];
 130                 dst1[1] = src1[0];
 131                 dst1[2] = (src1[0] == src0[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src0[0]) ? src1[0] : src1[0];
 132                 dst2[0] = src1[-1] == src2[0] ? src1[-1] : src1[0];
 133                 dst2[1] = (src1[-1] == src2[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src2[-1]) ? src2[0] : src1[0];
 134                 dst2[2] = src1[0];
 135         } else {
 136                 dst0[0] = src1[0];
 137                 dst0[1] = src1[0];
 138                 dst0[2] = src1[0];
 139                 dst1[0] = src1[0];
 140                 dst1[1] = src1[0];
 141                 dst1[2] = src1[0];
 142                 dst2[0] = src1[0];
 143                 dst2[1] = src1[0];
 144                 dst2[2] = src1[0];
 145         }
 146 }
 147
 148 #else /* USE_SCALE_RANDOMWRITE */
 149
 150 static inline void scale3x_8_def_border(scale3x_uint8* restrict dst, const scale3x_uint8* restrict src0, const scale3x_uint8* restrict src1, const scale3x_uint8* restrict src2, unsigned count)
 151 {
 152         assert(count >= 2);
 153
 154         /* first pixel */
 155         if (src0[0] != src2[0] && src1[0] != src1[1]) {
 156                 dst[0] = src1[0];
 157                 dst[1] = (src1[0] == src0[0] && src1[0] != src0[1]) || (src1[1] == src0[0] && src1[0] != src0[0]) ? src0[0] : src1[0];
 158                 dst[2] = src1[1] == src0[0] ? src1[1] : src1[0];
 159         } else {
 160                 dst[0] = src1[0];
 161                 dst[1] = src1[0];
 162                 dst[2] = src1[0];
 163         }
 164         ++src0;
 165         ++src1;
 166         ++src2;
 167         dst += 3;
 168
 169         /* central pixels */
 170         count -= 2;
 171         while (count) {
 172                 if (src0[0] != src2[0] && src1[-1] != src1[1]) {
 173                         dst[0] = src1[-1] == src0[0] ? src1[-1] : src1[0];
 174                         dst[1] = (src1[-1] == src0[0] && src1[0] != src0[1]) || (src1[1] == src0[0] && src1[0] != src0[-1]) ? src0[0] : src1[0];
 175                         dst[2] = src1[1] == src0[0] ? src1[1] : src1[0];
 176                 } else {
 177                         dst[0] = src1[0];
 178                         dst[1] = src1[0];
 179                         dst[2] = src1[0];
 180                 }
 181
 182                 ++src0;
 183                 ++src1;
 184                 ++src2;
 185                 dst += 3;
 186                 --count;
 187         }
 188
 189         /* last pixel */
 190         if (src0[0] != src2[0] && src1[-1] != src1[0]) {
 191                 dst[0] = src1[-1] == src0[0] ? src1[-1] : src1[0];
 192                 dst[1] = (src1[-1] == src0[0] && src1[0] != src0[0]) || (src1[0] == src0[0] && src1[0] != src0[-1]) ? src0[0] : src1[0];
 193                 dst[2] = src1[0];
 194         } else {
 195                 dst[0] = src1[0];
 196                 dst[1] = src1[0];
 197                 dst[2] = src1[0];
 198         }
 199 }
 200
 201 static inline void scale3x_8_def_center(scale3x_uint8* restrict dst, const scale3x_uint8* restrict src0, const scale3x_uint8* restrict src1, const scale3x_uint8* restrict src2, unsigned count)
 202 {
 203         assert(count >= 2);
 204
 205         /* first pixel */
 206         if (src0[0] != src2[0] && src1[0] != src1[1]) {
 207                 dst[0] = (src1[0] == src0[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src0[0]) ? src1[0] : src1[0];
 208                 dst[1] = src1[0];
 209                 dst[2] = (src1[1] == src0[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src0[1]) ? src1[1] : src1[0];
 210         } else {
 211                 dst[0] = src1[0];
 212                 dst[1] = src1[0];
 213                 dst[2] = src1[0];
 214         }
 215         ++src0;
 216         ++src1;
 217         ++src2;
 218         dst += 3;
 219
 220         /* central pixels */
 221         count -= 2;
 222         while (count) {
 223                 if (src0[0] != src2[0] && src1[-1] != src1[1]) {
 224                         dst[0] = (src1[-1] == src0[0] && src1[0] != src2[-1]) || (src1[-1] == src2[0] && src1[0] != src0[-1]) ? src1[-1] : src1[0];
 225                         dst[1] = src1[0];
 226                         dst[2] = (src1[1] == src0[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src0[1]) ? src1[1] : src1[0];
 227                 } else {
 228                         dst[0] = src1[0];
 229                         dst[1] = src1[0];
 230                         dst[2] = src1[0];
 231                 }
 232
 233                 ++src0;
 234                 ++src1;
 235                 ++src2;
 236                 dst += 3;
 237                 --count;
 238         }
 239
 240         /* last pixel */
 241         if (src0[0] != src2[0] && src1[-1] != src1[0]) {
 242                 dst[0] = (src1[-1] == src0[0] && src1[0] != src2[-1]) || (src1[-1] == src2[0] && src1[0] != src0[-1]) ? src1[-1] : src1[0];
 243                 dst[1] = src1[0];
 244                 dst[2] = (src1[0] == src0[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src0[0]) ? src1[0] : src1[0];
 245         } else {
 246                 dst[0] = src1[0];
 247                 dst[1] = src1[0];
 248                 dst[2] = src1[0];
 249         }
 250 }
 251
 252 #endif /* USE_SCALE_RANDOMWRITE */
 253
 254 #ifdef USE_SCALE_RANDOMWRITE
 255
 256 static inline void scale3x_16_def_whole(scale3x_uint16* restrict dst0, scale3x_uint16* restrict dst1, scale3x_uint16* restrict dst2, const scale3x_uint16* restrict src0, const scale3x_uint16* restrict src1, const scale3x_uint16* restrict src2, unsigned count)
 257 {
 258         assert(count >= 2);
 259
 260         /* first pixel */
 261         if (src0[0] != src2[0] && src1[0] != src1[1]) {
 262                 dst0[0] = src1[0];
 263                 dst0[1] = (src1[0] == src0[0] && src1[0] != src0[1]) || (src1[1] == src0[0] && src1[0] != src0[0]) ? src0[0] : src1[0];
 264                 dst0[2] = src1[1] == src0[0] ? src1[1] : src1[0];
 265                 dst1[0] = (src1[0] == src0[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src0[0]) ? src1[0] : src1[0];
 266                 dst1[1] = src1[0];
 267                 dst1[2] = (src1[1] == src0[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src0[1]) ? src1[1] : src1[0];
 268                 dst2[0] = src1[0];
 269                 dst2[1] = (src1[0] == src2[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src2[0]) ? src2[0] : src1[0];
 270                 dst2[2] = src1[1] == src2[0] ? src1[1] : src1[0];
 271         } else {
 272                 dst0[0] = src1[0];
 273                 dst0[1] = src1[0];
 274                 dst0[2] = src1[0];
 275                 dst1[0] = src1[0];
 276                 dst1[1] = src1[0];
 277                 dst1[2] = src1[0];
 278                 dst2[0] = src1[0];
 279                 dst2[1] = src1[0];
 280                 dst2[2] = src1[0];
 281         }
 282         ++src0;
 283         ++src1;
 284         ++src2;
 285         dst0 += 3;
 286         dst1 += 3;
 287         dst2 += 3;
 288
 289         /* central pixels */
 290         count -= 2;
 291         while (count) {
 292                 if (src0[0] != src2[0] && src1[-1] != src1[1]) {
 293                         dst0[0] = src1[-1] == src0[0] ? src1[-1] : src1[0];
 294                         dst0[1] = (src1[-1] == src0[0] && src1[0] != src0[1]) || (src1[1] == src0[0] && src1[0] != src0[-1]) ? src0[0] : src1[0];
 295                         dst0[2] = src1[1] == src0[0] ? src1[1] : src1[0];
 296                         dst1[0] = (src1[-1] == src0[0] && src1[0] != src2[-1]) || (src1[-1] == src2[0] && src1[0] != src0[-1]) ? src1[-1] : src1[0];
 297                         dst1[1] = src1[0];
 298                         dst1[2] = (src1[1] == src0[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src0[1]) ? src1[1] : src1[0];
 299                         dst2[0] = src1[-1] == src2[0] ? src1[-1] : src1[0];
 300                         dst2[1] = (src1[-1] == src2[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src2[-1]) ? src2[0] : src1[0];
 301                         dst2[2] = src1[1] == src2[0] ? src1[1] : src1[0];
 302                 } else {
 303                         dst0[0] = src1[0];
 304                         dst0[1] = src1[0];
 305                         dst0[2] = src1[0];
 306                         dst1[0] = src1[0];
 307                         dst1[1] = src1[0];
 308                         dst1[2] = src1[0];
 309                         dst2[0] = src1[0];
 310                         dst2[1] = src1[0];
 311                         dst2[2] = src1[0];
 312                 }
 313
 314                 ++src0;
 315                 ++src1;
 316                 ++src2;
 317                 dst0 += 3;
 318                 dst1 += 3;
 319                 dst2 += 3;
 320                 --count;
 321         }
 322
 323         /* last pixel */
 324         if (src0[0] != src2[0] && src1[-1] != src1[0]) {
 325                 dst0[0] = src1[-1] == src0[0] ? src1[-1] : src1[0];
 326                 dst0[1] = (src1[-1] == src0[0] && src1[0] != src0[0]) || (src1[0] == src0[0] && src1[0] != src0[-1]) ? src0[0] : src1[0];
 327                 dst0[2] = src1[0];
 328                 dst1[0] = (src1[-1] == src0[0] && src1[0] != src2[-1]) || (src1[-1] == src2[0] && src1[0] != src0[-1]) ? src1[-1] : src1[0];
 329                 dst1[1] = src1[0];
 330                 dst1[2] = (src1[0] == src0[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src0[0]) ? src1[0] : src1[0];
 331                 dst2[0] = src1[-1] == src2[0] ? src1[-1] : src1[0];
 332                 dst2[1] = (src1[-1] == src2[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src2[-1]) ? src2[0] : src1[0];
 333                 dst2[2] = src1[0];
 334         } else {
 335                 dst0[0] = src1[0];
 336                 dst0[1] = src1[0];
 337                 dst0[2] = src1[0];
 338                 dst1[0] = src1[0];
 339                 dst1[1] = src1[0];
 340                 dst1[2] = src1[0];
 341                 dst2[0] = src1[0];
 342                 dst2[1] = src1[0];
 343                 dst2[2] = src1[0];
 344         }
 345 }
 346
 347 #else /* USE_SCALE_RANDOMWRITE */
 348
 349 static inline void scale3x_16_def_border(scale3x_uint16* restrict dst, const scale3x_uint16* restrict src0, const scale3x_uint16* restrict src1, const scale3x_uint16* restrict src2, unsigned count)
 350 {
 351         assert(count >= 2);
 352
 353         /* first pixel */
 354         if (src0[0] != src2[0] && src1[0] != src1[1]) {
 355                 dst[0] = src1[0];
 356                 dst[1] = (src1[0] == src0[0] && src1[0] != src0[1]) || (src1[1] == src0[0] && src1[0] != src0[0]) ? src0[0] : src1[0];
 357                 dst[2] = src1[1] == src0[0] ? src1[1] : src1[0];
 358         } else {
 359                 dst[0] = src1[0];
 360                 dst[1] = src1[0];
 361                 dst[2] = src1[0];
 362         }
 363         ++src0;
 364         ++src1;
 365         ++src2;
 366         dst += 3;
 367
 368         /* central pixels */
 369         count -= 2;
 370         while (count) {
 371                 if (src0[0] != src2[0] && src1[-1] != src1[1]) {
 372                         dst[0] = src1[-1] == src0[0] ? src1[-1] : src1[0];
 373                         dst[1] = (src1[-1] == src0[0] && src1[0] != src0[1]) || (src1[1] == src0[0] && src1[0] != src0[-1]) ? src0[0] : src1[0];
 374                         dst[2] = src1[1] == src0[0] ? src1[1] : src1[0];
 375                 } else {
 376                         dst[0] = src1[0];
 377                         dst[1] = src1[0];
 378                         dst[2] = src1[0];
 379                 }
 380
 381                 ++src0;
 382                 ++src1;
 383                 ++src2;
 384                 dst += 3;
 385                 --count;
 386         }
 387
 388         /* last pixel */
 389         if (src0[0] != src2[0] && src1[-1] != src1[0]) {
 390                 dst[0] = src1[-1] == src0[0] ? src1[-1] : src1[0];
 391                 dst[1] = (src1[-1] == src0[0] && src1[0] != src0[0]) || (src1[0] == src0[0] && src1[0] != src0[-1]) ? src0[0] : src1[0];
 392                 dst[2] = src1[0];
 393         } else {
 394                 dst[0] = src1[0];
 395                 dst[1] = src1[0];
 396                 dst[2] = src1[0];
 397         }
 398 }
 399
 400 static inline void scale3x_16_def_center(scale3x_uint16* restrict dst, const scale3x_uint16* restrict src0, const scale3x_uint16* restrict src1, const scale3x_uint16* restrict src2, unsigned count)
 401 {
 402         assert(count >= 2);
 403
 404         /* first pixel */
 405         if (src0[0] != src2[0] && src1[0] != src1[1]) {
 406                 dst[0] = (src1[0] == src0[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src0[0]) ? src1[0] : src1[0];
 407                 dst[1] = src1[0];
 408                 dst[2] = (src1[1] == src0[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src0[1]) ? src1[1] : src1[0];
 409         } else {
 410                 dst[0] = src1[0];
 411                 dst[1] = src1[0];
 412                 dst[2] = src1[0];
 413         }
 414         ++src0;
 415         ++src1;
 416         ++src2;
 417         dst += 3;
 418
 419         /* central pixels */
 420         count -= 2;
 421         while (count) {
 422                 if (src0[0] != src2[0] && src1[-1] != src1[1]) {
 423                         dst[0] = (src1[-1] == src0[0] && src1[0] != src2[-1]) || (src1[-1] == src2[0] && src1[0] != src0[-1]) ? src1[-1] : src1[0];
 424                         dst[1] = src1[0];
 425                         dst[2] = (src1[1] == src0[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src0[1]) ? src1[1] : src1[0];
 426                 } else {
 427                         dst[0] = src1[0];
 428                         dst[1] = src1[0];
 429                         dst[2] = src1[0];
 430                 }
 431
 432                 ++src0;
 433                 ++src1;
 434                 ++src2;
 435                 dst += 3;
 436                 --count;
 437         }
 438
 439         /* last pixel */
 440         if (src0[0] != src2[0] && src1[-1] != src1[0]) {
 441                 dst[0] = (src1[-1] == src0[0] && src1[0] != src2[-1]) || (src1[-1] == src2[0] && src1[0] != src0[-1]) ? src1[-1] : src1[0];
 442                 dst[1] = src1[0];
 443                 dst[2] = (src1[0] == src0[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src0[0]) ? src1[0] : src1[0];
 444         } else {
 445                 dst[0] = src1[0];
 446                 dst[1] = src1[0];
 447                 dst[2] = src1[0];
 448         }
 449 }
 450
 451 #endif /* USE_SCALE_RANDOMWRITE */
 452
 453 #ifdef USE_SCALE_RANDOMWRITE
 454
 455 static inline void scale3x_32_def_whole(scale3x_uint32* restrict dst0, scale3x_uint32* restrict dst1, scale3x_uint32* restrict dst2, const scale3x_uint32* restrict src0, const scale3x_uint32* restrict src1, const scale3x_uint32* restrict src2, unsigned count)
 456 {
 457         assert(count >= 2);
 458
 459         /* first pixel */
 460         if (src0[0] != src2[0] && src1[0] != src1[1]) {
 461                 dst0[0] = src1[0];
 462                 dst0[1] = (src1[0] == src0[0] && src1[0] != src0[1]) || (src1[1] == src0[0] && src1[0] != src0[0]) ? src0[0] : src1[0];
 463                 dst0[2] = src1[1] == src0[0] ? src1[1] : src1[0];
 464                 dst1[0] = (src1[0] == src0[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src0[0]) ? src1[0] : src1[0];
 465                 dst1[1] = src1[0];
 466                 dst1[2] = (src1[1] == src0[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src0[1]) ? src1[1] : src1[0];
 467                 dst2[0] = src1[0];
 468                 dst2[1] = (src1[0] == src2[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src2[0]) ? src2[0] : src1[0];
 469                 dst2[2] = src1[1] == src2[0] ? src1[1] : src1[0];
 470         } else {
 471                 dst0[0] = src1[0];
 472                 dst0[1] = src1[0];
 473                 dst0[2] = src1[0];
 474                 dst1[0] = src1[0];
 475                 dst1[1] = src1[0];
 476                 dst1[2] = src1[0];
 477                 dst2[0] = src1[0];
 478                 dst2[1] = src1[0];
 479                 dst2[2] = src1[0];
 480         }
 481         ++src0;
 482         ++src1;
 483         ++src2;
 484         dst0 += 3;
 485         dst1 += 3;
 486         dst2 += 3;
 487
 488         /* central pixels */
 489         count -= 2;
 490         while (count) {
 491                 if (src0[0] != src2[0] && src1[-1] != src1[1]) {
 492                         dst0[0] = src1[-1] == src0[0] ? src1[-1] : src1[0];
 493                         dst0[1] = (src1[-1] == src0[0] && src1[0] != src0[1]) || (src1[1] == src0[0] && src1[0] != src0[-1]) ? src0[0] : src1[0];
 494                         dst0[2] = src1[1] == src0[0] ? src1[1] : src1[0];
 495                         dst1[0] = (src1[-1] == src0[0] && src1[0] != src2[-1]) || (src1[-1] == src2[0] && src1[0] != src0[-1]) ? src1[-1] : src1[0];
 496                         dst1[1] = src1[0];
 497                         dst1[2] = (src1[1] == src0[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src0[1]) ? src1[1] : src1[0];
 498                         dst2[0] = src1[-1] == src2[0] ? src1[-1] : src1[0];
 499                         dst2[1] = (src1[-1] == src2[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src2[-1]) ? src2[0] : src1[0];
 500                         dst2[2] = src1[1] == src2[0] ? src1[1] : src1[0];
 501                 } else {
 502                         dst0[0] = src1[0];
 503                         dst0[1] = src1[0];
 504                         dst0[2] = src1[0];
 505                         dst1[0] = src1[0];
 506                         dst1[1] = src1[0];
 507                         dst1[2] = src1[0];
 508                         dst2[0] = src1[0];
 509                         dst2[1] = src1[0];
 510                         dst2[2] = src1[0];
 511                 }
 512
 513                 ++src0;
 514                 ++src1;
 515                 ++src2;
 516                 dst0 += 3;
 517                 dst1 += 3;
 518                 dst2 += 3;
 519                 --count;
 520         }
 521
 522         /* last pixel */
 523         if (src0[0] != src2[0] && src1[-1] != src1[0]) {
 524                 dst0[0] = src1[-1] == src0[0] ? src1[-1] : src1[0];
 525                 dst0[1] = (src1[-1] == src0[0] && src1[0] != src0[0]) || (src1[0] == src0[0] && src1[0] != src0[-1]) ? src0[0] : src1[0];
 526                 dst0[2] = src1[0];
 527                 dst1[0] = (src1[-1] == src0[0] && src1[0] != src2[-1]) || (src1[-1] == src2[0] && src1[0] != src0[-1]) ? src1[-1] : src1[0];
 528                 dst1[1] = src1[0];
 529                 dst1[2] = (src1[0] == src0[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src0[0]) ? src1[0] : src1[0];
 530                 dst2[0] = src1[-1] == src2[0] ? src1[-1] : src1[0];
 531                 dst2[1] = (src1[-1] == src2[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src2[-1]) ? src2[0] : src1[0];
 532                 dst2[2] = src1[0];
 533         } else {
 534                 dst0[0] = src1[0];
 535                 dst0[1] = src1[0];
 536                 dst0[2] = src1[0];
 537                 dst1[0] = src1[0];
 538                 dst1[1] = src1[0];
 539                 dst1[2] = src1[0];
 540                 dst2[0] = src1[0];
 541                 dst2[1] = src1[0];
 542                 dst2[2] = src1[0];
 543         }
 544 }
 545
 546 #else /* USE_SCALE_RANDOMWRITE */
 547
 548 static inline void scale3x_32_def_border(scale3x_uint32* restrict dst, const scale3x_uint32* restrict src0, const scale3x_uint32* restrict src1, const scale3x_uint32* restrict src2, unsigned count)
 549 {
 550         assert(count >= 2);
 551
 552         /* first pixel */
 553         if (src0[0] != src2[0] && src1[0] != src1[1]) {
 554                 dst[0] = src1[0];
 555                 dst[1] = (src1[0] == src0[0] && src1[0] != src0[1]) || (src1[1] == src0[0] && src1[0] != src0[0]) ? src0[0] : src1[0];
 556                 dst[2] = src1[1] == src0[0] ? src1[1] : src1[0];
 557         } else {
 558                 dst[0] = src1[0];
 559                 dst[1] = src1[0];
 560                 dst[2] = src1[0];
 561         }
 562         ++src0;
 563         ++src1;
 564         ++src2;
 565         dst += 3;
 566
 567         /* central pixels */
 568         count -= 2;
 569         while (count) {
 570                 if (src0[0] != src2[0] && src1[-1] != src1[1]) {
 571                         dst[0] = src1[-1] == src0[0] ? src1[-1] : src1[0];
 572                         dst[1] = (src1[-1] == src0[0] && src1[0] != src0[1]) || (src1[1] == src0[0] && src1[0] != src0[-1]) ? src0[0] : src1[0];
 573                         dst[2] = src1[1] == src0[0] ? src1[1] : src1[0];
 574                 } else {
 575                         dst[0] = src1[0];
 576                         dst[1] = src1[0];
 577                         dst[2] = src1[0];
 578                 }
 579
 580                 ++src0;
 581                 ++src1;
 582                 ++src2;
 583                 dst += 3;
 584                 --count;
 585         }
 586
 587         /* last pixel */
 588         if (src0[0] != src2[0] && src1[-1] != src1[0]) {
 589                 dst[0] = src1[-1] == src0[0] ? src1[-1] : src1[0];
 590                 dst[1] = (src1[-1] == src0[0] && src1[0] != src0[0]) || (src1[0] == src0[0] && src1[0] != src0[-1]) ? src0[0] : src1[0];
 591                 dst[2] = src1[0];
 592         } else {
 593                 dst[0] = src1[0];
 594                 dst[1] = src1[0];
 595                 dst[2] = src1[0];
 596         }
 597 }
 598
 599 static inline void scale3x_32_def_center(scale3x_uint32* restrict dst, const scale3x_uint32* restrict src0, const scale3x_uint32* restrict src1, const scale3x_uint32* restrict src2, unsigned count)
 600 {
 601         assert(count >= 2);
 602
 603         /* first pixel */
 604         if (src0[0] != src2[0] && src1[0] != src1[1]) {
 605                 dst[0] = (src1[0] == src0[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src0[0]) ? src1[0] : src1[0];
 606                 dst[1] = src1[0];
 607                 dst[2] = (src1[1] == src0[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src0[1]) ? src1[1] : src1[0];
 608         } else {
 609                 dst[0] = src1[0];
 610                 dst[1] = src1[0];
 611                 dst[2] = src1[0];
 612         }
 613         ++src0;
 614         ++src1;
 615         ++src2;
 616         dst += 3;
 617
 618         /* central pixels */
 619         count -= 2;
 620         while (count) {
 621                 if (src0[0] != src2[0] && src1[-1] != src1[1]) {
 622                         dst[0] = (src1[-1] == src0[0] && src1[0] != src2[-1]) || (src1[-1] == src2[0] && src1[0] != src0[-1]) ? src1[-1] : src1[0];
 623                         dst[1] = src1[0];
 624                         dst[2] = (src1[1] == src0[0] && src1[0] != src2[1]) || (src1[1] == src2[0] && src1[0] != src0[1]) ? src1[1] : src1[0];
 625                 } else {
 626                         dst[0] = src1[0];
 627                         dst[1] = src1[0];
 628                         dst[2] = src1[0];
 629                 }
 630
 631                 ++src0;
 632                 ++src1;
 633                 ++src2;
 634                 dst += 3;
 635                 --count;
 636         }
 637
 638         /* last pixel */
 639         if (src0[0] != src2[0] && src1[-1] != src1[0]) {
 640                 dst[0] = (src1[-1] == src0[0] && src1[0] != src2[-1]) || (src1[-1] == src2[0] && src1[0] != src0[-1]) ? src1[-1] : src1[0];
 641                 dst[1] = src1[0];
 642                 dst[2] = (src1[0] == src0[0] && src1[0] != src2[0]) || (src1[0] == src2[0] && src1[0] != src0[0]) ? src1[0] : src1[0];
 643         } else {
 644                 dst[0] = src1[0];
 645                 dst[1] = src1[0];
 646                 dst[2] = src1[0];
 647         }
 648 }
 649
 650 #endif /* USE_SCALE_RANDOMWRITE */
 651
 652 /**
 653  * Scale by a factor of 3 a row of pixels of 8 bits.
 654  * The function is implemented in C.
 655  * The pixels over the left and right borders are assumed of the same color of
 656  * the pixels on the border.
 657  * \param src0 Pointer at the first pixel of the previous row.
 658  * \param src1 Pointer at the first pixel of the current row.
 659  * \param src2 Pointer at the first pixel of the next row.
 660  * \param count Length in pixels of the src0, src1 and src2 rows.
 661  * It must be at least 2.
 662  * \param dst0 First destination row, triple length in pixels.
 663  * \param dst1 Second destination row, triple length in pixels.
 664  * \param dst2 Third destination row, triple length in pixels.
 665  */
 666 void scale3x_8_def(scale3x_uint8* dst0, scale3x_uint8* dst1, scale3x_uint8* dst2, const scale3x_uint8* src0, const scale3x_uint8* src1, const scale3x_uint8* src2, unsigned count)
 667 {
 668 #ifdef USE_SCALE_RANDOMWRITE
 669         scale3x_8_def_whole(dst0, dst1, dst2, src0, src1, src2, count);
 670 #else
 671         scale3x_8_def_border(dst0, src0, src1, src2, count);
 672         scale3x_8_def_center(dst1, src0, src1, src2, count);
 673         scale3x_8_def_border(dst2, src2, src1, src0, count);
 674 #endif
 675 }
 676
 677 /**
 678  * Scale by a factor of 3 a row of pixels of 16 bits.
 679  * This function operates like scale3x_8_def() but for 16 bits pixels.
 680  * \param src0 Pointer at the first pixel of the previous row.
 681  * \param src1 Pointer at the first pixel of the current row.
 682  * \param src2 Pointer at the first pixel of the next row.
 683  * \param count Length in pixels of the src0, src1 and src2 rows.
 684  * It must be at least 2.
 685  * \param dst0 First destination row, triple length in pixels.
 686  * \param dst1 Second destination row, triple length in pixels.
 687  * \param dst2 Third destination row, triple length in pixels.
 688  */
 689 void scale3x_16_def(scale3x_uint16* dst0, scale3x_uint16* dst1, scale3x_uint16* dst2, const scale3x_uint16* src0, const scale3x_uint16* src1, const scale3x_uint16* src2, unsigned count)
 690 {
 691 #ifdef USE_SCALE_RANDOMWRITE
 692         scale3x_16_def_whole(dst0, dst1, dst2, src0, src1, src2, count);
 693 #else
 694         scale3x_16_def_border(dst0, src0, src1, src2, count);
 695         scale3x_16_def_center(dst1, src0, src1, src2, count);
 696         scale3x_16_def_border(dst2, src2, src1, src0, count);
 697 #endif
 698 }
 699
 700 /**
 701  * Scale by a factor of 3 a row of pixels of 32 bits.
 702  * This function operates like scale3x_8_def() but for 32 bits pixels.
 703  * \param src0 Pointer at the first pixel of the previous row.
 704  * \param src1 Pointer at the first pixel of the current row.
 705  * \param src2 Pointer at the first pixel of the next row.
 706  * \param count Length in pixels of the src0, src1 and src2 rows.
 707  * It must be at least 2.
 708  * \param dst0 First destination row, triple length in pixels.
 709  * \param dst1 Second destination row, triple length in pixels.
 710  * \param dst2 Third destination row, triple length in pixels.
 711  */
 712 void scale3x_32_def(scale3x_uint32* dst0, scale3x_uint32* dst1, scale3x_uint32* dst2, const scale3x_uint32* src0, const scale3x_uint32* src1, const scale3x_uint32* src2, unsigned count)
 713 {
 714 #ifdef USE_SCALE_RANDOMWRITE
 715         scale3x_32_def_whole(dst0, dst1, dst2, src0, src1, src2, count);
 716 #else
 717         scale3x_32_def_border(dst0, src0, src1, src2, count);
 718         scale3x_32_def_center(dst1, src0, src1, src2, count);
 719         scale3x_32_def_border(dst2, src2, src1, src0, count);
 720 #endif
 721 }
 722