src/libnr/nr-compose.cpp

   1 #define __NR_COMPOSE_C__
   2
   3 /*
   4  * Pixel buffer rendering library
   5  *
   6  * Authors:
   7  *   Lauris Kaplinski <lauris@kaplinski.com>
   8  *
   9  * This code is in public domain
  10  */
  11
  12 #ifdef HAVE_CONFIG_H
  13 # include "config.h"
  14 #endif
  15
  16 #include <string.h>
  17 #include "nr-pixops.h"
  18
  19 #ifdef WITH_MMX
  20 /* fixme: */
  21 #ifdef __cplusplus
  22 extern "C" {
  23 #endif /* __cplusplus */
  24 int nr_have_mmx (void);
  25 void nr_mmx_R8G8B8A8_P_EMPTY_A8_RGBAP (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned char *c);
  26 void nr_mmx_R8G8B8A8_P_R8G8B8A8_P_A8_RGBAP (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned char *c);
  27 void nr_mmx_R8G8B8_R8G8B8_R8G8B8A8_P (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned int alpha);
  28 #define NR_PIXOPS_MMX nr_have_mmx ()
  29 #ifdef __cplusplus
  30 }
  31 #endif /* __cplusplus */
  32 #endif
  33
  34 // Naming: nr_RESULT_BACKGROUND_FOREGROUND_extra
  35
  36 void
  37 nr_R8G8B8A8_N_EMPTY_R8G8B8A8_N (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned int alpha)
  38 {
  39         unsigned int r, c;
  40
  41         for (r = h; r > 0; r--) {
  42                 if (alpha == 0) {
  43                         memset(px, 0x0, 4 * w);
  44                 } else if (alpha == 255) {
  45                         memcpy(px, spx, 4 * w);
  46                 } else {
  47                         unsigned char *d = px;
  48                         const unsigned char *s = spx;
  49                         for (c = w; c > 0; c--) {
  50                                 *d++ = *s++;
  51                                 *d++ = *s++;
  52                                 *d++ = *s++;
  53                                 *d++ = NR_PREMUL_111(*s, alpha);
  54                                 s++;
  55                         }
  56                 }
  57                 px += rs;
  58                 spx += srs;
  59         }
  60 }
  61
  62 void
  63 nr_R8G8B8A8_N_EMPTY_R8G8B8A8_P (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned int alpha)
  64 {
  65         unsigned int r, c;
  66
  67         for (r = h; r > 0; r--) {
  68                 if (alpha == 0) {
  69                         memset(px, 0x0, 4 * w);
  70                 } else {
  71                         unsigned char *d = px;
  72                         const unsigned char *s = spx;
  73                         for (c = w; c > 0; c--) {
  74                                 if (s[3] == 0) {
  75                                         d[3] = 0;
  76                                 } else if (s[3] == 255) {
  77                                         memcpy(d, s, 4);
  78                                 } else {
  79                                         d[0] = NR_DEMUL_111(s[0], s[3]);
  80                                         d[1] = NR_DEMUL_111(s[1], s[3]);
  81                                         d[2] = NR_DEMUL_111(s[2], s[3]);
  82                                         d[3] = NR_PREMUL_111(s[3], alpha);
  83                                 }
  84                                 d += 4;
  85                                 s += 4;
  86                         }
  87                 }
  88                 px += rs;
  89                 spx += srs;
  90         }
  91 }
  92
  93 void
  94 nr_R8G8B8A8_P_EMPTY_R8G8B8A8_N (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned int alpha)
  95 {
  96         unsigned int r, c;
  97
  98         for (r = h; r > 0; r--) {
  99                 unsigned char *d = px;
 100                 const unsigned char *s = spx;
 101                 if (alpha == 0) {
 102                         memset(px, 0x0, 4 * w);
 103                 } else if (alpha == 255) {
 104                         for (c = w; c > 0; c--) {
 105                                 d[0] = NR_PREMUL_111(s[0], s[3]);
 106                                 d[1] = NR_PREMUL_111(s[1], s[3]);
 107                                 d[2] = NR_PREMUL_111(s[2], s[3]);
 108                                 d[3] = s[3];
 109                                 d += 4;
 110                                 s += 4;
 111                         }
 112                 } else {
 113                         for (c = w; c > 0; c--) {
 114                                 if (s[3] == 0) {
 115                                         memset(d, 0, 4);
 116                                 } else {
 117                                         unsigned int a;
 118                                         a = NR_PREMUL_112(s[3], alpha);
 119                                         d[0] = NR_PREMUL_121(s[0], a);
 120                                         d[1] = NR_PREMUL_121(s[1], a);
 121                                         d[2] = NR_PREMUL_121(s[2], a);
 122                                         d[3] = NR_NORMALIZE_21(a);
 123                                 }
 124                                 d += 4;
 125                                 s += 4;
 126                         }
 127                 }
 128                 px += rs;
 129                 spx += srs;
 130         }
 131 }
 132
 133 void
 134 nr_R8G8B8A8_P_EMPTY_R8G8B8A8_P (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned int alpha)
 135 {
 136         unsigned int r, c;
 137
 138         for (r = h; r > 0; r--) {
 139                 if (alpha == 0) {
 140                         memset(px, 0x0, 4 * w);
 141                 } else if (alpha == 255) {
 142                         memcpy(px, spx, 4 * w);
 143                 } else {
 144                         unsigned char *d = px;
 145                         const unsigned char *s = spx;
 146                         for (c = w; c > 0; c--) {
 147                                 d[0] = NR_PREMUL_111(s[0], alpha);
 148                                 d[1] = NR_PREMUL_111(s[1], alpha);
 149                                 d[2] = NR_PREMUL_111(s[2], alpha);
 150                                 d[3] = NR_PREMUL_111(s[3], alpha);
 151                                 d += 4;
 152                                 s += 4;
 153                         }
 154                 }
 155                 px += rs;
 156                 spx += srs;
 157         }
 158 }
 159
 160 void
 161 nr_R8G8B8A8_N_R8G8B8A8_N_R8G8B8A8_N (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned int alpha)
 162 {
 163         unsigned int r, c;
 164
 165         if (alpha == 0) {
 166                 /* NOP */
 167         } else if (alpha == 255) {
 168                 for (r = h; r > 0; r--) {
 169                         unsigned char *d = px;
 170                         const unsigned char *s = spx;
 171                         for (c = w; c > 0; c--) {
 172                                 if (s[3] == 0) {
 173                                         /* Transparent FG, NOP */
 174                                 } else if ((s[3] == 255) || (d[3] == 0)) {
 175                                         /* Full coverage, COPY */
 176                                         memcpy(d, s, 4);
 177                                 } else {
 178                                         /* Full composition */
 179                                         unsigned int ca;
 180                                         ca = NR_COMPOSEA_112(s[3], d[3]);
 181                                         d[0] = NR_COMPOSENNN_111121(s[0], s[3], d[0], d[3], ca);
 182                                         d[1] = NR_COMPOSENNN_111121(s[1], s[3], d[1], d[3], ca);
 183                                         d[2] = NR_COMPOSENNN_111121(s[2], s[3], d[2], d[3], ca);
 184                                         d[3] = NR_NORMALIZE_21(ca);
 185                                 }
 186                                 d += 4;
 187                                 s += 4;
 188                         }
 189                         px += rs;
 190                         spx += srs;
 191                 }
 192         } else {
 193                 for (r = h; r > 0; r--) {
 194                         unsigned char *d = px;
 195                         const unsigned char *s = spx;
 196                         for (c = w; c > 0; c--) {
 197                                 unsigned int a;
 198                                 a = NR_PREMUL_112(s[3], alpha);
 199                                 if (a == 0) {
 200                                         /* Transparent FG, NOP */
 201                                 } else if ((a == 255*255) || (d[3] == 0)) {
 202                                         /* Full coverage, COPY */
 203                                         d[0] = s[0];
 204                                         d[1] = s[1];
 205                                         d[2] = s[2];
 206                                         d[3] = NR_NORMALIZE_21(a);
 207                                 } else {
 208                                         /* Full composition */
 209                                         unsigned int ca;
 210                                         ca = NR_COMPOSEA_213(a, d[3]);
 211                                         d[0] = NR_COMPOSENNN_121131(s[0], a, d[0], d[3], ca);
 212                                         d[1] = NR_COMPOSENNN_121131(s[1], a, d[1], d[3], ca);
 213                                         d[2] = NR_COMPOSENNN_121131(s[2], a, d[2], d[3], ca);
 214                                         d[3] = NR_NORMALIZE_31(ca);
 215                                 }
 216                                 d += 4;
 217                                 s += 4;
 218                         }
 219                         px += rs;
 220                         spx += srs;
 221                 }
 222         }
 223 }
 224
 225 void
 226 nr_R8G8B8A8_N_R8G8B8A8_N_R8G8B8A8_P (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned int alpha)
 227 {
 228         unsigned int r, c;
 229
 230         if (alpha == 0) {
 231                 /* NOP */
 232         } else if (alpha == 255) {
 233                 for (r = h; r > 0; r--) {
 234                         unsigned char *d = px;
 235                         const unsigned char *s = spx;
 236                         for (c = w; c > 0; c--) {
 237                                 if (s[3] == 0) {
 238                                         /* Transparent FG, NOP */
 239                                 } else if (s[3] == 255) {
 240                                         /* Full coverage, demul src */
 241                                         //   dc' = ((1 - sa) * da*dc + sc)/da' = sc/da' = sc
 242                                         //   da' = 1 - (1 - sa) * (1 - da) = 1 - 0 * (1 - da) = 1
 243                                         memcpy(d, s, 4);
 244                                 } else if (d[3] == 0) {
 245                                         /* Full coverage, demul src */
 246                                         //   dc' = ((1 - sa) * da*dc + sc)/da' = sc/da' = sc/sa = sc/sa
 247                                         //   da' = 1 - (1 - sa) * (1 - da) = 1 - (1 - sa) = sa
 248                                         d[0] = NR_DEMUL_111(s[0], s[3]);
 249                                         d[1] = NR_DEMUL_111(s[1], s[3]);
 250                                         d[2] = NR_DEMUL_111(s[2], s[3]);
 251                                         d[3] = s[3];
 252                                 } else {
 253                                         /* Full composition */
 254                                         //   dc' = ((1 - sa) * da*dc + sc)/da' = ((1 - sa) * da*dc + sc)/da'
 255                                         //   da' = 1 - (1 - sa) * (1 - da) = 1 - (1 - sa) * (1 - da)
 256                                         unsigned int da = NR_COMPOSEA_112(s[3], d[3]);
 257                                         d[0] = NR_COMPOSEPNN_111121(s[0], s[3], d[0], d[3], da);
 258                                         d[1] = NR_COMPOSEPNN_111121(s[1], s[3], d[1], d[3], da);
 259                                         d[2] = NR_COMPOSEPNN_111121(s[2], s[3], d[2], d[3], da);
 260                                         d[3] = NR_NORMALIZE_21(da);
 261                                 }
 262                                 d += 4;
 263                                 s += 4;
 264                         }
 265                         px += rs;
 266                         spx += srs;
 267                 }
 268         } else {
 269                 for (r = h; r > 0; r--) {
 270                         unsigned char *d = px;
 271                         const unsigned char *s = spx;
 272                         for (c = w; c > 0; c--) {
 273                                 unsigned int a;
 274                                 a = NR_PREMUL_112(s[3], alpha);
 275                                 if (a == 0) {
 276                                         /* Transparent FG, NOP */
 277                                 } else if (d[3] == 0) {
 278                                         /* Full coverage, demul src */
 279                                         //   dc' = ((1 - alpha*sa) * da*dc + alpha*sc)/da' = alpha*sc/da' = alpha*sc/(alpha*sa) = sc/sa
 280                                         //   da' = 1 - (1 - alpha*sa) * (1 - da) = 1 - (1 - alpha*sa) = alpha*sa
 281                                         d[0] = NR_DEMUL_111(s[0], s[3]);
 282                                         d[1] = NR_DEMUL_111(s[1], s[3]);
 283                                         d[2] = NR_DEMUL_111(s[2], s[3]);
 284                                         d[3] = NR_NORMALIZE_21(a);
 285                                 } else {
 286                                         //   dc' = ((1 - alpha*sa) * da*dc + alpha*sc)/da'
 287                                         //   da' = 1 - (1 - alpha*sa) * (1 - da)
 288                                         unsigned int da = NR_COMPOSEA_213(a, d[3]);
 289                                         d[0] = NR_COMPOSEPNN_221131(NR_PREMUL_112(s[0], alpha), a, d[0], d[3], da);
 290                                         d[1] = NR_COMPOSEPNN_221131(NR_PREMUL_112(s[1], alpha), a, d[1], d[3], da);
 291                                         d[2] = NR_COMPOSEPNN_221131(NR_PREMUL_112(s[2], alpha), a, d[2], d[3], da);
 292                                         d[3] = NR_NORMALIZE_31(da);
 293                                 }
 294                                 d += 4;
 295                                 s += 4;
 296                         }
 297                         px += rs;
 298                         spx += srs;
 299                 }
 300         }
 301 }
 302
 303 void
 304 nr_R8G8B8A8_P_R8G8B8A8_P_R8G8B8A8_N (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned int alpha)
 305 {
 306         unsigned int r, c;
 307
 308         if (alpha == 0) {
 309                 /* NOP */
 310         } else if (alpha == 255) {
 311                 for (r = h; r > 0; r--) {
 312                         unsigned char *d = px;
 313                         const unsigned char *s = spx;
 314                         for (c = w; c > 0; c--) {
 315                                 if (s[3] == 0) {
 316                                         /* Transparent FG, NOP */
 317                                 } else if (s[3] == 255) {
 318                                         /* Opaque FG, COPY */
 319                                         //   dc' = (1 - sa) * dc + sa*sc = sa*sc = sc
 320                                         //   da' = 1 - (1 - sa) * (1 - da) = 1 - 0 * (1 - da) = 1 (= sa)
 321                                         memcpy(d, s, 4);
 322                                 } else if (d[3] == 0) {
 323                                         /* Transparent BG, premul src */
 324                                         //   dc' = (1 - sa) * dc + sa*sc = sa*sc
 325                                         //   da' = 1 - (1 - sa) * (1 - da) = 1 - (1 - sa) = sa
 326                                         d[0] = NR_PREMUL_111(s[0], s[3]);
 327                                         d[1] = NR_PREMUL_111(s[1], s[3]);
 328                                         d[2] = NR_PREMUL_111(s[2], s[3]);
 329                                         d[3] = s[3];
 330                                 } else {
 331                                         //   dc' = (1 - sa) * dc + sa*sc
 332                                         //   da' = 1 - (1 - sa) * (1 - da)
 333                                         d[0] = NR_COMPOSENPP_1111(s[0], s[3], d[0]);
 334                                         d[1] = NR_COMPOSENPP_1111(s[1], s[3], d[1]);
 335                                         d[2] = NR_COMPOSENPP_1111(s[2], s[3], d[2]);
 336                                         d[3] = NR_COMPOSEA_111(s[3], d[3]);
 337                                 }
 338                                 d += 4;
 339                                 s += 4;
 340                         }
 341                         px += rs;
 342                         spx += srs;
 343                 }
 344         } else {
 345                 for (r = h; r > 0; r--) {
 346                         unsigned char *d = px;
 347                         const unsigned char *s = spx;
 348                         for (c = w; c > 0; c--) {
 349                                 unsigned int a;
 350                                 a = NR_PREMUL_112 (s[3], alpha);
 351                                 if (a == 0) {
 352                                         /* Transparent FG, NOP */
 353                                 } else if (d[3] == 0) {
 354                                         /* Transparent BG, premul src */
 355                                         //   dc' = (1 - alpha*sa) * dc + alpha*sa*sc = alpha*sa*sc
 356                                         //   da' = 1 - (1 - alpha*sa) * (1 - da) = 1 - (1 - alpha*sa) = alpha*sa
 357                                         d[0] = NR_PREMUL_121(s[0], a);
 358                                         d[1] = NR_PREMUL_121(s[1], a);
 359                                         d[2] = NR_PREMUL_121(s[2], a);
 360                                         d[3] = NR_NORMALIZE_21(a);
 361                                 } else {
 362                                         //   dc' = (1 - alpha*sa) * dc + alpha*sa*sc
 363                                         //   da' = 1 - (1 - alpha*sa) * (1 - da)
 364                                         d[0] = NR_COMPOSENPP_1211(s[0], a, d[0]);
 365                                         d[1] = NR_COMPOSENPP_1211(s[1], a, d[1]);
 366                                         d[2] = NR_COMPOSENPP_1211(s[2], a, d[2]);
 367                                         d[3] = NR_COMPOSEA_211(a, d[3]);
 368                                 }
 369                                 d += 4;
 370                                 s += 4;
 371                         }
 372                         px += rs;
 373                         spx += srs;
 374                 }
 375         }
 376 }
 377
 378 void
 379 nr_R8G8B8A8_P_R8G8B8A8_P_R8G8B8A8_P (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned int alpha)
 380 {
 381         unsigned int r, c;
 382
 383         if (alpha == 0) {
 384                 /* Transparent FG, NOP */
 385         } else if (alpha == 255) {
 386                 /* Simple */
 387                 for (r = h; r > 0; r--) {
 388                         unsigned char *d = px;
 389                         const unsigned char *s = spx;
 390                         for (c = w; c > 0; c--) {
 391                                 if (s[3] == 0) {
 392                                         /* Transparent FG, NOP */
 393                                 } else if ((s[3] == 255) || (d[3] == 0)) {
 394                                         /* Transparent BG, COPY */
 395                                         memcpy(d, s, 4);
 396                                 } else {
 397                                         d[0] = NR_COMPOSEPPP_1111(s[0], s[3], d[0]);
 398                                         d[1] = NR_COMPOSEPPP_1111(s[1], s[3], d[1]);
 399                                         d[2] = NR_COMPOSEPPP_1111(s[2], s[3], d[2]);
 400                                         d[3] = NR_COMPOSEA_111(s[3], d[3]);
 401                                 }
 402                                 d += 4;
 403                                 s += 4;
 404                         }
 405                         px += rs;
 406                         spx += srs;
 407                 }
 408         } else {
 409                 for (r = h; r > 0; r--) {
 410                         unsigned char *d = px;
 411                         const unsigned char *s = spx;
 412                         for (c = w; c > 0; c--) {
 413                                 if (s[3] == 0) {
 414                                         /* Transparent FG, NOP */
 415                                 } else if (d[3] == 0) {
 416                                         /* Transparent BG, COPY */
 417                                         d[0] = NR_PREMUL_111(s[0], alpha);
 418                                         d[1] = NR_PREMUL_111(s[1], alpha);
 419                                         d[2] = NR_PREMUL_111(s[2], alpha);
 420                                         d[3] = NR_PREMUL_111(s[3], alpha);
 421                                 } else {
 422                                         //   dc' = (1 - alpha*sa) * dc + alpha*sc
 423                                         //   da' = 1 - (1 - alpha*sa) * (1 - da)
 424                                         unsigned int a;
 425                                         a = NR_PREMUL_112(s[3], alpha);
 426                                         d[0] = NR_COMPOSEPPP_2211(NR_PREMUL_112(alpha, s[0]), a, d[0]);
 427                                         d[1] = NR_COMPOSEPPP_2211(NR_PREMUL_112(alpha, s[1]), a, d[1]);
 428                                         d[2] = NR_COMPOSEPPP_2211(NR_PREMUL_112(alpha, s[2]), a, d[2]);
 429                                         d[3] = NR_COMPOSEA_211(a, d[3]);
 430                                 }
 431                                 d += 4;
 432                                 s += 4;
 433                         }
 434                         px += rs;
 435                         spx += srs;
 436                 }
 437         }
 438 }
 439
 440 /* Masked operations */
 441
 442 void
 443 nr_R8G8B8A8_N_EMPTY_R8G8B8A8_N_A8 (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, const unsigned char *mpx, int mrs)
 444 {
 445         unsigned int r, c;
 446
 447         for (r = h; r > 0; r--) {
 448                 unsigned char *d = px;
 449                 const unsigned char *s = spx;
 450                 const unsigned char *m = mpx;
 451                 for (c = w; c > 0; c--) {
 452                         d[0] = s[0];
 453                         d[1] = s[1];
 454                         d[2] = s[2];
 455                         d[3] = NR_PREMUL_111(s[3], m[0]);
 456                         d += 4;
 457                         s += 4;
 458                         m += 1;
 459                 }
 460                 px += rs;
 461                 spx += srs;
 462                 mpx += mrs;
 463         }
 464 }
 465
 466 void
 467 nr_R8G8B8A8_N_EMPTY_R8G8B8A8_P_A8 (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, const unsigned char *mpx, int mrs)
 468 {
 469         unsigned int r, c;
 470
 471         for (r = h; r > 0; r--) {
 472                 unsigned char *d = px;
 473                 const unsigned char *s = spx;
 474                 const unsigned char *m = mpx;
 475                 for (c = w; c > 0; c--) {
 476                         unsigned int a;
 477                         a = NR_PREMUL_112 (s[3], m[0]);
 478                         if (a == 0) {
 479                                 d[3] = 0;
 480                         } else if (a == 255*255) {
 481                                 memcpy(d, s, 4);
 482                         } else {
 483                                 //   dc' = ((1 - m*sa) * da*dc + m*sc)/da' = m*sc/da' = m*sc/(m*sa) = sc/sa
 484                                 //   da' = 1 - (1 - m*sa) * (1 - da) = 1 - (1 - m*sa) = m*sa
 485                                 d[0] = NR_DEMUL_111(s[0], s[3]);
 486                                 d[1] = NR_DEMUL_111(s[1], s[3]);
 487                                 d[2] = NR_DEMUL_111(s[2], s[3]);
 488                                 d[3] = NR_NORMALIZE_21(a);
 489                         }
 490                         d += 4;
 491                         s += 4;
 492                         m += 1;
 493                 }
 494                 px += rs;
 495                 spx += srs;
 496                 mpx += mrs;
 497         }
 498 }
 499
 500 void
 501 nr_R8G8B8A8_P_EMPTY_R8G8B8A8_N_A8 (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, const unsigned char *mpx, int mrs)
 502 {
 503         unsigned int r, c;
 504
 505         for (r = h; r > 0; r--) {
 506                 unsigned char *d = px;
 507                 const unsigned char *s = spx;
 508                 const unsigned char *m = mpx;
 509                 for (c = w; c > 0; c--) {
 510                         unsigned int a;
 511                         a = NR_PREMUL_112(s[3], m[0]);
 512                         if (a == 0) {
 513                                 memset(d, 0, 4);
 514                         } else if (a == 255*255) {
 515                                 memcpy(d, s, 4);
 516                         } else {
 517                                 d[0] = NR_PREMUL_121(s[0], a);
 518                                 d[1] = NR_PREMUL_121(s[1], a);
 519                                 d[2] = NR_PREMUL_121(s[2], a);
 520                                 d[3] = NR_NORMALIZE_21(a);
 521                         }
 522                         d += 4;
 523                         s += 4;
 524                         m += 1;
 525                 }
 526                 px += rs;
 527                 spx += srs;
 528                 mpx += mrs;
 529         }
 530 }
 531
 532 void
 533 nr_R8G8B8A8_P_EMPTY_R8G8B8A8_P_A8 (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, const unsigned char *mpx, int mrs)
 534 {
 535         unsigned int r, c;
 536
 537         for (r = h; r > 0; r--) {
 538                 unsigned char *d = px;
 539                 const unsigned char *s = spx;
 540                 const unsigned char *m = mpx;
 541                 for (c = w; c > 0; c--) {
 542                         d[0] = NR_PREMUL_111(s[0], m[0]);
 543                         d[1] = NR_PREMUL_111(s[1], m[0]);
 544                         d[2] = NR_PREMUL_111(s[2], m[0]);
 545                         d[3] = NR_PREMUL_111(s[3], m[0]);
 546                         d += 4;
 547                         s += 4;
 548                         m += 1;
 549                 }
 550                 px += rs;
 551                 spx += srs;
 552                 mpx += mrs;
 553         }
 554 }
 555
 556 void
 557 nr_R8G8B8A8_N_R8G8B8A8_N_R8G8B8A8_N_A8 (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, const unsigned char *mpx, int mrs)
 558 {
 559         unsigned int r, c;
 560
 561         for (r = h; r > 0; r--) {
 562                 unsigned char *d = px;
 563                 const unsigned char *s = spx;
 564                 const unsigned char *m = mpx;
 565                 for (c = w; c > 0; c--) {
 566                         unsigned int a;
 567                         a = NR_PREMUL_112(s[3], m[0]);
 568                         if (a == 0) {
 569                                 /* Transparent FG, NOP */
 570                         } else if ((a == 255*255) || (d[3] == 0)) {
 571                                 /* Full coverage, COPY */
 572                                 d[0] = s[0];
 573                                 d[1] = s[1];
 574                                 d[2] = s[2];
 575                                 d[3] = NR_NORMALIZE_21(a);
 576                         } else {
 577                                 /* Full composition */
 578                                 unsigned int ca;
 579                                 ca = NR_COMPOSEA_213(a, d[3]);
 580                                 d[0] = NR_COMPOSENNN_121131(s[0], a, d[0], d[3], ca);
 581                                 d[1] = NR_COMPOSENNN_121131(s[1], a, d[1], d[3], ca);
 582                                 d[2] = NR_COMPOSENNN_121131(s[2], a, d[2], d[3], ca);
 583                                 d[3] = NR_NORMALIZE_31(ca);
 584                         }
 585                         d += 4;
 586                         s += 4;
 587                         m += 1;
 588                 }
 589                 px += rs;
 590                 spx += srs;
 591                 mpx += mrs;
 592         }
 593 }
 594
 595 void
 596 nr_R8G8B8A8_N_R8G8B8A8_N_R8G8B8A8_P_A8 (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, const unsigned char *mpx, int mrs)
 597 {
 598         unsigned int r, c;
 599
 600         for (r = h; r > 0; r--) {
 601                 unsigned char *d = px;
 602                 const unsigned char *s = spx;
 603                 const unsigned char *m = mpx;
 604                 for (c = w; c > 0; c--) {
 605                         unsigned int a;
 606                         a = NR_PREMUL_112(s[3], m[0]);
 607                         if (a == 0) {
 608                                 /* Transparent FG, NOP */
 609                         } else if (a == 255*255) {
 610                                 /* Opaque FG, COPY */
 611                                 memcpy(d, s, 4);
 612                         } else if (d[3] == 0) {
 613                                 /* Full coverage, demul src */
 614                                 //   dc' = ((1 - m*sa) * da*dc + m*sc)/da' = m*sc/da' = m*sc/(m*sa) = sc/sa
 615                                 //   da' = 1 - (1 - m*sa) * (1 - da) = 1 - (1 - m*sa) = m*sa
 616                                 d[0] = NR_DEMUL_111(s[0], s[3]);
 617                                 d[1] = NR_DEMUL_111(s[1], s[3]);
 618                                 d[2] = NR_DEMUL_111(s[2], s[3]);
 619                                 d[3] = NR_NORMALIZE_21(a);
 620                         } else if (m[0] == 255) {
 621                                 /* Full composition */
 622                                 //   dc' = ((1 - m*sa) * da*dc + m*sc)/da' = ((1 - sa) * da*dc + sc)/da'
 623                                 //   da' = 1 - (1 - m*sa) * (1 - da) = 1 - (1 - sa) * (1 - da)
 624                                 unsigned int da = NR_COMPOSEA_112(s[3], d[3]);
 625                                 d[0] = NR_COMPOSEPNN_111121(s[0], s[3], d[0], d[3], da);
 626                                 d[1] = NR_COMPOSEPNN_111121(s[1], s[3], d[1], d[3], da);
 627                                 d[2] = NR_COMPOSEPNN_111121(s[2], s[3], d[2], d[3], da);
 628                                 d[3] = NR_NORMALIZE_21(da);
 629                         } else {
 630                                 //   dc' = ((1 - m*sa) * da*dc + m*sc)/da'
 631                                 //   da' = 1 - (1 - m*sa) * (1 - da)
 632                                 unsigned int da = NR_COMPOSEA_213(a, d[3]);
 633                                 d[0] = NR_COMPOSEPNN_221131(NR_PREMUL_112(s[0], m[0]), a, d[0], d[3], da);
 634                                 d[1] = NR_COMPOSEPNN_221131(NR_PREMUL_112(s[1], m[0]), a, d[1], d[3], da);
 635                                 d[2] = NR_COMPOSEPNN_221131(NR_PREMUL_112(s[2], m[0]), a, d[2], d[3], da);
 636                                 d[3] = NR_NORMALIZE_31(da);
 637                         }
 638                         d += 4;
 639                         s += 4;
 640                         m += 1;
 641                 }
 642                 px += rs;
 643                 spx += srs;
 644                 mpx += mrs;
 645         }
 646 }
 647
 648 void
 649 nr_R8G8B8A8_P_R8G8B8A8_P_R8G8B8A8_N_A8 (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, const unsigned char *mpx, int mrs)
 650 {
 651         unsigned int r, c;
 652
 653         for (r = h; r>0; r--) {
 654                 unsigned char *d = px;
 655                 const unsigned char *s = spx;
 656                 const unsigned char *m = mpx;
 657                 for (c = w; c>0; c--) {
 658                         unsigned int a;
 659                         a = NR_PREMUL_112(s[3], m[0]);
 660                         if (a == 0) {
 661                                 /* Transparent FG, NOP */
 662                         } else if (a == 255*255) {
 663                                 memcpy(d, s, 4);
 664                         } else {
 665                                 d[0] = NR_COMPOSENPP_1211(s[0], a, d[0]);
 666                                 d[1] = NR_COMPOSENPP_1211(s[1], a, d[1]);
 667                                 d[2] = NR_COMPOSENPP_1211(s[2], a, d[2]);
 668                                 d[3] = NR_COMPOSEA_211(a, d[3]);
 669                         }
 670                         d += 4;
 671                         s += 4;
 672                         m += 1;
 673                 }
 674                 px += rs;
 675                 spx += srs;
 676                 mpx += mrs;
 677         }
 678 }
 679
 680 void
 681 nr_R8G8B8A8_P_R8G8B8A8_P_R8G8B8A8_P_A8 (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, const unsigned char *mpx, int mrs)
 682 {
 683         unsigned int r, c;
 684
 685         for (r = h; r > 0; r--) {
 686                 unsigned char *d = px;
 687                 const unsigned char *s = spx;
 688                 const unsigned char *m = mpx;
 689                 for (c = w; c > 0; c--) {
 690                         unsigned int a;
 691                         a = NR_PREMUL_112 (s[3], m[0]);
 692                         if (a == 0) {
 693                                 /* Transparent FG, NOP */
 694                         } else if (a == 255*255) {
 695                                 /* Opaque FG, COPY */
 696                                 memcpy(d, s, 4);
 697                         } else if (d[3] == 0) {
 698                                 /* Transparent BG, COPY */
 699                                 //   dc' = (1 - m*sa) * dc + m*sc = m*sc
 700                                 //   da' = 1 - (1 - m*sa) * (1 - da) = 1 - (1 - m*sa)  = m*sa
 701                                 d[0] = NR_PREMUL_111 (s[0], m[0]);
 702                                 d[1] = NR_PREMUL_111 (s[1], m[0]);
 703                                 d[2] = NR_PREMUL_111 (s[2], m[0]);
 704                                 d[3] = NR_NORMALIZE_21(a);
 705                         } else {
 706                                 //   dc' = (1 - m*sa) * dc + m*sc
 707                                 //   da' = 1 - (1 - m*sa) * (1 - da)
 708                                 d[0] = NR_COMPOSEPPP_2211 (NR_PREMUL_112 (s[0], m[0]), a, d[0]);
 709                                 d[1] = NR_COMPOSEPPP_2211 (NR_PREMUL_112 (s[1], m[0]), a, d[1]);
 710                                 d[2] = NR_COMPOSEPPP_2211 (NR_PREMUL_112 (s[2], m[0]), a, d[2]);
 711                                 d[3] = NR_COMPOSEA_211(a, d[3]);
 712                         }
 713                         d += 4;
 714                         s += 4;
 715                         m += 1;
 716                 }
 717                 px += rs;
 718                 spx += srs;
 719                 mpx += mrs;
 720         }
 721 }
 722
 723 /* FINAL DST MASK COLOR */
 724
 725 void
 726 nr_R8G8B8A8_N_EMPTY_A8_RGBA32 (unsigned char *px, int w, int h, int rs, const unsigned char *mpx, int mrs, unsigned long rgba)
 727 {
 728         unsigned int r, g, b, a;
 729         unsigned int x, y;
 730
 731         r = NR_RGBA32_R (rgba);
 732         g = NR_RGBA32_G (rgba);
 733         b = NR_RGBA32_B (rgba);
 734         a = NR_RGBA32_A (rgba);
 735
 736         for (y = h; y > 0; y--) {
 737                 if (a == 0) {
 738                         memset(px, 0, w*4);
 739                 } else {
 740                         unsigned char *d = px;
 741                         const unsigned char *m = mpx;
 742                         for (x = w; x > 0; x--) {
 743                                 d[0] = r;
 744                                 d[1] = g;
 745                                 d[2] = b;
 746                                 d[3] = NR_PREMUL_111 (m[0], a);
 747                                 d += 4;
 748                                 m += 1;
 749                         }
 750                 }
 751                 px += rs;
 752                 mpx += mrs;
 753         }
 754 }
 755
 756 void
 757 nr_R8G8B8A8_P_EMPTY_A8_RGBA32 (unsigned char *px, int w, int h, int rs, const unsigned char *mpx, int mrs, unsigned long rgba)
 758 {
 759         unsigned int r, g, b, a;
 760         unsigned int x, y;
 761
 762         r = NR_RGBA32_R (rgba);
 763         g = NR_RGBA32_G (rgba);
 764         b = NR_RGBA32_B (rgba);
 765         a = NR_RGBA32_A (rgba);
 766
 767 #ifdef WITH_MMX
 768         if (NR_PIXOPS_MMX) {
 769                 unsigned char c[4];
 770                 c[0] = NR_PREMUL_111 (r, a);
 771                 c[1] = NR_PREMUL_111 (g, a);
 772                 c[2] = NR_PREMUL_111 (b, a);
 773                 c[3] = a;
 774                 /* WARNING: MMX composer REQUIRES w > 0 and h > 0 */
 775                 nr_mmx_R8G8B8A8_P_EMPTY_A8_RGBAP (px, w, h, rs, mpx, mrs, c);
 776                 return;
 777         }
 778 #endif
 779
 780         if ( a != 255 ){
 781                 // Pre-premultiply color values
 782                 r *= a;
 783                 g *= a;
 784                 b *= a;
 785         }
 786
 787         for (y = h; y > 0; y--) {
 788                 unsigned char *d = px;
 789                 const unsigned char *m = mpx;
 790                 if (a == 0) {
 791                         memset(px, 0, w*4);
 792                 } else if (a == 255) {
 793                         for (x = w; x > 0; x--) {
 794                                 d[0] = NR_PREMUL_111(m[0], r);
 795                                 d[1] = NR_PREMUL_111(m[0], g);
 796                                 d[2] = NR_PREMUL_111(m[0], b);
 797                                 d[3] = m[0];
 798                                 d += 4;
 799                                 m += 1;
 800                         }
 801                 } else {
 802                         for (x = w; x > 0; x--) {
 803                                 // Color values are already premultiplied with a
 804                                 d[0] = NR_PREMUL_121(m[0], r);
 805                                 d[1] = NR_PREMUL_121(m[0], g);
 806                                 d[2] = NR_PREMUL_121(m[0], b);
 807                                 d[3] = NR_PREMUL_111(m[0], a);
 808                                 d += 4;
 809                                 m += 1;
 810                         }
 811                 }
 812                 px += rs;
 813                 mpx += mrs;
 814         }
 815 }
 816
 817 void
 818 nr_R8G8B8_R8G8B8_A8_RGBA32 (unsigned char *px, int w, int h, int rs, const unsigned char *mpx, int mrs, unsigned long rgba)
 819 {
 820         unsigned int r, g, b, a;
 821         unsigned int x, y;
 822
 823         r = NR_RGBA32_R (rgba);
 824         g = NR_RGBA32_G (rgba);
 825         b = NR_RGBA32_B (rgba);
 826         a = NR_RGBA32_A (rgba);
 827
 828         if (a == 0) {
 829                 /* NOP */
 830         } else if (a == 255) {
 831                 for (y = h; y > 0; y--) {
 832                         unsigned char *d = px;
 833                         const unsigned char *m = mpx;
 834                         for (x = w; x > 0; x--) {
 835                                 d[0] = NR_COMPOSEN11_1111 (r, m[0], d[0]);
 836                                 d[1] = NR_COMPOSEN11_1111 (g, m[0], d[1]);
 837                                 d[2] = NR_COMPOSEN11_1111 (b, m[0], d[2]);
 838                                 d += 3;
 839                                 m += 1;
 840                         }
 841                         px += rs;
 842                         mpx += mrs;
 843                 }
 844         } else {
 845                 for (y = h; y > 0; y--) {
 846                         unsigned char *d = px;
 847                         const unsigned char *m = mpx;
 848                         for (x = w; x > 0; x--) {
 849                                 //   dc' = (1 - m*sa) * dc + m*sa*sc
 850                                 unsigned int alpha;
 851                                 alpha = NR_PREMUL_112 (a, m[0]);
 852                                 d[0] = NR_COMPOSEN11_1211 (r, alpha, d[0]);
 853                                 d[1] = NR_COMPOSEN11_1211 (g, alpha, d[1]);
 854                                 d[2] = NR_COMPOSEN11_1211 (b, alpha, d[2]);
 855                                 d += 3;
 856                                 m += 1;
 857                         }
 858                         px += rs;
 859                         mpx += mrs;
 860                 }
 861         }
 862 }
 863
 864 void
 865 nr_R8G8B8A8_N_R8G8B8A8_N_A8_RGBA32 (unsigned char *px, int w, int h, int rs, const unsigned char *mpx, int mrs, unsigned long rgba)
 866 {
 867         unsigned int r, g, b, a;
 868         unsigned int x, y;
 869
 870         r = NR_RGBA32_R (rgba);
 871         g = NR_RGBA32_G (rgba);
 872         b = NR_RGBA32_B (rgba);
 873         a = NR_RGBA32_A (rgba);
 874
 875         if (a == 0) {
 876                 /* NOP */
 877         } else if (a == 255) {
 878                 for (y = h; y > 0; y--) {
 879                         unsigned char *d = px;
 880                         const unsigned char *m = mpx;
 881                         for (x = w; x > 0; x--) {
 882                                 if (m[0] == 0) {
 883                                         /* Transparent FG, NOP */
 884                                 } else if (m[0] == 255 || d[3] == 0) {
 885                                         /* Full coverage, COPY */
 886                                         d[0] = r;
 887                                         d[1] = g;
 888                                         d[2] = b;
 889                                         d[3] = m[0];
 890                                 } else {
 891                                         /* Full composition */
 892                                         unsigned int da = NR_COMPOSEA_112(m[0], d[3]);
 893                                         d[0] = NR_COMPOSENNN_111121(r, m[0], d[0], d[3], da);
 894                                         d[1] = NR_COMPOSENNN_111121(g, m[0], d[1], d[3], da);
 895                                         d[2] = NR_COMPOSENNN_111121(b, m[0], d[2], d[3], da);
 896                                         d[3] = NR_NORMALIZE_21(da);
 897                                 }
 898                                 d += 4;
 899                                 m += 1;
 900                         }
 901                         px += rs;
 902                         mpx += mrs;
 903                 }
 904         } else {
 905                 for (y = h; y > 0; y--) {
 906                         unsigned char *d = px;
 907                         const unsigned char *m = mpx;
 908                         for (x = w; x > 0; x--) {
 909                                 unsigned int ca;
 910                                 ca = NR_PREMUL_112 (m[0], a);
 911                                 if (ca == 0) {
 912                                         /* Transparent FG, NOP */
 913                                 } else if (d[3] == 0) {
 914                                         /* Full coverage, COPY */
 915                                         d[0] = r;
 916                                         d[1] = g;
 917                                         d[2] = b;
 918                                         d[3] = NR_NORMALIZE_21(ca);
 919                                 } else {
 920                                         /* Full composition */
 921                                         unsigned int da = NR_COMPOSEA_213(ca, d[3]);
 922                                         d[0] = NR_COMPOSENNN_121131(r, ca, d[0], d[3], da);
 923                                         d[1] = NR_COMPOSENNN_121131(g, ca, d[1], d[3], da);
 924                                         d[2] = NR_COMPOSENNN_121131(b, ca, d[2], d[3], da);
 925                                         d[3] = NR_NORMALIZE_31(da);
 926                                 }
 927                                 d += 4;
 928                                 m += 1;
 929                         }
 930                         px += rs;
 931                         mpx += mrs;
 932                 }
 933         }
 934 }
 935
 936 void
 937 nr_R8G8B8A8_P_R8G8B8A8_P_A8_RGBA32 (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned long rgba)
 938 {
 939         unsigned int r, g, b, a;
 940         unsigned int x, y;
 941
 942         r = NR_RGBA32_R (rgba);
 943         g = NR_RGBA32_G (rgba);
 944         b = NR_RGBA32_B (rgba);
 945         a = NR_RGBA32_A (rgba);
 946
 947 #ifdef WITH_MMX
 948         if (NR_PIXOPS_MMX && a != 0) {
 949                 unsigned char c[4];
 950                 c[0] = NR_PREMUL_111 (r, a);
 951                 c[1] = NR_PREMUL_111 (g, a);
 952                 c[2] = NR_PREMUL_111 (b, a);
 953                 c[3] = a;
 954                 /* WARNING: MMX composer REQUIRES w > 0 and h > 0 */
 955                 nr_mmx_R8G8B8A8_P_R8G8B8A8_P_A8_RGBAP (px, w, h, rs, spx, srs, c);
 956                 return;
 957         }
 958 #endif
 959
 960         if (a == 0) {
 961                 /* Transparent FG, NOP */
 962         } else if (a == 255) {
 963                 /* Simple */
 964                 for (y = h; y > 0; y--) {
 965                         unsigned char *d, *s;
 966                         d = (unsigned char *) px;
 967                         s = (unsigned char *) spx;
 968                         for (x = w; x > 0; x--) {
 969                                 if (s[0] == 0) {
 970                                         /* Transparent FG, NOP */
 971                                 } else {
 972                                         /* Full composition */
 973                                         unsigned int invca = 255-s[0]; // By swapping the arguments GCC can better optimize these calls
 974                                         d[0] = NR_COMPOSENPP_1111(d[0], invca, r);
 975                                         d[1] = NR_COMPOSENPP_1111(d[1], invca, g);
 976                                         d[2] = NR_COMPOSENPP_1111(d[2], invca, b);
 977                                         d[3] = NR_COMPOSEA_111(s[0], d[3]);
 978                                 }
 979                                 d += 4;
 980                                 s += 1;
 981                         }
 982                         px += rs;
 983                         spx += srs;
 984                 }
 985         } else {
 986                 for (y = h; y > 0; y--) {
 987                         unsigned char *d, *s;
 988                         d = (unsigned char *) px;
 989                         s = (unsigned char *) spx;
 990                         for (x = w; x > 0; x--) {
 991                                 unsigned int ca;
 992                                 ca = NR_PREMUL_112 (s[0], a);
 993                                 if (ca == 0) {
 994                                         /* Transparent FG, NOP */
 995                                 } else {
 996                                         /* Full composition */
 997                                         unsigned int invca = 255*255-ca; // By swapping the arguments GCC can better optimize these calls
 998                                         d[0] = NR_COMPOSENPP_1211(d[0], invca, r);
 999                                         d[1] = NR_COMPOSENPP_1211(d[1], invca, g);
1000                                         d[2] = NR_COMPOSENPP_1211(d[2], invca, b);
1001                                         d[3] = NR_COMPOSEA_211(ca, d[3]);
1002                                 }
1003                                 d += 4;
1004                                 s += 1;
1005                         }
1006                         px += rs;
1007                         spx += srs;
1008                 }
1009         }
1010 }
1011
1012 /* RGB */
1013
1014 void
1015 nr_R8G8B8_R8G8B8_R8G8B8A8_P (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned int alpha)
1016 {
1017         unsigned int r, c;
1018
1019 #ifdef WITH_MMX
1020         if (NR_PIXOPS_MMX && alpha != 0) {
1021                 /* WARNING: MMX composer REQUIRES w > 0 and h > 0 */
1022                 nr_mmx_R8G8B8_R8G8B8_R8G8B8A8_P (px, w, h, rs, spx, srs, alpha);
1023                 return;
1024         }
1025 #endif
1026
1027         if (alpha == 0) {
1028                 /* NOP */
1029         } else if (alpha == 255) {
1030                 for (r = h; r > 0; r--) {
1031                         unsigned char *d = px;
1032                         const unsigned char *s = spx;
1033                         for (c = w; c > 0; c--) {
1034                                 //   dc' = (1 - alpha*sa) * dc + alpha*sc = (1 - sa) * dc + sc
1035                                 if (s[3] == 0) {
1036                                         /* NOP */
1037                                 } else if (s[3] == 255) {
1038                                         d[0] = s[0];
1039                                         d[1] = s[1];
1040                                         d[2] = s[2];
1041                                 } else {
1042                                         d[0] = NR_COMPOSEP11_1111(s[0], s[3], d[0]);
1043                                         d[1] = NR_COMPOSEP11_1111(s[1], s[3], d[1]);
1044                                         d[2] = NR_COMPOSEP11_1111(s[2], s[3], d[2]);
1045                                 }
1046                                 d += 3;
1047                                 s += 4;
1048                         }
1049                         px += rs;
1050                         spx += srs;
1051                 }
1052         } else {
1053                 for (r = h; r > 0; r--) {
1054                         unsigned char *d = px;
1055                         const unsigned char *s = spx;
1056                         for (c = w; c > 0; c--) {
1057                                 unsigned int a;
1058                                 a = NR_PREMUL_112(s[3], alpha);
1059                                 //   dc' = (1 - alpha*sa) * dc + alpha*sc
1060                                 if (a == 0) {
1061                                         /* NOP */
1062                                 } else {
1063                                         d[0] = NR_COMPOSEP11_2211(NR_PREMUL_112(s[0], alpha), a, d[0]);
1064                                         d[1] = NR_COMPOSEP11_2211(NR_PREMUL_112(s[1], alpha), a, d[1]);
1065                                         d[2] = NR_COMPOSEP11_2211(NR_PREMUL_112(s[2], alpha), a, d[2]);
1066                                 }
1067                                 /* a == 255 is impossible, because alpha < 255 */
1068                                 d += 3;
1069                                 s += 4;
1070                         }
1071                         px += rs;
1072                         spx += srs;
1073                 }
1074         }
1075 }
1076
1077 void
1078 nr_R8G8B8_R8G8B8_R8G8B8A8_N (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, unsigned int alpha)
1079 {
1080         unsigned int r, c;
1081
1082         if (alpha == 0) {
1083                 /* NOP */
1084         } else if (alpha == 255) {
1085                 for (r = h; r > 0; r--) {
1086                         unsigned char *d = px;
1087                         const unsigned char *s = spx;
1088                         for (c = w; c > 0; c--) {
1089                                 //   dc' = (1 - alpha*sa) * dc + alpha*sa*sc = (1 - sa) * dc + sa*sc
1090                                 if (s[3] == 0) {
1091                                         /* NOP */
1092                                 } else if (s[3] == 255) {
1093                                         d[0] = s[0];
1094                                         d[1] = s[1];
1095                                         d[2] = s[2];
1096                                 } else {
1097                                         d[0] = NR_COMPOSEN11_1111(s[0], s[3], d[0]);
1098                                         d[1] = NR_COMPOSEN11_1111(s[1], s[3], d[1]);
1099                                         d[2] = NR_COMPOSEN11_1111(s[2], s[3], d[2]);
1100                                 }
1101                                 d += 3;
1102                                 s += 4;
1103                         }
1104                         px += rs;
1105                         spx += srs;
1106                 }
1107         } else {
1108                 for (r = h; r > 0; r--) {
1109                         unsigned char *d = px;
1110                         const unsigned char *s = spx;
1111                         for (c = w; c > 0; c--) {
1112                                 unsigned int a;
1113                                 a = NR_PREMUL_112(s[3], alpha);
1114                                 //   dc' = (1 - alpha*sa) * dc + alpha*sa*sc
1115                                 if (a == 0) {
1116                                         /* NOP */
1117                                 } else {
1118                                         d[0] = NR_COMPOSEN11_1211(s[0], a, d[0]);
1119                                         d[1] = NR_COMPOSEN11_1211(s[1], a, d[1]);
1120                                         d[2] = NR_COMPOSEN11_1211(s[2], a, d[2]);
1121                                 }
1122                                 /* a == 255 is impossible, because alpha < 255 */
1123                                 d += 3;
1124                                 s += 4;
1125                         }
1126                         px += rs;
1127                         spx += srs;
1128                 }
1129         }
1130 }
1131
1132 void
1133 nr_R8G8B8_R8G8B8_R8G8B8A8_P_A8 (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, const unsigned char *mpx, int mrs)
1134 {
1135         unsigned int x, y;
1136
1137         for (y = h; y > 0; y--) {
1138                 unsigned char* d = px;
1139                 const unsigned char* s = spx;
1140                 const unsigned char* m = mpx;
1141                 for (x = w; x > 0; x--) {
1142                         unsigned int a;
1143                         a = NR_PREMUL_112(s[3], m[0]);
1144                         if (a == 0) {
1145                                 /* NOP */
1146                         } else if (a == 255*255) {
1147                                 memcpy(d, s, 3);
1148                         } else {
1149                                 //   dc' = (1 - m*sa) * dc + m*sc
1150                                 d[0] = NR_COMPOSEP11_2211(NR_PREMUL_112(s[0], m[0]), a, d[0]);
1151                                 d[1] = NR_COMPOSEP11_2211(NR_PREMUL_112(s[1], m[0]), a, d[1]);
1152                                 d[2] = NR_COMPOSEP11_2211(NR_PREMUL_112(s[2], m[0]), a, d[2]);
1153                         }
1154                         d += 3;
1155                         s += 4;
1156                         m += 1;
1157                 }
1158                 px += rs;
1159                 spx += srs;
1160                 mpx += mrs;
1161         }
1162 }
1163
1164 void
1165 nr_R8G8B8_R8G8B8_R8G8B8A8_N_A8 (unsigned char *px, int w, int h, int rs, const unsigned char *spx, int srs, const unsigned char *mpx, int mrs)
1166 {
1167         unsigned int x, y;
1168
1169         for (y = h; y > 0; y--) {
1170                 unsigned char* d = px;
1171                 const unsigned char* s = spx;
1172                 const unsigned char* m = mpx;
1173                 for (x = w; x > 0; x--) {
1174                         unsigned int a;
1175                         a = NR_PREMUL_112(s[3], m[0]);
1176                         if (a == 0) {
1177                                 /* NOP */
1178                         } else if (a == 255*255) {
1179                                 memcpy(d, s, 3);
1180                         } else {
1181                                 //   dc' = (1 - m*sa) * dc + m*sa*sc
1182                                 d[0] = NR_COMPOSEN11_1211(s[0], a, d[0]);
1183                                 d[1] = NR_COMPOSEN11_1211(s[1], a, d[1]);
1184                                 d[2] = NR_COMPOSEN11_1211(s[2], a, d[2]);
1185                         }
1186                         d += 3;
1187                         s += 4;
1188                         m += 1;
1189                 }
1190                 px += rs;
1191                 spx += srs;
1192                 mpx += mrs;
1193         }
1194 }
1195
1196