src/admame.c

   1 /*
   2  * This file is part of the Advance project.
   3  *
   4  * Copyright (C) 1999-2002 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  * In addition, as a special exception, Andrea Mazzoleni
  21  * gives permission to link the code of this program with
  22  * the MAME library (or with modified versions of MAME that use the
  23  * same license as MAME), and distribute linked combinations including
  24  * the two.  You must obey the GNU General Public License in all
  25  * respects for all of the code used other than MAME.  If you modify
  26  * this file, you may extend this exception to your version of the
  27  * file, but you are not obligated to do so.  If you do not wish to
  28  * do so, delete this exception statement from your version.
  29  */
  30
  31 /*
  32  * Alternatively at the previous license terms, you are allowed to use this
  33  * code in your program with these conditions:
  34  * - the program is not used in commercial activities.
  35  * - the whole source code of the program is released with the binary.
  36  */
  37
  38 #include "System.h"
  39
  40 #ifdef MMX
  41 extern "C" bool cpu_mmx;
  42 #endif
  43
  44 static void internal_scale2x_16_def(u16 *dst, const u16* src0, const u16* src1, const u16* src2, unsigned count) {
  45   /* first pixel */
  46   dst[0] = src1[0];
  47   if (src1[1] == src0[0] && src2[0] != src0[0])
  48     dst[1] = src0[0];
  49   else
  50     dst[1] = src1[0];
  51   ++src0;
  52   ++src1;
  53   ++src2;
  54   dst += 2;
  55
  56   /* central pixels */
  57   count -= 2;
  58   while (count) {
  59     if (src0[0] != src2[0] && src1[-1] != src1[1]) {
  60       dst[0] = src1[-1] == src0[0] ? src0[0] : src1[0];
  61       dst[1] = src1[1] == src0[0] ? src0[0] : src1[0];
  62     } else {
  63       dst[0] = src1[0];
  64       dst[1] = src1[0];
  65     }
  66
  67     ++src0;
  68     ++src1;
  69     ++src2;
  70     dst += 2;
  71     --count;
  72   }
  73
  74   /* last pixel */
  75   if (src1[-1] == src0[0] && src2[0] != src0[0])
  76     dst[0] = src0[0];
  77   else
  78     dst[0] = src1[0];
  79   dst[1] = src1[0];
  80 }
  81
  82 static void internal_scale2x_32_def(u32* dst,
  83                                     const u32* src0,
  84                                     const u32* src1,
  85                                     const u32* src2,
  86                                     unsigned count)
  87 {
  88   /* first pixel */
  89   dst[0] = src1[0];
  90   if (src1[1] == src0[0] && src2[0] != src0[0])
  91     dst[1] = src0[0];
  92   else
  93     dst[1] = src1[0];
  94   ++src0;
  95   ++src1;
  96   ++src2;
  97   dst += 2;
  98
  99   /* central pixels */
 100   count -= 2;
 101   while (count) {
 102     if (src0[0] != src2[0] && src1[-1] != src1[1]) {
 103       dst[0] = src1[-1] == src0[0] ? src0[0] : src1[0];
 104       dst[1] = src1[1] == src0[0] ? src0[0] : src1[0];
 105     } else {
 106       dst[0] = src1[0];
 107       dst[1] = src1[0];
 108     }
 109
 110     ++src0;
 111     ++src1;
 112     ++src2;
 113     dst += 2;
 114     --count;
 115   }
 116
 117   /* last pixel */
 118   if (src1[-1] == src0[0] && src2[0] != src0[0])
 119     dst[0] = src0[0];
 120   else
 121     dst[0] = src1[0];
 122   dst[1] = src1[0];
 123 }
 124
 125 #ifdef MMX
 126 static void internal_scale2x_16_mmx_single(u16* dst, const u16* src0, const u16* src1, const u16* src2, unsigned count) {
 127   /* always do the first and last run */
 128   count -= 2*4;
 129
 130 #ifdef __GNUC__
 131   __asm__ __volatile__(
 132                        /* first run */
 133                        /* set the current, current_pre, current_next registers */
 134                        "movq 0(%1), %%mm0\n"
 135                        "movq 0(%1),%%mm7\n"
 136                        "movq 8(%1),%%mm1\n"
 137                        "psllq $48,%%mm0\n"
 138                        "psllq $48,%%mm1\n"
 139                        "psrlq $48, %%mm0\n"
 140                        "movq %%mm7,%%mm2\n"
 141                        "movq %%mm7,%%mm3\n"
 142                        "psllq $16,%%mm2\n"
 143                        "psrlq $16,%%mm3\n"
 144                        "por %%mm2,%%mm0\n"
 145                        "por %%mm3,%%mm1\n"
 146
 147                        /* current_upper */
 148                        "movq (%0),%%mm6\n"
 149
 150                        /* compute the upper-left pixel for dst on %%mm2 */
 151                        /* compute the upper-right pixel for dst on %%mm4 */
 152                        "movq %%mm0,%%mm2\n"
 153                        "movq %%mm1,%%mm4\n"
 154                        "movq %%mm0,%%mm3\n"
 155                        "movq %%mm1,%%mm5\n"
 156                        "pcmpeqw %%mm6,%%mm2\n"
 157                        "pcmpeqw %%mm6,%%mm4\n"
 158                        "pcmpeqw (%2),%%mm3\n"
 159                        "pcmpeqw (%2),%%mm5\n"
 160                        "pandn %%mm2,%%mm3\n"
 161                        "pandn %%mm4,%%mm5\n"
 162                        "movq %%mm0,%%mm2\n"
 163                        "movq %%mm1,%%mm4\n"
 164                        "pcmpeqw %%mm1,%%mm2\n"
 165                        "pcmpeqw %%mm0,%%mm4\n"
 166                        "pandn %%mm3,%%mm2\n"
 167                        "pandn %%mm5,%%mm4\n"
 168                        "movq %%mm2,%%mm3\n"
 169                        "movq %%mm4,%%mm5\n"
 170                        "pand %%mm6,%%mm2\n"
 171                        "pand %%mm6,%%mm4\n"
 172                        "pandn %%mm7,%%mm3\n"
 173                        "pandn %%mm7,%%mm5\n"
 174                        "por %%mm3,%%mm2\n"
 175                        "por %%mm5,%%mm4\n"
 176
 177                        /* set *dst */
 178                        "movq %%mm2,%%mm3\n"
 179                        "punpcklwd %%mm4,%%mm2\n"
 180                        "punpckhwd %%mm4,%%mm3\n"
 181                        "movq %%mm2,(%3)\n"
 182                        "movq %%mm3,8(%3)\n"
 183
 184                        /* next */
 185                        "addl $8,%0\n"
 186                        "addl $8,%1\n"
 187                        "addl $8,%2\n"
 188                        "addl $16,%3\n"
 189
 190                        /* central runs */
 191                        "shrl $2,%4\n"
 192                        "jz 1f\n"
 193
 194                        "0:\n"
 195
 196                        /* set the current, current_pre, current_next registers */
 197                        "movq -8(%1),%%mm0\n"
 198                        "movq (%1),%%mm7\n"
 199                        "movq 8(%1),%%mm1\n"
 200                        "psrlq $48,%%mm0\n"
 201                        "psllq $48,%%mm1\n"
 202                        "movq %%mm7,%%mm2\n"
 203                        "movq %%mm7,%%mm3\n"
 204                        "psllq $16,%%mm2\n"
 205                        "psrlq $16,%%mm3\n"
 206                        "por %%mm2,%%mm0\n"
 207                        "por %%mm3,%%mm1\n"
 208
 209                        /* current_upper */
 210                        "movq (%0),%%mm6\n"
 211
 212                        /* compute the upper-left pixel for dst on %%mm2 */
 213                        /* compute the upper-right pixel for dst on %%mm4 */
 214                        "movq %%mm0,%%mm2\n"
 215                        "movq %%mm1,%%mm4\n"
 216                        "movq %%mm0,%%mm3\n"
 217                        "movq %%mm1,%%mm5\n"
 218                        "pcmpeqw %%mm6,%%mm2\n"
 219                        "pcmpeqw %%mm6,%%mm4\n"
 220                        "pcmpeqw (%2),%%mm3\n"
 221                        "pcmpeqw (%2),%%mm5\n"
 222                        "pandn %%mm2,%%mm3\n"
 223                        "pandn %%mm4,%%mm5\n"
 224                        "movq %%mm0,%%mm2\n"
 225                        "movq %%mm1,%%mm4\n"
 226                        "pcmpeqw %%mm1,%%mm2\n"
 227                        "pcmpeqw %%mm0,%%mm4\n"
 228                        "pandn %%mm3,%%mm2\n"
 229                        "pandn %%mm5,%%mm4\n"
 230                        "movq %%mm2,%%mm3\n"
 231                        "movq %%mm4,%%mm5\n"
 232                        "pand %%mm6,%%mm2\n"
 233                        "pand %%mm6,%%mm4\n"
 234                        "pandn %%mm7,%%mm3\n"
 235                        "pandn %%mm7,%%mm5\n"
 236                        "por %%mm3,%%mm2\n"
 237                        "por %%mm5,%%mm4\n"
 238
 239                        /* set *dst */
 240                        "movq %%mm2,%%mm3\n"
 241                        "punpcklwd %%mm4,%%mm2\n"
 242                        "punpckhwd %%mm4,%%mm3\n"
 243                        "movq %%mm2,(%3)\n"
 244                        "movq %%mm3,8(%3)\n"
 245
 246                        /* next */
 247                        "addl $8,%0\n"
 248                        "addl $8,%1\n"
 249                        "addl $8,%2\n"
 250                        "addl $16,%3\n"
 251
 252                        "decl %4\n"
 253                        "jnz 0b\n"
 254                        "1:\n"
 255
 256                        /* final run */
 257                        /* set the current, current_pre, current_next registers */
 258                        "movq (%1),%%mm1\n"
 259                        "movq (%1),%%mm7\n"
 260                        "movq -8(%1),%%mm0\n"
 261                        "psrlq $48,%%mm1\n"
 262                        "psrlq $48,%%mm0\n"
 263                        "psllq $48,%%mm1\n"
 264                        "movq %%mm7,%%mm2\n"
 265                        "movq %%mm7,%%mm3\n"
 266                        "psllq $16,%%mm2\n"
 267                        "psrlq $16,%%mm3\n"
 268                        "por %%mm2,%%mm0\n"
 269                        "por %%mm3,%%mm1\n"
 270
 271                        /* current_upper */
 272                        "movq (%0),%%mm6\n"
 273
 274                        /* compute the upper-left pixel for dst on %%mm2 */
 275                        /* compute the upper-right pixel for dst on %%mm4 */
 276                        "movq %%mm0,%%mm2\n"
 277                        "movq %%mm1,%%mm4\n"
 278                        "movq %%mm0,%%mm3\n"
 279                        "movq %%mm1,%%mm5\n"
 280                        "pcmpeqw %%mm6,%%mm2\n"
 281                        "pcmpeqw %%mm6,%%mm4\n"
 282                        "pcmpeqw (%2),%%mm3\n"
 283                        "pcmpeqw (%2),%%mm5\n"
 284                        "pandn %%mm2,%%mm3\n"
 285                        "pandn %%mm4,%%mm5\n"
 286                        "movq %%mm0,%%mm2\n"
 287                        "movq %%mm1,%%mm4\n"
 288                        "pcmpeqw %%mm1,%%mm2\n"
 289                        "pcmpeqw %%mm0,%%mm4\n"
 290                        "pandn %%mm3,%%mm2\n"
 291                        "pandn %%mm5,%%mm4\n"
 292                        "movq %%mm2,%%mm3\n"
 293                        "movq %%mm4,%%mm5\n"
 294                        "pand %%mm6,%%mm2\n"
 295                        "pand %%mm6,%%mm4\n"
 296                        "pandn %%mm7,%%mm3\n"
 297                        "pandn %%mm7,%%mm5\n"
 298                        "por %%mm3,%%mm2\n"
 299                        "por %%mm5,%%mm4\n"
 300
 301                        /* set *dst */
 302                        "movq %%mm2,%%mm3\n"
 303                        "punpcklwd %%mm4,%%mm2\n"
 304                        "punpckhwd %%mm4,%%mm3\n"
 305                        "movq %%mm2,(%3)\n"
 306                        "movq %%mm3,8(%3)\n"
 307                        "emms\n"
 308
 309                        : "+r" (src0), "+r" (src1), "+r" (src2), "+r" (dst), "+r" (count)
 310                        :
 311                        : "cc"
 312                        );
 313 #else
 314   __asm {
 315     mov eax, src0;
 316     mov ebx, src1;
 317     mov ecx, src2;
 318     mov edx, dst;
 319     mov esi, count;
 320
 321     /* first run */
 322     /* set the current, current_pre, current_next registers */
 323     movq mm0, qword ptr [ebx];
 324     movq mm7, qword ptr [ebx];
 325     movq mm1, qword ptr [ebx + 8];
 326     psllq mm0, 48;
 327     psllq mm1, 48;
 328     psrlq mm0, 48;
 329     movq mm2, mm7;
 330     movq mm3, mm7;
 331     psllq mm2, 16;
 332     psrlq mm3, 16;
 333     por mm0, mm2;
 334     por mm1, mm3;
 335
 336     /* current_upper */
 337     movq mm6, qword ptr [eax];
 338
 339     /* compute the upper-left pixel for dst on %%mm2 */
 340     /* compute the upper-right pixel for dst on %%mm4 */
 341     movq mm2, mm0;
 342     movq mm4, mm1;
 343     movq mm3, mm0;
 344     movq mm5, mm1;
 345     pcmpeqw mm2, mm6;
 346     pcmpeqw mm4, mm6;
 347     pcmpeqw mm3, qword ptr [ecx];
 348     pcmpeqw mm5, qword ptr [ecx];
 349     pandn mm3,mm2;
 350     pandn mm5,mm4;
 351     movq mm2,mm0;
 352     movq mm4,mm1;
 353     pcmpeqw mm2,mm1;
 354     pcmpeqw mm4,mm0;
 355     pandn mm2,mm3;
 356     pandn mm4,mm5;
 357     movq mm3,mm2;
 358     movq mm5,mm4;
 359     pand mm2,mm6;
 360     pand mm4,mm6;
 361     pandn mm3,mm7;
 362     pandn mm5,mm7;
 363     por mm2,mm3;
 364     por mm4,mm5;
 365
 366     /* set *dst0 */
 367     movq mm3,mm2;
 368     punpcklwd mm2,mm4;
 369     punpckhwd mm3,mm4;
 370     movq qword ptr [edx], mm2;
 371     movq qword ptr [edx + 8], mm3;
 372
 373     /* next */
 374     add eax, 8;
 375     add ebx, 8;
 376     add ecx, 8;
 377     add edx, 16;
 378
 379     /* central runs */
 380     shr esi, 2;
 381     jz label1;
 382     align 4;
 383   label0:
 384
 385     /* set the current, current_pre, current_next registers */
 386     movq mm0, qword ptr [ebx-8];
 387     movq mm7, qword ptr [ebx];
 388     movq mm1, qword ptr [ebx+8];
 389     psrlq mm0,48;
 390     psllq mm1,48;
 391     movq mm2,mm7;
 392     movq mm3,mm7;
 393     psllq mm2,16;
 394     psrlq mm3,16;
 395     por mm0,mm2;
 396     por mm1,mm3;
 397
 398     /* current_upper */
 399     movq mm6, qword ptr [eax];
 400
 401     /* compute the upper-left pixel for dst on %%mm2 */
 402     /* compute the upper-right pixel for dst on %%mm4 */
 403     movq mm2,mm0;
 404     movq mm4,mm1;
 405     movq mm3,mm0;
 406     movq mm5,mm1;
 407     pcmpeqw mm2,mm6;
 408     pcmpeqw mm4,mm6;
 409     pcmpeqw mm3, qword ptr [ecx];
 410     pcmpeqw mm5, qword ptr [ecx];
 411     pandn mm3,mm2;
 412     pandn mm5,mm4;
 413     movq mm2,mm0;
 414     movq mm4,mm1;
 415     pcmpeqw mm2,mm1;
 416     pcmpeqw mm4,mm0;
 417     pandn mm2,mm3;
 418     pandn mm4,mm5;
 419     movq mm3,mm2;
 420     movq mm5,mm4;
 421     pand mm2,mm6;
 422     pand mm4,mm6;
 423     pandn mm3,mm7;
 424     pandn mm5,mm7;
 425     por mm2,mm3;
 426     por mm4,mm5;
 427
 428     /* set *dst */
 429     movq mm3,mm2;
 430     punpcklwd mm2,mm4;
 431     punpckhwd mm3,mm4;
 432     movq qword ptr [edx], mm2;
 433     movq qword ptr [edx+8], mm3;
 434
 435     /* next */
 436     add eax,8;
 437     add ebx,8;
 438     add ecx,8;
 439     add edx,16;
 440
 441     dec esi;
 442     jnz label0;
 443   label1:
 444
 445     /* final run */
 446     /* set the current, current_pre, current_next registers */
 447     movq mm1, qword ptr [ebx];
 448     movq mm7, qword ptr [ebx];
 449     movq mm0, qword ptr [ebx-8];
 450     psrlq mm1,48;
 451     psrlq mm0,48;
 452     psllq mm1,48;
 453     movq mm2,mm7;
 454     movq mm3,mm7;
 455     psllq mm2,16;
 456     psrlq mm3,16;
 457     por mm0,mm2;
 458     por mm1,mm3;
 459
 460     /* current_upper */
 461     movq mm6, qword ptr [eax];
 462
 463     /* compute the upper-left pixel for dst on %%mm2 */
 464     /* compute the upper-right pixel for dst on %%mm4 */
 465     movq mm2,mm0;
 466     movq mm4,mm1;
 467     movq mm3,mm0;
 468     movq mm5,mm1;
 469     pcmpeqw mm2,mm6;
 470     pcmpeqw mm4,mm6;
 471     pcmpeqw mm3, qword ptr [ecx];
 472     pcmpeqw mm5, qword ptr [ecx];
 473     pandn mm3,mm2;
 474     pandn mm5,mm4;
 475     movq mm2,mm0;
 476     movq mm4,mm1;
 477     pcmpeqw mm2,mm1;
 478     pcmpeqw mm4,mm0;
 479     pandn mm2,mm3;
 480     pandn mm4,mm5;
 481     movq mm3,mm2;
 482     movq mm5,mm4;
 483     pand mm2,mm6;
 484     pand mm4,mm6;
 485     pandn mm3,mm7;
 486     pandn mm5,mm7;
 487     por mm2,mm3;
 488     por mm4,mm5;
 489
 490     /* set *dst */
 491     movq mm3,mm2;
 492     punpcklwd mm2,mm4;
 493     punpckhwd mm3,mm4;
 494     movq qword ptr [edx], mm2;
 495     movq qword ptr [edx+8], mm3;
 496
 497     mov src0, eax;
 498     mov src1, ebx;
 499     mov src2, ecx;
 500     mov dst, edx;
 501     mov count, esi;
 502
 503     emms;
 504   }
 505 #endif
 506 }
 507
 508 static void internal_scale2x_32_mmx_single(u32* dst, const u32* src0, const u32* src1, const u32* src2, unsigned count) {
 509   /* always do the first and last run */
 510   count -= 2*2;
 511
 512 #ifdef __GNUC__
 513   __asm__ __volatile__(
 514                        /* first run */
 515                        /* set the current, current_pre, current_next registers */
 516                        "movq 0(%1),%%mm0\n"
 517                        "movq 0(%1),%%mm7\n"
 518                        "movq 8(%1),%%mm1\n"
 519                        "psllq $32,%%mm0\n"
 520                        "psllq $32,%%mm1\n"
 521                        "psrlq $32,%%mm0\n"
 522                        "movq %%mm7,%%mm2\n"
 523                        "movq %%mm7,%%mm3\n"
 524                        "psllq $32,%%mm2\n"
 525                        "psrlq $32,%%mm3\n"
 526                        "por %%mm2,%%mm0\n"
 527                        "por %%mm3,%%mm1\n"
 528
 529                        /* current_upper */
 530                        "movq (%0),%%mm6\n"
 531
 532                        /* compute the upper-left pixel for dst on %%mm2 */
 533                        /* compute the upper-right pixel for dst on %%mm4 */
 534                        "movq %%mm0,%%mm2\n"
 535                        "movq %%mm1,%%mm4\n"
 536                        "movq %%mm0,%%mm3\n"
 537                        "movq %%mm1,%%mm5\n"
 538                        "pcmpeqd %%mm6,%%mm2\n"
 539                        "pcmpeqd %%mm6,%%mm4\n"
 540                        "pcmpeqd (%2),%%mm3\n"
 541                        "pcmpeqd (%2),%%mm5\n"
 542                        "pandn %%mm2,%%mm3\n"
 543                        "pandn %%mm4,%%mm5\n"
 544                        "movq %%mm0,%%mm2\n"
 545                        "movq %%mm1,%%mm4\n"
 546                        "pcmpeqd %%mm1,%%mm2\n"
 547                        "pcmpeqd %%mm0,%%mm4\n"
 548                        "pandn %%mm3,%%mm2\n"
 549                        "pandn %%mm5,%%mm4\n"
 550                        "movq %%mm2,%%mm3\n"
 551                        "movq %%mm4,%%mm5\n"
 552                        "pand %%mm6,%%mm2\n"
 553                        "pand %%mm6,%%mm4\n"
 554                        "pandn %%mm7,%%mm3\n"
 555                        "pandn %%mm7,%%mm5\n"
 556                        "por %%mm3,%%mm2\n"
 557                        "por %%mm5,%%mm4\n"
 558
 559                        /* set *dst */
 560                        "movq %%mm2,%%mm3\n"
 561                        "punpckldq %%mm4,%%mm2\n"
 562                        "punpckhdq %%mm4,%%mm3\n"
 563                        "movq %%mm2,(%3)\n"
 564                        "movq %%mm3, 8(%3)\n"
 565
 566                        /* next */
 567                        "addl $8,%0\n"
 568                        "addl $8,%1\n"
 569                        "addl $8,%2\n"
 570                        "addl $16,%3\n"
 571
 572                        /* central runs */
 573                        "shrl $1,%4\n"
 574                        "jz 1f\n"
 575
 576                        "0:\n"
 577
 578                        /* set the current, current_pre, current_next registers */
 579                        "movq -8(%1),%%mm0\n"
 580                        "movq (%1),%%mm7\n"
 581                        "movq 8(%1),%%mm1\n"
 582                        "psrlq $32,%%mm0\n"
 583                        "psllq $32,%%mm1\n"
 584                        "movq %%mm7,%%mm2\n"
 585                        "movq %%mm7,%%mm3\n"
 586                        "psllq $32,%%mm2\n"
 587                        "psrlq $32,%%mm3\n"
 588                        "por %%mm2,%%mm0\n"
 589                        "por %%mm3,%%mm1\n"
 590
 591                        /* current_upper */
 592                        "movq (%0),%%mm6\n"
 593
 594                        /* compute the upper-left pixel for dst on %%mm2 */
 595                        /* compute the upper-right pixel for dst on %%mm4 */
 596                        "movq %%mm0,%%mm2\n"
 597                        "movq %%mm1,%%mm4\n"
 598                        "movq %%mm0,%%mm3\n"
 599                        "movq %%mm1,%%mm5\n"
 600                        "pcmpeqd %%mm6,%%mm2\n"
 601                        "pcmpeqd %%mm6,%%mm4\n"
 602                        "pcmpeqd (%2),%%mm3\n"
 603                        "pcmpeqd (%2),%%mm5\n"
 604                        "pandn %%mm2,%%mm3\n"
 605                        "pandn %%mm4,%%mm5\n"
 606                        "movq %%mm0,%%mm2\n"
 607                        "movq %%mm1,%%mm4\n"
 608                        "pcmpeqd %%mm1,%%mm2\n"
 609                        "pcmpeqd %%mm0,%%mm4\n"
 610                        "pandn %%mm3,%%mm2\n"
 611                        "pandn %%mm5,%%mm4\n"
 612                        "movq %%mm2,%%mm3\n"
 613                        "movq %%mm4,%%mm5\n"
 614                        "pand %%mm6,%%mm2\n"
 615                        "pand %%mm6,%%mm4\n"
 616                        "pandn %%mm7,%%mm3\n"
 617                        "pandn %%mm7,%%mm5\n"
 618                        "por %%mm3,%%mm2\n"
 619                        "por %%mm5,%%mm4\n"
 620
 621                        /* set *dst */
 622                        "movq %%mm2,%%mm3\n"
 623                        "punpckldq %%mm4,%%mm2\n"
 624                        "punpckhdq %%mm4,%%mm3\n"
 625                        "movq %%mm2,(%3)\n"
 626                        "movq %%mm3,8(%3)\n"
 627
 628                        /* next */
 629                        "addl $8,%0\n"
 630                        "addl $8,%1\n"
 631                        "addl $8,%2\n"
 632                        "addl $16,%3\n"
 633
 634                        "decl %4\n"
 635                        "jnz 0b\n"
 636                        "1:\n"
 637
 638                        /* final run */
 639                        /* set the current, current_pre, current_next registers */
 640                        "movq (%1),%%mm1\n"
 641                        "movq (%1),%%mm7\n"
 642                        "movq -8(%1), %%mm0\n"
 643                        "psrlq $32,%%mm1\n"
 644                        "psrlq $32,%%mm0\n"
 645                        "psllq $32,%%mm1\n"
 646                        "movq %%mm7,%%mm2\n"
 647                        "movq %%mm7,%%mm3\n"
 648                        "psllq $32,%%mm2\n"
 649                        "psrlq $32,%%mm3\n"
 650                        "por %%mm2,%%mm0\n"
 651                        "por %%mm3,%%mm1\n"
 652
 653                        /* current_upper */
 654                        "movq (%0),%%mm6\n"
 655
 656                        /* compute the upper-left pixel for dst on %%mm2 */
 657                        /* compute the upper-right pixel for dst on %%mm4 */
 658                        "movq %%mm0,%%mm2\n"
 659                        "movq %%mm1,%%mm4\n"
 660                        "movq %%mm0,%%mm3\n"
 661                        "movq %%mm1,%%mm5\n"
 662                        "pcmpeqd %%mm6,%%mm2\n"
 663                        "pcmpeqd %%mm6,%%mm4\n"
 664                        "pcmpeqd (%2),%%mm3\n"
 665                        "pcmpeqd (%2),%%mm5\n"
 666                        "pandn %%mm2,%%mm3\n"
 667                        "pandn %%mm4,%%mm5\n"
 668                        "movq %%mm0,%%mm2\n"
 669                        "movq %%mm1,%%mm4\n"
 670                        "pcmpeqd %%mm1,%%mm2\n"
 671                        "pcmpeqd %%mm0,%%mm4\n"
 672                        "pandn %%mm3,%%mm2\n"
 673                        "pandn %%mm5,%%mm4\n"
 674                        "movq %%mm2,%%mm3\n"
 675                        "movq %%mm4,%%mm5\n"
 676                        "pand %%mm6,%%mm2\n"
 677                        "pand %%mm6,%%mm4\n"
 678                        "pandn %%mm7,%%mm3\n"
 679                        "pandn %%mm7,%%mm5\n"
 680                        "por %%mm3,%%mm2\n"
 681                        "por %%mm5,%%mm4\n"
 682
 683                        /* set *dst */
 684                        "movq %%mm2,%%mm3\n"
 685                        "punpckldq %%mm4,%%mm2\n"
 686                        "punpckhdq %%mm4,%%mm3\n"
 687                        "movq %%mm2,(%3)\n"
 688                        "movq %%mm3,8(%3)\n"
 689                        "emms\n"
 690
 691                        : "+r" (src0), "+r" (src1), "+r" (src2), "+r" (dst), "+r" (count)
 692                        :
 693                        : "cc"
 694                        );
 695 #else
 696   __asm {
 697     mov eax, src0;
 698     mov ebx, src1;
 699     mov ecx, src2;
 700     mov edx, dst;
 701     mov esi, count;
 702
 703     /* first run */
 704     /* set the current, current_pre, current_next registers */
 705     movq mm0,qword ptr [ebx];
 706     movq mm7,qword ptr [ebx];
 707     movq mm1,qword ptr [ebx + 8];
 708     psllq mm0,32;
 709     psllq mm1,32;
 710     psrlq mm0,32;
 711     movq mm2,mm7;
 712     movq mm3,mm7;
 713     psllq mm2,32;
 714     psrlq mm3,32;
 715     por mm0,mm2;
 716     por mm1,mm3;
 717
 718     /* current_upper */
 719     movq mm6,qword ptr [eax];
 720
 721     /* compute the upper-left pixel for dst on %%mm2 */
 722     /* compute the upper-right pixel for dst on %%mm4 */
 723     movq mm2,mm0;
 724     movq mm4,mm1;
 725     movq mm3,mm0;
 726     movq mm5,mm1;
 727     pcmpeqd mm2,mm6;
 728     pcmpeqd mm4,mm6;
 729     pcmpeqd mm3,qword ptr [ecx];
 730     pcmpeqd mm5,qword ptr [ecx];
 731     pandn mm3,mm2;
 732     pandn mm5,mm4;
 733     movq mm2,mm0;
 734     movq mm4,mm1;
 735     pcmpeqd mm2,mm1;
 736     pcmpeqd mm4,mm0;
 737     pandn mm2,mm3;
 738     pandn mm4,mm5;
 739     movq mm3,mm2;
 740     movq mm5,mm4;
 741     pand mm2,mm6;
 742     pand mm4,mm6;
 743     pandn mm3,mm7;
 744     pandn mm5,mm7;
 745     por mm2,mm3;
 746     por mm4,mm5;
 747
 748     /* set *dst */
 749     movq mm3,mm2;
 750     punpckldq mm2,mm4;
 751     punpckhdq mm3,mm4;
 752     movq qword ptr [edx],mm2;
 753     movq qword ptr [edx+8],mm3;
 754
 755     /* next */
 756     add eax,8;
 757     add ebx,8;
 758     add ecx,8;
 759     add edx,16;
 760
 761     /* central runs */
 762     shr esi,1;
 763     jz label1;
 764 label0:
 765
 766   /* set the current, current_pre, current_next registers */
 767     movq mm0,qword ptr [ebx-8];
 768     movq mm7,qword ptr [ebx];
 769     movq mm1,qword ptr [ebx+8];
 770     psrlq mm0,32;
 771     psllq mm1,32;
 772     movq mm2,mm7;
 773     movq mm3,mm7;
 774     psllq mm2,32;
 775     psrlq mm3,32;
 776     por mm0,mm2;
 777     por mm1,mm3;
 778
 779     /* current_upper */
 780     movq mm6,qword ptr[eax];
 781
 782     /* compute the upper-left pixel for dst on %%mm2 */
 783     /* compute the upper-right pixel for dst on %%mm4 */
 784     movq mm2,mm0;
 785     movq mm4,mm1;
 786     movq mm3,mm0;
 787     movq mm5,mm1;
 788     pcmpeqd mm2,mm6;
 789     pcmpeqd mm4,mm6;
 790     pcmpeqd mm3,qword ptr[ecx];
 791     pcmpeqd mm5,qword ptr[ecx];
 792     pandn mm3,mm2;
 793     pandn mm5,mm4;
 794     movq mm2,mm0;
 795     movq mm4,mm1;
 796     pcmpeqd mm2,mm1;
 797     pcmpeqd mm4,mm0;
 798     pandn mm2,mm3;
 799     pandn mm4,mm5;
 800     movq mm3,mm2;
 801     movq mm5,mm4;
 802     pand mm2,mm6;
 803     pand mm4,mm6;
 804     pandn mm3,mm7;
 805     pandn mm5,mm7;
 806     por mm2,mm3;
 807     por mm4,mm5;
 808
 809     /* set *dst */
 810     movq mm3,mm2;
 811     punpckldq mm2,mm4;
 812     punpckhdq mm3,mm4;
 813     movq qword ptr [edx],mm2;
 814     movq qword ptr [edx+8],mm3;
 815
 816     /* next */
 817     add eax,8;
 818     add ebx,8;
 819     add ecx,8;
 820     add edx,16;
 821
 822     dec esi;
 823     jnz label0;
 824 label1:
 825
 826     /* final run */
 827     /* set the current, current_pre, current_next registers */
 828     movq mm1,qword ptr [ebx];
 829     movq mm7,qword ptr [ebx];
 830     movq mm0,qword ptr [ebx-8];
 831     psrlq mm1,32;
 832     psrlq mm0,32;
 833     psllq mm1,32;
 834     movq mm2,mm7;
 835     movq mm3,mm7;
 836     psllq mm2,32;
 837     psrlq mm3,32;
 838     por mm0,mm2;
 839     por mm1,mm3;
 840
 841     /* current_upper */
 842     movq mm6,qword ptr [eax];
 843
 844     /* compute the upper-left pixel for dst on %%mm2 */
 845     /* compute the upper-right pixel for dst on %%mm4 */
 846     movq mm2,mm0;
 847     movq mm4,mm1;
 848     movq mm3,mm0;
 849     movq mm5,mm1;
 850     pcmpeqd mm2,mm6;
 851     pcmpeqd mm4,mm6;
 852     pcmpeqd mm3,qword ptr [ecx];
 853     pcmpeqd mm5,qword ptr [ecx];
 854     pandn mm3,mm2;
 855     pandn mm5,mm4;
 856     movq mm2,mm0;
 857     movq mm4,mm1;
 858     pcmpeqd mm2,mm1;
 859     pcmpeqd mm4,mm0;
 860     pandn mm2,mm3;
 861     pandn mm4,mm5;
 862     movq mm3,mm2;
 863     movq mm5,mm4;
 864     pand mm2,mm6;
 865     pand mm4,mm6;
 866     pandn mm3,mm7;
 867     pandn mm5,mm7;
 868     por mm2,mm3;
 869     por mm4,mm5;
 870
 871     /* set *dst */
 872     movq mm3,mm2;
 873     punpckldq mm2,mm4;
 874     punpckhdq mm3,mm4;
 875     movq qword ptr [edx],mm2;
 876     movq qword ptr [edx+8],mm3;
 877
 878     mov src0, eax;
 879     mov src1, ebx;
 880     mov src2, ecx;
 881     mov dst, edx;
 882     mov count, esi;
 883
 884     emms;
 885   }
 886 #endif
 887 }
 888
 889 static void internal_scale2x_16_mmx(u16* dst0, u16* dst1, const u16* src0, const u16* src1, const u16* src2, unsigned count) {
 890   //    assert( count >= 2*4 );
 891   internal_scale2x_16_mmx_single(dst0, src0, src1, src2, count);
 892   internal_scale2x_16_mmx_single(dst1, src2, src1, src0, count);
 893 }
 894
 895 static void internal_scale2x_32_mmx(u32* dst0, u32* dst1, const u32* src0, const u32* src1, const u32* src2, unsigned count) {
 896   //    assert( count >= 2*2 );
 897   internal_scale2x_32_mmx_single(dst0, src0, src1, src2, count);
 898   internal_scale2x_32_mmx_single(dst1, src2, src1, src0, count);
 899 }
 900 #endif
 901
 902 void AdMame2x(u8 *srcPtr, u32 srcPitch, u8 * /* deltaPtr */,
 903               u8 *dstPtr, u32 dstPitch, int width, int height)
 904 {
 905   u16 *dst0 = (u16 *)dstPtr;
 906   u16 *dst1 = dst0 + (dstPitch >> 1);
 907
 908   u16 *src0 = (u16 *)srcPtr;
 909   u16 *src1 = src0 + (srcPitch >> 1);
 910   u16 *src2 = src1 + (srcPitch >> 1);
 911 #ifdef MMX
 912   if(cpu_mmx) {
 913     internal_scale2x_16_mmx(dst0, dst1, src0, src0, src1, width);
 914
 915     int count = height;
 916
 917     count -= 2;
 918     while(count) {
 919       dst0 += dstPitch;
 920       dst1 += dstPitch;
 921       internal_scale2x_16_mmx(dst0, dst1, src0, src1, src2, width);
 922       src0 = src1;
 923       src1 = src2;
 924       src2 += srcPitch >> 1;
 925       --count;
 926     }
 927     dst0 += dstPitch;
 928     dst1 += dstPitch;
 929     internal_scale2x_16_mmx(dst0, dst1, src0, src1, src1, width);
 930   } else {
 931 #endif
 932     internal_scale2x_16_def(dst0, src0, src0, src1, width);
 933     internal_scale2x_16_def(dst1, src1, src0, src0, width);
 934
 935     int count = height;
 936
 937     count -= 2;
 938     while(count) {
 939       dst0 += dstPitch;
 940       dst1 += dstPitch;
 941       internal_scale2x_16_def(dst0, src0, src1, src2, width);
 942       internal_scale2x_16_def(dst1, src2, src1, src0, width);
 943       src0 = src1;
 944       src1 = src2;
 945       src2 += srcPitch >> 1;
 946       --count;
 947     }
 948     dst0 += dstPitch;
 949     dst1 += dstPitch;
 950     internal_scale2x_16_def(dst0, src0, src1, src1, width);
 951     internal_scale2x_16_def(dst1, src1, src1, src0, width);
 952 #ifdef MMX
 953   }
 954 #endif
 955 }
 956
 957 void AdMame2x32(u8 *srcPtr, u32 srcPitch, u8 * /* deltaPtr */,
 958                 u8 *dstPtr, u32 dstPitch, int width, int height)
 959 {
 960   u32 *dst0 = (u32 *)dstPtr;
 961   u32 *dst1 = dst0 + (dstPitch >> 2);
 962
 963   u32 *src0 = (u32 *)srcPtr;
 964   u32 *src1 = src0 + (srcPitch >> 2);
 965   u32 *src2 = src1 + (srcPitch >> 2);
 966 #ifdef MMX
 967   if(cpu_mmx) {
 968     internal_scale2x_32_mmx(dst0, dst1, src0, src0, src1, width);
 969
 970     int count = height;
 971
 972     count -= 2;
 973     while(count) {
 974       dst0 += dstPitch >> 1;
 975       dst1 += dstPitch >> 1;
 976       internal_scale2x_32_mmx(dst0, dst1, src0, src1, src2, width);
 977       src0 = src1;
 978       src1 = src2;
 979       src2 += srcPitch >> 2;
 980       --count;
 981     }
 982     dst0 += dstPitch >> 1;
 983     dst1 += dstPitch >> 1;
 984     internal_scale2x_32_mmx(dst0, dst1, src0, src1, src1, width);
 985   } else {
 986 #endif
 987     internal_scale2x_32_def(dst0, src0, src0, src1, width);
 988     internal_scale2x_32_def(dst1, src1, src0, src0, width);
 989
 990     int count = height;
 991
 992     count -= 2;
 993     while(count) {
 994       dst0 += dstPitch >> 1;
 995       dst1 += dstPitch >> 1;
 996       internal_scale2x_32_def(dst0, src0, src1, src2, width);
 997       internal_scale2x_32_def(dst1, src2, src1, src0, width);
 998       src0 = src1;
 999       src1 = src2;
1000       src2 += srcPitch >> 2;
1001       --count;
1002     }
1003     dst0 += dstPitch >> 1;
1004     dst1 += dstPitch >> 1;
1005     internal_scale2x_32_def(dst0, src0, src1, src1, width);
1006     internal_scale2x_32_def(dst1, src1, src1, src0, width);
1007 #ifdef MMX
1008   }
1009 #endif
1010 }