[vba-commit] VisualBoyAdvance/src admame.cpp,1.9,1.10

[<for...@us...>]

Update of /cvsroot/vba/VisualBoyAdvance/src
In directory sc8-pr-cvs1.sourceforge.net:/tmp/cvs-serv19684

Modified Files:
	admame.cpp 
Log Message:
Sync code with Scale2x V2.0

Index: admame.cpp
===================================================================
RCS file: /cvsroot/vba/VisualBoyAdvance/src/admame.cpp,v
retrieving revision 1.9
retrieving revision 1.10
diff -C2 -d -r1.9 -r1.10
*** admame.cpp	20 Jan 2004 22:02:03 -0000	1.9
--- admame.cpp	24 Feb 2004 11:22:35 -0000	1.10
***************
*** 42,171 ****
  #endif
  
! /* Suggested in "Intel Optimization" for Pentium II */
! #define ASM_JUMP_ALIGN ".p2align 4\n"
! 
! static void internal_scale2x_16_def(u16 *dst0, u16* dst1, const u16* src0, const u16* src1, const u16* src2, unsigned count) {
    /* first pixel */
!   dst0[0] = src1[0];
!   dst1[0] = src1[0];
    if (src1[1] == src0[0] && src2[0] != src0[0])
!     dst0[1] =src0[0];
!   else
!     dst0[1] =src1[0];
!   if (src1[1] == src2[0] && src0[0] != src2[0])
!     dst1[1] =src2[0];
    else
!     dst1[1] =src1[0];
    ++src0;
    ++src1;
    ++src2;
!   dst0 += 2;
!   dst1 += 2;
  
    /* central pixels */
    count -= 2;
    while (count) {
!     if (src1[-1] == src0[0] && src2[0] != src0[0] && src1[1] != src0[0])
!       dst0[0] = src0[0];
!     else
!       dst0[0] = src1[0];
!     if (src1[1] == src0[0] && src2[0] != src0[0] && src1[-1] != src0[0])
!       dst0[1] =src0[0];
!     else
!       dst0[1] =src1[0];
! 
!     if (src1[-1] == src2[0] && src0[0] != src2[0] && src1[1] != src2[0])
!       dst1[0] =src2[0];
!     else
!       dst1[0] =src1[0];
!     if (src1[1] == src2[0] && src0[0] != src2[0] && src1[-1] != src2[0])
!       dst1[1] =src2[0];
!     else
!       dst1[1] =src1[0];
! 
      ++src0;
      ++src1;
      ++src2;
!     dst0 += 2;
!     dst1 += 2;
      --count;
    }
! 
    /* last pixel */
    if (src1[-1] == src0[0] && src2[0] != src0[0])
!     dst0[0] =src0[0];
!   else
!     dst0[0] =src1[0];
!   if (src1[-1] == src2[0] && src0[0] != src2[0])
!     dst1[0] =src2[0];
    else
!     dst1[0] =src1[0];
!   dst0[1] =src1[0];
!   dst1[1] =src1[0];
  }
  
! static void internal_scale2x_32_def(u32* dst0,
!                                     u32* dst1,
                                      const u32* src0,
                                      const u32* src1,
                                      const u32* src2,
!                                     unsigned count) {
    /* first pixel */
!   dst0[0] = src1[0];
!   dst1[0] = src1[0];
    if (src1[1] == src0[0] && src2[0] != src0[0])
!     dst0[1] = src0[0];
!   else
!     dst0[1] = src1[0];
!   if (src1[1] == src2[0] && src0[0] != src2[0])
!     dst1[1] = src2[0];
    else
!     dst1[1] = src1[0];
    ++src0;
    ++src1;
    ++src2;
!   dst0 += 2;
!   dst1 += 2;
! 
    /* central pixels */
    count -= 2;
    while (count) {
!     if (src1[-1] == src0[0] && src2[0] != src0[0] && src1[1] != src0[0])
!       dst0[0] = src0[0];
!     else
!       dst0[0] = src1[0];
!     if (src1[1] == src0[0] && src2[0] != src0[0] && src1[-1] != src0[0])
!       dst0[1] = src0[0];
!     else
!       dst0[1] = src1[0];
! 
!     if (src1[-1] == src2[0] && src0[0] != src2[0] && src1[1] != src2[0])
!       dst1[0] = src2[0];
!     else
!       dst1[0] = src1[0];
!     if (src1[1] == src2[0] && src0[0] != src2[0] && src1[-1] != src2[0])
!       dst1[1] = src2[0];
!     else
!       dst1[1] = src1[0];
! 
      ++src0;
      ++src1;
      ++src2;
!     dst0 += 2;
!     dst1 += 2;
      --count;
    }
! 
    /* last pixel */
    if (src1[-1] == src0[0] && src2[0] != src0[0])
!     dst0[0] = src0[0];
!   else
!     dst0[0] = src1[0];
!   if (src1[-1] == src2[0] && src0[0] != src2[0])
!     dst1[0] = src2[0];
    else
!     dst1[0] = src1[0];
!   dst0[1] = src1[0];
!   dst1[1] = src1[0];
  }
  
--- 42,124 ----
  #endif
  
! static void internal_scale2x_16_def(u16 *dst, const u16* src0, const u16* src1, const u16* src2, unsigned count) {
    /* first pixel */
!   dst[0] = src1[0];
    if (src1[1] == src0[0] && src2[0] != src0[0])
!     dst[1] = src0[0];
    else
!     dst[1] = src1[0];
    ++src0;
    ++src1;
    ++src2;
!   dst += 2;
  
    /* central pixels */
    count -= 2;
    while (count) {
!     if (src0[0] != src2[0] && src1[-1] != src1[1]) {
!       dst[0] = src1[-1] == src0[0] ? src0[0] : src1[0];
!       dst[1] = src1[1] == src0[0] ? src0[0] : src1[0];
!     } else {
!       dst[0] = src1[0];
!       dst[1] = src1[0];
!     }
!     
      ++src0;
      ++src1;
      ++src2;
!     dst += 2;
      --count;
    }
!   
    /* last pixel */
    if (src1[-1] == src0[0] && src2[0] != src0[0])
!     dst[0] = src0[0];
    else
!     dst[0] = src1[0];
!   dst[1] = src1[0];
  }
  
! static void internal_scale2x_32_def(u32* dst,
                                      const u32* src0,
                                      const u32* src1,
                                      const u32* src2,
!                                     unsigned count) 
! {
    /* first pixel */
!   dst[0] = src1[0];
    if (src1[1] == src0[0] && src2[0] != src0[0])
!     dst[1] = src0[0];
    else
!     dst[1] = src1[0];
    ++src0;
    ++src1;
    ++src2;
!   dst += 2;
!   
    /* central pixels */
    count -= 2;
    while (count) {
!     if (src0[0] != src2[0] && src1[-1] != src1[1]) {
!       dst[0] = src1[-1] == src0[0] ? src0[0] : src1[0];
!       dst[1] = src1[1] == src0[0] ? src0[0] : src1[0];
!     } else {
!       dst[0] = src1[0];
!       dst[1] = src1[0];
!     }
!     
      ++src0;
      ++src1;
      ++src2;
!     dst += 2;
      --count;
    }
!   
    /* last pixel */
    if (src1[-1] == src0[0] && src2[0] != src0[0])
!     dst[0] = src0[0];
    else
!     dst[0] = src1[0];
!   dst[1] = src1[0];
  }
  
***************
*** 179,187 ****
                         /* first run */
                         /* set the current, current_pre, current_next registers */
!                        "pxor %%mm0,%%mm0\n" /* use a fake black out of screen */
                         "movq 0(%1),%%mm7\n"
                         "movq 8(%1),%%mm1\n"
!                        "psrlq $48,%%mm0\n"
                         "psllq $48,%%mm1\n"
                         "movq %%mm7,%%mm2\n"
                         "movq %%mm7,%%mm3\n"
--- 132,141 ----
                         /* first run */
                         /* set the current, current_pre, current_next registers */
!                        "movq 0(%1), %%mm0\n"
                         "movq 0(%1),%%mm7\n"
                         "movq 8(%1),%%mm1\n"
!                        "psllq $48,%%mm0\n"
                         "psllq $48,%%mm1\n"
+                        "psrlq $48, %%mm0\n"
                         "movq %%mm7,%%mm2\n"
                         "movq %%mm7,%%mm3\n"
***************
*** 194,199 ****
                         "movq (%0),%%mm6\n"
  
!                        /* compute the upper-left pixel for dst0 on %%mm2 */
!                        /* compute the upper-right pixel for dst0 on %%mm4 */
                         "movq %%mm0,%%mm2\n"
                         "movq %%mm1,%%mm4\n"
--- 148,153 ----
                         "movq (%0),%%mm6\n"
  
!                        /* compute the upper-left pixel for dst on %%mm2 */
!                        /* compute the upper-right pixel for dst on %%mm4 */
                         "movq %%mm0,%%mm2\n"
                         "movq %%mm1,%%mm4\n"
***************
*** 221,225 ****
                         "por %%mm5,%%mm4\n"
  
!                        /* set *dst0 */
                         "movq %%mm2,%%mm3\n"
                         "punpcklwd %%mm4,%%mm2\n"
--- 175,179 ----
                         "por %%mm5,%%mm4\n"
  
!                        /* set *dst */
                         "movq %%mm2,%%mm3\n"
                         "punpcklwd %%mm4,%%mm2\n"
***************
*** 237,241 ****
                         "shrl $2,%4\n"
                         "jz 1f\n"
!                        ASM_JUMP_ALIGN
                         "0:\n"
  
--- 191,195 ----
                         "shrl $2,%4\n"
                         "jz 1f\n"
! 
                         "0:\n"
  
***************
*** 256,261 ****
                         "movq (%0),%%mm6\n"
  
!                        /* compute the upper-left pixel for dst0 on %%mm2 */
!                        /* compute the upper-right pixel for dst0 on %%mm4 */
                         "movq %%mm0,%%mm2\n"
                         "movq %%mm1,%%mm4\n"
--- 210,215 ----
                         "movq (%0),%%mm6\n"
  
!                        /* compute the upper-left pixel for dst on %%mm2 */
!                        /* compute the upper-right pixel for dst on %%mm4 */
                         "movq %%mm0,%%mm2\n"
                         "movq %%mm1,%%mm4\n"
***************
*** 283,287 ****
                         "por %%mm5,%%mm4\n"
  
!                        /* set *dst0 */
                         "movq %%mm2,%%mm3\n"
                         "punpcklwd %%mm4,%%mm2\n"
--- 237,241 ----
                         "por %%mm5,%%mm4\n"
  
!                        /* set *dst */
                         "movq %%mm2,%%mm3\n"
                         "punpcklwd %%mm4,%%mm2\n"
***************
*** 302,308 ****
                         /* final run */
                         /* set the current, current_pre, current_next registers */
!                        "movq -8(%1),%%mm0\n"
                         "movq (%1),%%mm7\n"
!                        "pxor %%mm1,%%mm1\n" /* use a fake black out of screen */
                         "psrlq $48,%%mm0\n"
                         "psllq $48,%%mm1\n"
--- 256,263 ----
                         /* final run */
                         /* set the current, current_pre, current_next registers */
!                        "movq (%1),%%mm1\n"
                         "movq (%1),%%mm7\n"
!                        "movq -8(%1),%%mm0\n"
!                        "psrlq $48,%%mm1\n"
                         "psrlq $48,%%mm0\n"
                         "psllq $48,%%mm1\n"
***************
*** 317,322 ****
                         "movq (%0),%%mm6\n"
  
!                        /* compute the upper-left pixel for dst0 on %%mm2 */
!                        /* compute the upper-right pixel for dst0 on %%mm4 */
                         "movq %%mm0,%%mm2\n"
                         "movq %%mm1,%%mm4\n"
--- 272,277 ----
                         "movq (%0),%%mm6\n"
  
!                        /* compute the upper-left pixel for dst on %%mm2 */
!                        /* compute the upper-right pixel for dst on %%mm4 */
                         "movq %%mm0,%%mm2\n"
                         "movq %%mm1,%%mm4\n"
***************
*** 344,348 ****
                         "por %%mm5,%%mm4\n"
  
!                        /* set *dst0 */
                         "movq %%mm2,%%mm3\n"
                         "punpcklwd %%mm4,%%mm2\n"
--- 299,303 ----
                         "por %%mm5,%%mm4\n"
  
!                        /* set *dst */
                         "movq %%mm2,%%mm3\n"
                         "punpcklwd %%mm4,%%mm2\n"
***************
*** 366,374 ****
      /* first run */
      /* set the current, current_pre, current_next registers */
!     pxor mm0,mm0; /* use a fake black out of screen */
      movq mm7, qword ptr [ebx];
      movq mm1, qword ptr [ebx + 8];
!     psrlq mm0, 48;
      psllq mm1, 48;
      movq mm2, mm7;
      movq mm3, mm7;
--- 321,330 ----
      /* first run */
      /* set the current, current_pre, current_next registers */
!     movq mm0, qword ptr [ebx];
      movq mm7, qword ptr [ebx];
      movq mm1, qword ptr [ebx + 8];
!     psllq mm0, 48;
      psllq mm1, 48;
+     psrlq mm0, 48;
      movq mm2, mm7;
      movq mm3, mm7;
***************
*** 381,386 ****
      movq mm6, qword ptr [eax];
  
!     /* compute the upper-left pixel for dst0 on %%mm2 */
!     /* compute the upper-right pixel for dst0 on %%mm4 */
      movq mm2, mm0;
      movq mm4, mm1;
--- 337,342 ----
      movq mm6, qword ptr [eax];
  
!     /* compute the upper-left pixel for dst on %%mm2 */
!     /* compute the upper-right pixel for dst on %%mm4 */
      movq mm2, mm0;
      movq mm4, mm1;
***************
*** 443,448 ****
      movq mm6, qword ptr [eax];
  
!     /* compute the upper-left pixel for dst0 on %%mm2 */
!     /* compute the upper-right pixel for dst0 on %%mm4 */
      movq mm2,mm0;
      movq mm4,mm1;
--- 399,404 ----
      movq mm6, qword ptr [eax];
  
!     /* compute the upper-left pixel for dst on %%mm2 */
!     /* compute the upper-right pixel for dst on %%mm4 */
      movq mm2,mm0;
      movq mm4,mm1;
***************
*** 470,474 ****
      por mm4,mm5;
  
!     /* set *dst0 */
      movq mm3,mm2;
      punpcklwd mm2,mm4;
--- 426,430 ----
      por mm4,mm5;
  
!     /* set *dst */
      movq mm3,mm2;
      punpcklwd mm2,mm4;
***************
*** 489,495 ****
      /* final run */
      /* set the current, current_pre, current_next registers */
!     movq mm0, qword ptr [ebx-8];
      movq mm7, qword ptr [ebx];
!     pxor mm1,mm1; /* use a fake black out of screen */
      psrlq mm0,48;
      psllq mm1,48;
--- 445,452 ----
      /* final run */
      /* set the current, current_pre, current_next registers */
!     movq mm1, qword ptr [ebx];
      movq mm7, qword ptr [ebx];
!     movq mm0, qword ptr [ebx-8];
!     psrlq mm1,48;
      psrlq mm0,48;
      psllq mm1,48;
***************
*** 504,509 ****
      movq mm6, qword ptr [eax];
  
!     /* compute the upper-left pixel for dst0 on %%mm2 */
!     /* compute the upper-right pixel for dst0 on %%mm4 */
      movq mm2,mm0;
      movq mm4,mm1;
--- 461,466 ----
      movq mm6, qword ptr [eax];
  
!     /* compute the upper-left pixel for dst on %%mm2 */
!     /* compute the upper-right pixel for dst on %%mm4 */
      movq mm2,mm0;
      movq mm4,mm1;
***************
*** 531,535 ****
      por mm4,mm5;
  
!     /* set *dst0 */
      movq mm3,mm2;
      punpcklwd mm2,mm4;
--- 488,492 ----
      por mm4,mm5;
  
!     /* set *dst */
      movq mm3,mm2;
      punpcklwd mm2,mm4;
***************
*** 557,565 ****
                         /* first run */
                         /* set the current, current_pre, current_next registers */
!                        "pxor %%mm0,%%mm0\n" /* use a fake black out of screen */
                         "movq 0(%1),%%mm7\n"
                         "movq 8(%1),%%mm1\n"
!                        "psrlq $32,%%mm0\n"
                         "psllq $32,%%mm1\n"
                         "movq %%mm7,%%mm2\n"
                         "movq %%mm7,%%mm3\n"
--- 514,523 ----
                         /* first run */
                         /* set the current, current_pre, current_next registers */
!                        "movq 0(%1),%%mm0\n"
                         "movq 0(%1),%%mm7\n"
                         "movq 8(%1),%%mm1\n"
!                        "psllq $32,%%mm0\n"
                         "psllq $32,%%mm1\n"
+                        "psrlq $32,%%mm0\n"
                         "movq %%mm7,%%mm2\n"
                         "movq %%mm7,%%mm3\n"
***************
*** 572,577 ****
                         "movq (%0),%%mm6\n"
  
!                        /* compute the upper-left pixel for dst0 on %%mm2 */
!                        /* compute the upper-right pixel for dst0 on %%mm4 */
                         "movq %%mm0,%%mm2\n"
                         "movq %%mm1,%%mm4\n"
--- 530,535 ----
                         "movq (%0),%%mm6\n"
  
!                        /* compute the upper-left pixel for dst on %%mm2 */
!                        /* compute the upper-right pixel for dst on %%mm4 */
                         "movq %%mm0,%%mm2\n"
                         "movq %%mm1,%%mm4\n"
***************
*** 599,608 ****
                         "por %%mm5,%%mm4\n"
  
!                        /* set *dst0 */
                         "movq %%mm2,%%mm3\n"
                         "punpckldq %%mm4,%%mm2\n"
                         "punpckhdq %%mm4,%%mm3\n"
                         "movq %%mm2,(%3)\n"
!                        "movq %%mm3,8(%3)\n"
  
                         /* next */
--- 557,566 ----
                         "por %%mm5,%%mm4\n"
  
!                        /* set *dst */
                         "movq %%mm2,%%mm3\n"
                         "punpckldq %%mm4,%%mm2\n"
                         "punpckhdq %%mm4,%%mm3\n"
                         "movq %%mm2,(%3)\n"
!                        "movq %%mm3, 8(%3)\n"
  
                         /* next */
***************
*** 615,619 ****
                         "shrl $1,%4\n"
                         "jz 1f\n"
!                        ASM_JUMP_ALIGN
                         "0:\n"
  
--- 573,577 ----
                         "shrl $1,%4\n"
                         "jz 1f\n"
! 
                         "0:\n"
  
***************
*** 634,639 ****
                         "movq (%0),%%mm6\n"
  
!                        /* compute the upper-left pixel for dst0 on %%mm2 */
!                        /* compute the upper-right pixel for dst0 on %%mm4 */
                         "movq %%mm0,%%mm2\n"
                         "movq %%mm1,%%mm4\n"
--- 592,597 ----
                         "movq (%0),%%mm6\n"
  
!                        /* compute the upper-left pixel for dst on %%mm2 */
!                        /* compute the upper-right pixel for dst on %%mm4 */
                         "movq %%mm0,%%mm2\n"
                         "movq %%mm1,%%mm4\n"
***************
*** 661,665 ****
                         "por %%mm5,%%mm4\n"
  
!                        /* set *dst0 */
                         "movq %%mm2,%%mm3\n"
                         "punpckldq %%mm4,%%mm2\n"
--- 619,623 ----
                         "por %%mm5,%%mm4\n"
  
!                        /* set *dst */
                         "movq %%mm2,%%mm3\n"
                         "punpckldq %%mm4,%%mm2\n"
***************
*** 680,686 ****
                         /* final run */
                         /* set the current, current_pre, current_next registers */
!                        "movq -8(%1),%%mm0\n"
                         "movq (%1),%%mm7\n"
!                        "pxor %%mm1,%%mm1\n" /* use a fake black out of screen */
                         "psrlq $32,%%mm0\n"
                         "psllq $32,%%mm1\n"
--- 638,645 ----
                         /* final run */
                         /* set the current, current_pre, current_next registers */
!                        "movq (%1),%%mm1\n"
                         "movq (%1),%%mm7\n"
!                        "movq -8(%1), %%mm0\n"
!                        "psrlq $32,%%mm1\n"
                         "psrlq $32,%%mm0\n"
                         "psllq $32,%%mm1\n"
***************
*** 695,700 ****
                         "movq (%0),%%mm6\n"
  
!                        /* compute the upper-left pixel for dst0 on %%mm2 */
!                        /* compute the upper-right pixel for dst0 on %%mm4 */
                         "movq %%mm0,%%mm2\n"
                         "movq %%mm1,%%mm4\n"
--- 654,659 ----
                         "movq (%0),%%mm6\n"
  
!                        /* compute the upper-left pixel for dst on %%mm2 */
!                        /* compute the upper-right pixel for dst on %%mm4 */
                         "movq %%mm0,%%mm2\n"
                         "movq %%mm1,%%mm4\n"
***************
*** 722,726 ****
                         "por %%mm5,%%mm4\n"
  
!                        /* set *dst0 */
                         "movq %%mm2,%%mm3\n"
                         "punpckldq %%mm4,%%mm2\n"
--- 681,685 ----
                         "por %%mm5,%%mm4\n"
  
!                        /* set *dst */
                         "movq %%mm2,%%mm3\n"
                         "punpckldq %%mm4,%%mm2\n"
***************
*** 744,752 ****
      /* first run */
      /* set the current, current_pre, current_next registers */
!     pxor mm0,mm0;
      movq mm7,qword ptr [ebx];
      movq mm1,qword ptr [ebx + 8];
!     psrlq mm0,32;
      psllq mm1,32;
      movq mm2,mm7;
      movq mm3,mm7;
--- 703,712 ----
      /* first run */
      /* set the current, current_pre, current_next registers */
!     movq mm0,qword ptr [ebx];
      movq mm7,qword ptr [ebx];
      movq mm1,qword ptr [ebx + 8];
!     psllq mm0,32;
      psllq mm1,32;
+     psrlq mm0,32;
      movq mm2,mm7;
      movq mm3,mm7;
***************
*** 759,764 ****
      movq mm6,qword ptr [eax];
  
!     /* compute the upper-left pixel for dst0 on %%mm2 */
!     /* compute the upper-right pixel for dst0 on %%mm4 */
      movq mm2,mm0;
      movq mm4,mm1;
--- 719,724 ----
      movq mm6,qword ptr [eax];
  
!     /* compute the upper-left pixel for dst on %%mm2 */
!     /* compute the upper-right pixel for dst on %%mm4 */
      movq mm2,mm0;
      movq mm4,mm1;
***************
*** 786,790 ****
      por mm4,mm5;
  
!     /* set *dst0 */
      movq mm3,mm2;
      punpckldq mm2,mm4;
--- 746,750 ----
      por mm4,mm5;
  
!     /* set *dst */
      movq mm3,mm2;
      punpckldq mm2,mm4;
***************
*** 820,825 ****
      movq mm6,qword ptr[eax];
  
!     /* compute the upper-left pixel for dst0 on %%mm2 */
!     /* compute the upper-right pixel for dst0 on %%mm4 */
      movq mm2,mm0;
      movq mm4,mm1;
--- 780,785 ----
      movq mm6,qword ptr[eax];
  
!     /* compute the upper-left pixel for dst on %%mm2 */
!     /* compute the upper-right pixel for dst on %%mm4 */
      movq mm2,mm0;
      movq mm4,mm1;
***************
*** 847,851 ****
      por mm4,mm5;
  
!     /* set *dst0 */
      movq mm3,mm2;
      punpckldq mm2,mm4;
--- 807,811 ----
      por mm4,mm5;
  
!     /* set *dst */
      movq mm3,mm2;
      punpckldq mm2,mm4;
***************
*** 866,872 ****
      /* final run */
      /* set the current, current_pre, current_next registers */
!     movq mm0,qword ptr [ebx-8];
      movq mm7,qword ptr [ebx];
!     pxor mm1,mm1;
      psrlq mm0,32;
      psllq mm1,32;
--- 826,833 ----
      /* final run */
      /* set the current, current_pre, current_next registers */
!     movq mm1,qword ptr [ebx];
      movq mm7,qword ptr [ebx];
!     movq mm0,qword ptr [ebx-8];
!     psrlq mm1,32;
      psrlq mm0,32;
      psllq mm1,32;
***************
*** 881,886 ****
      movq mm6,qword ptr [eax];
  
!     /* compute the upper-left pixel for dst0 on %%mm2 */
!     /* compute the upper-right pixel for dst0 on %%mm4 */
      movq mm2,mm0;
      movq mm4,mm1;
--- 842,847 ----
      movq mm6,qword ptr [eax];
  
!     /* compute the upper-left pixel for dst on %%mm2 */
!     /* compute the upper-right pixel for dst on %%mm4 */
      movq mm2,mm0;
      movq mm4,mm1;
***************
*** 908,912 ****
      por mm4,mm5;
  
!     /* set *dst0 */
      movq mm3,mm2;
      punpckldq mm2,mm4;
--- 869,873 ----
      por mm4,mm5;
  
!     /* set *dst */
      movq mm3,mm2;
      punpckldq mm2,mm4;
***************
*** 969,973 ****
    } else {
  #endif
!     internal_scale2x_16_def(dst0, dst1, src0, src0, src1, width);
    
      int count = height;
--- 930,935 ----
    } else {
  #endif
!     internal_scale2x_16_def(dst0, src0, src0, src1, width);
!     internal_scale2x_16_def(dst1, src1, src0, src0, width);
    
      int count = height;
***************
*** 977,981 ****
        dst0 += dstPitch;
        dst1 += dstPitch;
!       internal_scale2x_16_def(dst0, dst1, src0, src1, src2, width);
        src0 = src1;
        src1 = src2;
--- 939,944 ----
        dst0 += dstPitch;
        dst1 += dstPitch;
!       internal_scale2x_16_def(dst0, src0, src1, src2, width);
!       internal_scale2x_16_def(dst1, src2, src1, src0, width);
        src0 = src1;
        src1 = src2;
***************
*** 985,989 ****
      dst0 += dstPitch;
      dst1 += dstPitch;
!     internal_scale2x_16_def(dst0, dst1, src0, src1, src1, width);
  #ifdef MMX
    }
--- 948,953 ----
      dst0 += dstPitch;
      dst1 += dstPitch;
!     internal_scale2x_16_def(dst0, src0, src1, src1, width);
!     internal_scale2x_16_def(dst1, src1, src1, src0, width);
  #ifdef MMX
    }
***************
*** 1021,1025 ****
    } else {
  #endif
!     internal_scale2x_32_def(dst0, dst1, src0, src0, src1, width);
    
      int count = height;
--- 985,990 ----
    } else {
  #endif
!     internal_scale2x_32_def(dst0, src0, src0, src1, width);
!     internal_scale2x_32_def(dst1, src1, src0, src0, width);
    
      int count = height;
***************
*** 1029,1033 ****
        dst0 += dstPitch >> 1;
        dst1 += dstPitch >> 1;
!       internal_scale2x_32_def(dst0, dst1, src0, src1, src2, width);
        src0 = src1;
        src1 = src2;
--- 994,999 ----
        dst0 += dstPitch >> 1;
        dst1 += dstPitch >> 1;
!       internal_scale2x_32_def(dst0, src0, src1, src2, width);
!       internal_scale2x_32_def(dst1, src2, src1, src0, width);
        src0 = src1;
        src1 = src2;
***************
*** 1037,1041 ****
      dst0 += dstPitch >> 1;
      dst1 += dstPitch >> 1;
!     internal_scale2x_32_def(dst0, dst1, src0, src1, src1, width);
  #ifdef MMX
    }
--- 1003,1008 ----
      dst0 += dstPitch >> 1;
      dst1 += dstPitch >> 1;
!     internal_scale2x_32_def(dst0, src0, src1, src1, width);
!     internal_scale2x_32_def(dst1, src1, src1, src0, width);
  #ifdef MMX
    }