diff --git a/doc/ReadMe.txt b/doc/ReadMe.txt
index 23fe9347..e8c60512 100644
--- a/doc/ReadMe.txt
+++ b/doc/ReadMe.txt
@@ -63,9 +63,6 @@ Important:
 
 - Improve automatic 64k/128k flash save detection
 
-- HQ3x/4x ASM implementation produces wrong interpolation on the image's border
- - This has already been fixed in the C version; look at hq_base.h / line 343 - 372. The ASM version most likely only has something like skipLine instead of skipLinePlus and skipLineMinus, which is however necessary in order to work correctly.
-
 - Fix OpenGL issues
 
 - Remove 16 bit hack for filters
@@ -107,7 +104,7 @@ Performance:
 
 - Apply pixel filter only to changed parts of the image
 
-- Make use of multi-core CPUs
+- Make even more use of multi-core CPUs
 
 - Make use of 64 bit CPUs
 
@@ -150,4 +147,6 @@ Jonas Quinn
 Spacy
 
 Special Thanks go to:
-chrono - ASM genius, fixed a bug in the HQ3x/4x filters and made them thread-safe
+chrono:
+ASM genius, fixed a bug in the HQ3x/4x filters.
+Made HQ3x/4x and Bilinear filters (ASM versions) thread-safe
diff --git a/src/bilinear.cpp b/src/bilinear.cpp
index 00e7a1c8..3bd72d0a 100644
--- a/src/bilinear.cpp
+++ b/src/bilinear.cpp
@@ -24,12 +24,6 @@
 
 #include "System.h"
 
-static u8 row_cur[3*322];
-static u8 row_next[3*322];
-
-static u8 *rgb_row_cur = row_cur;
-static u8 *rgb_row_next = row_next;
-
 #define RGB(r,g,b) ((r)>>3) << systemRedShift |\
   ((g) >> 3) << systemGreenShift |\
   ((b) >> 3) << systemBlueShift\
@@ -83,6 +77,11 @@ static void fill_rgb_row_32(u32 *from, int src_width, u8 *row, int width)
 void Bilinear(u8 *srcPtr, u32 srcPitch, u8 * /* deltaPtr */,
               u8 *dstPtr, u32 dstPitch, int width, int height)
 {
+  u8 row_cur[3*322];
+  u8 row_next[3*322];
+  u8 *rgb_row_cur = row_cur;
+  u8 *rgb_row_next = row_next;
+
   u16 *to = (u16 *)dstPtr;
   u16 *to_odd = (u16 *)(dstPtr + dstPitch);
 
@@ -161,6 +160,11 @@ void Bilinear(u8 *srcPtr, u32 srcPitch, u8 * /* deltaPtr */,
 void BilinearPlus(u8 *srcPtr, u32 srcPitch, u8 * /* deltaPtr */,
                   u8 *dstPtr, u32 dstPitch, int width, int height)
 {
+  u8 row_cur[3*322];
+  u8 row_next[3*322];
+  u8 *rgb_row_cur = row_cur;
+  u8 *rgb_row_next = row_next;
+
   u16 *to = (u16 *)dstPtr;
   u16 *to_odd = (u16 *)(dstPtr + dstPitch);
 
@@ -250,6 +254,11 @@ void BilinearPlus(u8 *srcPtr, u32 srcPitch, u8 * /* deltaPtr */,
 void Bilinear32(u8 *srcPtr, u32 srcPitch, u8 * /* deltaPtr */,
                 u8 *dstPtr, u32 dstPitch, int width, int height)
 {
+  u8 row_cur[3*322];
+  u8 row_next[3*322];
+  u8 *rgb_row_cur = row_cur;
+  u8 *rgb_row_next = row_next;
+
   u32 *to = (u32 *)dstPtr;
   u32 *to_odd = (u32 *)(dstPtr + dstPitch);
 
@@ -330,6 +339,11 @@ void Bilinear32(u8 *srcPtr, u32 srcPitch, u8 * /* deltaPtr */,
 void BilinearPlus32(u8 *srcPtr, u32 srcPitch, u8 * /* deltaPtr */,
                     u8 *dstPtr, u32 dstPitch, int width, int height)
 {
+  u8 row_cur[3*322];
+  u8 row_next[3*322];
+  u8 *rgb_row_cur = row_cur;
+  u8 *rgb_row_next = row_next;
+
   u32 *to = (u32 *)dstPtr;
   u32 *to_odd = (u32 *)(dstPtr + dstPitch);
 
diff --git a/src/hq/asm/hq3x_16.asm b/src/hq/asm/hq3x_16.asm
index f32c5272..85f0160b 100644
--- a/src/hq/asm/hq3x_16.asm
+++ b/src/hq/asm/hq3x_16.asm
@@ -1,4 +1,4 @@
-;hq3x filter
+;hq3x filter (thread-safe version)
 ;16bpp output
 ;----------------------------------------------------------
 ;Copyright (C) 2003 MaxSt ( maxst@hiend3d.com )
@@ -23,20 +23,6 @@
 EXTSYM LUT16to32,RGBtoYUV
 
 SECTION .bss
-linesleft resd 1
-xcounter  resd 1
-cross     resd 1
-nextline  resd 1
-prevline  resd 1
-w1        resd 1
-w2        resd 1
-w3        resd 1
-w4        resd 1
-w5        resd 1
-w6        resd 1
-w7        resd 1
-w8        resd 1
-w9        resd 1
 
 SECTION .data
 
@@ -217,27 +203,27 @@ SECTION .text
 %endmacro
 
 %macro PIXEL00_1M 0
-    Interp1 [edi],eax,[w1]
+    Interp1 [edi],eax,dword[ebp-w1]
 %endmacro
 
 %macro PIXEL00_1U 0
-    Interp1 [edi],eax,[w2]
+    Interp1 [edi],eax,dword[ebp-w2]
 %endmacro
 
 %macro PIXEL00_1L 0
-    Interp1 [edi],eax,[w4]
+    Interp1 [edi],eax,dword[ebp-w4]
 %endmacro
 
 %macro PIXEL00_2 0
-    Interp2 [edi],eax,[w4],[w2]
+    Interp2 [edi],eax,dword[ebp-w4],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL00_4 0
-    Interp4 [edi],[w4],[w2]
+    Interp4 [edi],dword[ebp-w4],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL00_5 0
-    Interp5 [edi],[w4],[w2]
+    Interp5 [edi],dword[ebp-w4],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL00_C 0
@@ -245,15 +231,15 @@ SECTION .text
 %endmacro
 
 %macro PIXEL01_1 0
-    Interp1 [edi+2],eax,[w2]
+    Interp1 [edi+2],eax,dword[ebp-w2]
 %endmacro
 
 %macro PIXEL01_3 0
-    Interp3 [edi+2],[w2]
+    Interp3 [edi+2],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL01_6 0
-    Interp1 [edi+2],[w2],eax
+    Interp1 [edi+2],dword[ebp-w2],eax
 %endmacro
 
 %macro PIXEL01_C 0
@@ -261,27 +247,27 @@ SECTION .text
 %endmacro
 
 %macro PIXEL02_1M 0
-    Interp1 [edi+4],eax,[w3]
+    Interp1 [edi+4],eax,dword[ebp-w3]
 %endmacro
 
 %macro PIXEL02_1U 0
-    Interp1 [edi+4],eax,[w2]
+    Interp1 [edi+4],eax,dword[ebp-w2]
 %endmacro
 
 %macro PIXEL02_1R 0
-    Interp1 [edi+4],eax,[w6]
+    Interp1 [edi+4],eax,dword[ebp-w6]
 %endmacro
 
 %macro PIXEL02_2 0
-    Interp2 [edi+4],eax,[w2],[w6]
+    Interp2 [edi+4],eax,dword[ebp-w2],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL02_4 0
-    Interp4 [edi+4],[w2],[w6]
+    Interp4 [edi+4],dword[ebp-w2],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL02_5 0
-    Interp5 [edi+4],[w2],[w6]
+    Interp5 [edi+4],dword[ebp-w2],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL02_C 0
@@ -289,15 +275,15 @@ SECTION .text
 %endmacro
 
 %macro PIXEL10_1 0
-    Interp1 [edi+ebx],eax,[w4]
+    Interp1 [edi+ebx],eax,dword[ebp-w4]
 %endmacro
 
 %macro PIXEL10_3 0
-    Interp3 [edi+ebx],[w4]
+    Interp3 [edi+ebx],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL10_6 0
-    Interp1 [edi+ebx],[w4],eax
+    Interp1 [edi+ebx],dword[ebp-w4],eax
 %endmacro
 
 %macro PIXEL10_C 0
@@ -309,15 +295,15 @@ SECTION .text
 %endmacro
 
 %macro PIXEL12_1 0
-    Interp1 [edi+ebx+4],eax,[w6]
+    Interp1 [edi+ebx+4],eax,dword[ebp-w6]
 %endmacro
 
 %macro PIXEL12_3 0
-    Interp3 [edi+ebx+4],[w6]
+    Interp3 [edi+ebx+4],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL12_6 0
-    Interp1 [edi+ebx+4],[w6],eax
+    Interp1 [edi+ebx+4],dword[ebp-w6],eax
 %endmacro
 
 %macro PIXEL12_C 0
@@ -325,27 +311,27 @@ SECTION .text
 %endmacro
 
 %macro PIXEL20_1M 0
-    Interp1 [edi+ebx*2],eax,[w7]
+    Interp1 [edi+ebx*2],eax,dword[ebp-w7]
 %endmacro
 
 %macro PIXEL20_1D 0
-    Interp1 [edi+ebx*2],eax,[w8]
+    Interp1 [edi+ebx*2],eax,dword[ebp-w8]
 %endmacro
 
 %macro PIXEL20_1L 0
-    Interp1 [edi+ebx*2],eax,[w4]
+    Interp1 [edi+ebx*2],eax,dword[ebp-w4]
 %endmacro
 
 %macro PIXEL20_2 0
-    Interp2 [edi+ebx*2],eax,[w8],[w4]
+    Interp2 [edi+ebx*2],eax,dword[ebp-w8],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL20_4 0
-    Interp4 [edi+ebx*2],[w8],[w4]
+    Interp4 [edi+ebx*2],dword[ebp-w8],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL20_5 0
-    Interp5 [edi+ebx*2],[w8],[w4]
+    Interp5 [edi+ebx*2],dword[ebp-w8],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL20_C 0
@@ -353,15 +339,15 @@ SECTION .text
 %endmacro
 
 %macro PIXEL21_1 0
-    Interp1 [edi+ebx*2+2],eax,[w8]
+    Interp1 [edi+ebx*2+2],eax,dword[ebp-w8]
 %endmacro
 
 %macro PIXEL21_3 0
-    Interp3 [edi+ebx*2+2],[w8]
+    Interp3 [edi+ebx*2+2],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL21_6 0
-    Interp1 [edi+ebx*2+2],[w8],eax
+    Interp1 [edi+ebx*2+2],dword[ebp-w8],eax
 %endmacro
 
 %macro PIXEL21_C 0
@@ -369,27 +355,27 @@ SECTION .text
 %endmacro
 
 %macro PIXEL22_1M 0
-    Interp1 [edi+ebx*2+4],eax,[w9]
+    Interp1 [edi+ebx*2+4],eax,dword[ebp-w9]
 %endmacro
 
 %macro PIXEL22_1D 0
-    Interp1 [edi+ebx*2+4],eax,[w8]
+    Interp1 [edi+ebx*2+4],eax,dword[ebp-w8]
 %endmacro
 
 %macro PIXEL22_1R 0
-    Interp1 [edi+ebx*2+4],eax,[w6]
+    Interp1 [edi+ebx*2+4],eax,dword[ebp-w6]
 %endmacro
 
 %macro PIXEL22_2 0
-    Interp2 [edi+ebx*2+4],eax,[w6],[w8]
+    Interp2 [edi+ebx*2+4],eax,dword[ebp-w6],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL22_4 0
-    Interp4 [edi+ebx*2+4],[w6],[w8]
+    Interp4 [edi+ebx*2+4],dword[ebp-w6],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL22_5 0
-    Interp5 [edi+ebx*2+4],[w6],[w8]
+    Interp5 [edi+ebx*2+4],dword[ebp-w6],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL22_C 0
@@ -403,95 +389,111 @@ Yres         equ 20
 pitch        equ 24
 offset        equ 28
 
+linesleft equ 4
+xcounter  equ 8
+cross     equ 12
+nextline  equ 16
+prevline  equ 20
+w1        equ 24
+w2        equ 28
+w3        equ 32
+w4        equ 36
+w5        equ 40
+w6        equ 44
+w7        equ 48
+w8        equ 52
+w9        equ 56
+localsize equ 56
+
 NEWSYM hq3x_16
     push ebp
     mov ebp,esp
+    sub esp, localsize
     pushad
 
     mov     esi,[ebp+inbuffer]
     mov     edi,[ebp+outbuffer]
     mov     edx,[ebp+Yres]
-    mov     [linesleft],edx
+    mov     [ebp-linesleft],edx
     mov     ebx,[ebp+Xres]
     shl     ebx,1
-    mov     dword[prevline],0
-    mov     eax, ebx
-    add     eax, [ebp+offset]
-    mov     dword[nextline],eax
-    
+    mov     dword[ebp-prevline],0
+    mov     eax, [ebp+offset]
+    add     eax, ebx
+    mov     dword[ebp-nextline],eax
 .loopy
     mov     ecx,[ebp+Xres]
     sub     ecx,2                 ; x={Xres-2, Xres-1} are special cases.
-    mov     dword[xcounter],ecx
+    mov     dword[ebp-xcounter],ecx
     ; x=0 - special case
-    mov     ebx,[prevline]
+    mov     ebx,[ebp-prevline]
     movq    mm5,[esi+ebx]
     movq    mm6,[esi]
-    mov     ebx,[nextline]
+    mov     ebx,[ebp-nextline]
     movq    mm7,[esi+ebx]
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w1],edx
-    mov     [w2],edx
+    mov     [ebp-w1],edx
+    mov     [ebp-w2],edx
     shr     eax,16
-    mov     [w3],eax
+    mov     [ebp-w3],eax
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w4],edx
-    mov     [w5],edx
+    mov     [ebp-w4],edx
+    mov     [ebp-w5],edx
     shr     eax,16
-    mov     [w6],eax
+    mov     [ebp-w6],eax
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w7],edx
-    mov     [w8],edx
+    mov     [ebp-w7],edx
+    mov     [ebp-w8],edx
     shr     eax,16
-    mov     [w9],eax
+    mov     [ebp-w9],eax
     jmp     .flags
 .loopx
-    mov     ebx,[prevline]
+    mov     ebx,[ebp-prevline]
     movq    mm5,[esi+ebx-2]
     movq    mm6,[esi-2]
-    mov     ebx,[nextline]
+    mov     ebx,[ebp-nextline]
     movq    mm7,[esi+ebx-2]
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w1],edx
+    mov     [ebp-w1],edx
     shr     eax,16
-    mov     [w2],eax
+    mov     [ebp-w2],eax
     psrlq   mm5,32
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w3],edx
+    mov     [ebp-w3],edx
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w4],edx
+    mov     [ebp-w4],edx
     shr     eax,16
-    mov     [w5],eax
+    mov     [ebp-w5],eax
     psrlq   mm6,32
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w6],edx
+    mov     [ebp-w6],edx
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w7],edx
+    mov     [ebp-w7],edx
     shr     eax,16
-    mov     [w8],eax
+    mov     [ebp-w8],eax
     psrlq   mm7,32
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w9],edx
+    mov     [ebp-w9],edx
 .flags
     mov     ebx,RGBtoYUV
-    mov     eax,[w5]
+    mov     eax,[ebp-w5]
     xor     ecx,ecx
     movd    mm5,[ebx+eax*4]
-    mov     dword[cross],0
+    mov     dword[ebp-cross],0
 
-    mov     edx,[w2]
+    mov     edx,[ebp-w2]
     cmp     eax,edx
     je      .noflag2
-    or      dword[cross],1
+    or      dword[ebp-cross],1
     movq    mm1,mm5
     movd    mm2,[ebx+edx*4]
     psubusb mm1,mm2
@@ -503,10 +505,10 @@ NEWSYM hq3x_16
     jz      .noflag2
     or      ecx,2
 .noflag2
-    mov     edx,[w4]
+    mov     edx,[ebp-w4]
     cmp     eax,edx
     je      .noflag4
-    or      dword[cross],2
+    or      dword[ebp-cross],2
     movq    mm1,mm5
     movd    mm2,[ebx+edx*4]
     psubusb mm1,mm2
@@ -518,10 +520,10 @@ NEWSYM hq3x_16
     jz      .noflag4
     or      ecx,8
 .noflag4
-    mov     edx,[w6]
+    mov     edx,[ebp-w6]
     cmp     eax,edx
     je      .noflag6
-    or      dword[cross],4
+    or      dword[ebp-cross],4
     movq    mm1,mm5
     movd    mm2,[ebx+edx*4]
     psubusb mm1,mm2
@@ -533,10 +535,10 @@ NEWSYM hq3x_16
     jz      .noflag6
     or      ecx,16
 .noflag6
-    mov     edx,[w8]
+    mov     edx,[ebp-w8]
     cmp     eax,edx
     je      .noflag8
-    or      dword[cross],8
+    or      dword[ebp-cross],8
     movq    mm1,mm5
     movd    mm2,[ebx+edx*4]
     psubusb mm1,mm2
@@ -550,11 +552,11 @@ NEWSYM hq3x_16
 .noflag8
     test    ecx,ecx
     jnz     .testflag1
-    mov     ecx,[cross]
+    mov     ecx,[ebp-cross]
     mov     ebx,[ebp+pitch]
     jmp     [FuncTable2+ecx*4]
 .testflag1
-    mov     edx,[w1]
+    mov     edx,[ebp-w1]
     cmp     eax,edx
     je      .noflag1
     movq    mm1,mm5
@@ -568,7 +570,7 @@ NEWSYM hq3x_16
     jz      .noflag1
     or      ecx,1
 .noflag1
-    mov     edx,[w3]
+    mov     edx,[ebp-w3]
     cmp     eax,edx
     je      .noflag3
     movq    mm1,mm5
@@ -582,7 +584,7 @@ NEWSYM hq3x_16
     jz      .noflag3
     or      ecx,4
 .noflag3
-    mov     edx,[w7]
+    mov     edx,[ebp-w7]
     cmp     eax,edx
     je      .noflag7
     movq    mm1,mm5
@@ -596,7 +598,7 @@ NEWSYM hq3x_16
     jz      .noflag7
     or      ecx,32
 .noflag7
-    mov     edx,[w9]
+    mov     edx,[ebp-w9]
     cmp     eax,edx
     je      .noflag9
     movq    mm1,mm5
@@ -810,7 +812,7 @@ NEWSYM hq3x_16
 ..@flag18
 ..@flag50
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_1M,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_1M,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_1
     PIXEL11
     PIXEL20_2
@@ -825,7 +827,7 @@ NEWSYM hq3x_16
     PIXEL10_1
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_1M,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_1M,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag72
 ..@flag76
@@ -834,12 +836,12 @@ NEWSYM hq3x_16
     PIXEL02_2
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_1M,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_1M,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag10
 ..@flag138
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
@@ -924,7 +926,7 @@ NEWSYM hq3x_16
 ..@flag22
 ..@flag54
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_1
     PIXEL11
     PIXEL20_2
@@ -939,7 +941,7 @@ NEWSYM hq3x_16
     PIXEL10_1
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag104
 ..@flag108
@@ -948,12 +950,12 @@ NEWSYM hq3x_16
     PIXEL02_2
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag11
 ..@flag139
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
@@ -963,7 +965,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag19
 ..@flag51
-    DiffOrNot w2,w6,PIXEL00_1L,PIXEL01_C,PIXEL02_1M,PIXEL12_C,PIXEL00_2,PIXEL01_6,PIXEL02_5,PIXEL12_1
+    DiffOrNot ebp-w2,ebp-w6,PIXEL00_1L,PIXEL01_C,PIXEL02_1M,PIXEL12_C,PIXEL00_2,PIXEL01_6,PIXEL02_5,PIXEL12_1
     PIXEL10_1
     PIXEL11
     PIXEL20_2
@@ -972,7 +974,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag146
 ..@flag178
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_1M,PIXEL12_C,PIXEL22_1D,PIXEL01_1,PIXEL02_5,PIXEL12_6,PIXEL22_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_1M,PIXEL12_C,PIXEL22_1D,PIXEL01_1,PIXEL02_5,PIXEL12_6,PIXEL22_2
     PIXEL00_1M
     PIXEL10_1
     PIXEL11
@@ -981,7 +983,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag84
 ..@flag85
-    DiffOrNot w6,w8,PIXEL02_1U,PIXEL12_C,PIXEL21_C,PIXEL22_1M,PIXEL02_2,PIXEL12_6,PIXEL21_1,PIXEL22_5
+    DiffOrNot ebp-w6,ebp-w8,PIXEL02_1U,PIXEL12_C,PIXEL21_C,PIXEL22_1M,PIXEL02_2,PIXEL12_6,PIXEL21_1,PIXEL22_5
     PIXEL00_2
     PIXEL01_1
     PIXEL10_1
@@ -990,7 +992,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag112
 ..@flag113
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL20_1L,PIXEL21_C,PIXEL22_1M,PIXEL12_1,PIXEL20_2,PIXEL21_6,PIXEL22_5
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL20_1L,PIXEL21_C,PIXEL22_1M,PIXEL12_1,PIXEL20_2,PIXEL21_6,PIXEL22_5
     PIXEL00_2
     PIXEL01_1
     PIXEL02_1M
@@ -999,7 +1001,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag200
 ..@flag204
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_1M,PIXEL21_C,PIXEL22_1R,PIXEL10_1,PIXEL20_5,PIXEL21_6,PIXEL22_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_1M,PIXEL21_C,PIXEL22_1R,PIXEL10_1,PIXEL20_5,PIXEL21_6,PIXEL22_2
     PIXEL00_1M
     PIXEL01_1
     PIXEL02_2
@@ -1008,7 +1010,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag73
 ..@flag77
-    DiffOrNot w8,w4,PIXEL00_1U,PIXEL10_C,PIXEL20_1M,PIXEL21_C,PIXEL00_2,PIXEL10_6,PIXEL20_5,PIXEL21_1
+    DiffOrNot ebp-w8,ebp-w4,PIXEL00_1U,PIXEL10_C,PIXEL20_1M,PIXEL21_C,PIXEL00_2,PIXEL10_6,PIXEL20_5,PIXEL21_1
     PIXEL01_1
     PIXEL02_2
     PIXEL11
@@ -1017,7 +1019,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag42
 ..@flag170
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL01_C,PIXEL10_C,PIXEL20_1D,PIXEL00_5,PIXEL01_1,PIXEL10_6,PIXEL20_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL01_C,PIXEL10_C,PIXEL20_1D,PIXEL00_5,PIXEL01_1,PIXEL10_6,PIXEL20_2
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
@@ -1026,7 +1028,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag14
 ..@flag142
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL01_C,PIXEL02_1R,PIXEL10_C,PIXEL00_5,PIXEL01_6,PIXEL02_2,PIXEL10_1
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL01_C,PIXEL02_1R,PIXEL10_C,PIXEL00_5,PIXEL01_6,PIXEL02_2,PIXEL10_1
     PIXEL11
     PIXEL12_1
     PIXEL20_1M
@@ -1123,9 +1125,9 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag26
 ..@flag31
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
     PIXEL11
     PIXEL20_1M
     PIXEL21_1
@@ -1134,12 +1136,12 @@ NEWSYM hq3x_16
 ..@flag82
 ..@flag214
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag88
 ..@flag248
@@ -1147,22 +1149,22 @@ NEWSYM hq3x_16
     PIXEL01_1
     PIXEL02_1M
     PIXEL11
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
     jmp .loopx_end
 ..@flag74
 ..@flag107
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
     PIXEL02_1M
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag27
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL12_C
@@ -1172,7 +1174,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag86
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_1
     PIXEL11
     PIXEL20_1M
@@ -1186,7 +1188,7 @@ NEWSYM hq3x_16
     PIXEL10_C
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag106
     PIXEL00_1M
@@ -1194,12 +1196,12 @@ NEWSYM hq3x_16
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag30
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_C
     PIXEL11
     PIXEL20_1M
@@ -1213,7 +1215,7 @@ NEWSYM hq3x_16
     PIXEL10_1
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag120
     PIXEL00_1M
@@ -1221,11 +1223,11 @@ NEWSYM hq3x_16
     PIXEL02_1M
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag75
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
@@ -1366,9 +1368,9 @@ NEWSYM hq3x_16
     PIXEL22_1D
     jmp .loopx_end
 ..@flag58
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_C
     PIXEL11
     PIXEL12_C
@@ -1379,13 +1381,13 @@ NEWSYM hq3x_16
 ..@flag83
     PIXEL00_1L
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1M
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag92
     PIXEL00_1M
@@ -1394,36 +1396,36 @@ NEWSYM hq3x_16
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag202
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
     PIXEL02_1M
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
     PIXEL22_1R
     jmp .loopx_end
 ..@flag78
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
     PIXEL02_1R
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
     PIXEL22_1M
     jmp .loopx_end
 ..@flag154
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_C
     PIXEL11
     PIXEL12_C
@@ -1434,13 +1436,13 @@ NEWSYM hq3x_16
 ..@flag114
     PIXEL00_1M
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1L
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag89
     PIXEL00_1U
@@ -1449,24 +1451,24 @@ NEWSYM hq3x_16
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag90
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag55
 ..@flag23
-    DiffOrNot w2,w6,PIXEL00_1L,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL00_2,PIXEL01_6,PIXEL02_5,PIXEL12_1
+    DiffOrNot ebp-w2,ebp-w6,PIXEL00_1L,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL00_2,PIXEL01_6,PIXEL02_5,PIXEL12_1
     PIXEL10_1
     PIXEL11
     PIXEL20_2
@@ -1475,7 +1477,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag182
 ..@flag150
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL22_1D,PIXEL01_1,PIXEL02_5,PIXEL12_6,PIXEL22_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL22_1D,PIXEL01_1,PIXEL02_5,PIXEL12_6,PIXEL22_2
     PIXEL00_1M
     PIXEL10_1
     PIXEL11
@@ -1484,7 +1486,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag213
 ..@flag212
-    DiffOrNot w6,w8,PIXEL02_1U,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL02_2,PIXEL12_6,PIXEL21_1,PIXEL22_5
+    DiffOrNot ebp-w6,ebp-w8,PIXEL02_1U,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL02_2,PIXEL12_6,PIXEL21_1,PIXEL22_5
     PIXEL00_2
     PIXEL01_1
     PIXEL10_1
@@ -1493,7 +1495,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag241
 ..@flag240
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL20_1L,PIXEL21_C,PIXEL22_C,PIXEL12_1,PIXEL20_2,PIXEL21_6,PIXEL22_5
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL20_1L,PIXEL21_C,PIXEL22_C,PIXEL12_1,PIXEL20_2,PIXEL21_6,PIXEL22_5
     PIXEL00_2
     PIXEL01_1
     PIXEL02_1M
@@ -1502,7 +1504,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag236
 ..@flag232
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL22_1R,PIXEL10_1,PIXEL20_5,PIXEL21_6,PIXEL22_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL22_1R,PIXEL10_1,PIXEL20_5,PIXEL21_6,PIXEL22_2
     PIXEL00_1M
     PIXEL01_1
     PIXEL02_2
@@ -1511,7 +1513,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag109
 ..@flag105
-    DiffOrNot w8,w4,PIXEL00_1U,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL00_2,PIXEL10_6,PIXEL20_5,PIXEL21_1
+    DiffOrNot ebp-w8,ebp-w4,PIXEL00_1U,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL00_2,PIXEL10_6,PIXEL20_5,PIXEL21_1
     PIXEL01_1
     PIXEL02_2
     PIXEL11
@@ -1520,7 +1522,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag171
 ..@flag43
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL20_1D,PIXEL00_5,PIXEL01_1,PIXEL10_6,PIXEL20_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL20_1D,PIXEL00_5,PIXEL01_1,PIXEL10_6,PIXEL20_2
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
@@ -1529,7 +1531,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag143
 ..@flag15
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL02_1R,PIXEL10_C,PIXEL00_5,PIXEL01_6,PIXEL02_2,PIXEL10_1
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL02_1R,PIXEL10_C,PIXEL00_5,PIXEL01_6,PIXEL02_2,PIXEL10_1
     PIXEL11
     PIXEL12_1
     PIXEL20_1M
@@ -1542,11 +1544,11 @@ NEWSYM hq3x_16
     PIXEL02_1U
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag203
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
@@ -1556,7 +1558,7 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag62
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_C
     PIXEL11
     PIXEL20_1D
@@ -1570,11 +1572,11 @@ NEWSYM hq3x_16
     PIXEL10_1
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag118
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_1
     PIXEL11
     PIXEL20_1L
@@ -1588,7 +1590,7 @@ NEWSYM hq3x_16
     PIXEL10_C
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag110
     PIXEL00_1M
@@ -1596,11 +1598,11 @@ NEWSYM hq3x_16
     PIXEL02_1R
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag155
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL12_C
@@ -1702,12 +1704,12 @@ NEWSYM hq3x_16
     PIXEL02_1U
     PIXEL10_C
     PIXEL11
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag158
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_C
     PIXEL11
     PIXEL20_1M
@@ -1715,26 +1717,26 @@ NEWSYM hq3x_16
     PIXEL22_1D
     jmp .loopx_end
 ..@flag234
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1R
     jmp .loopx_end
 ..@flag242
     PIXEL00_1M
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL20_1L
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag59
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL11
     PIXEL12_C
     PIXEL20_1D
@@ -1747,62 +1749,62 @@ NEWSYM hq3x_16
     PIXEL02_1M
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag87
     PIXEL00_1L
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_1
     PIXEL11
     PIXEL20_1M
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag79
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1R
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
     PIXEL22_1M
     jmp .loopx_end
 ..@flag122
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag94
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_C
     PIXEL11
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag218
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_C
     PIXEL11
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag91
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag229
     PIXEL00_2
@@ -1849,9 +1851,9 @@ NEWSYM hq3x_16
     PIXEL22_1D
     jmp .loopx_end
 ..@flag186
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_C
     PIXEL11
     PIXEL12_C
@@ -1862,13 +1864,13 @@ NEWSYM hq3x_16
 ..@flag115
     PIXEL00_1L
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1L
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag93
     PIXEL00_1U
@@ -1877,18 +1879,18 @@ NEWSYM hq3x_16
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag206
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
     PIXEL02_1R
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
     PIXEL22_1R
     jmp .loopx_end
@@ -1900,13 +1902,13 @@ NEWSYM hq3x_16
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
     PIXEL22_1R
     jmp .loopx_end
 ..@flag174
 ..@flag46
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
     PIXEL02_1R
     PIXEL10_C
@@ -1920,7 +1922,7 @@ NEWSYM hq3x_16
 ..@flag147
     PIXEL00_1L
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
@@ -1938,7 +1940,7 @@ NEWSYM hq3x_16
     PIXEL12_C
     PIXEL20_1L
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag189
     PIXEL00_1U
@@ -1964,20 +1966,20 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag126
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL11
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag219
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag125
-    DiffOrNot w8,w4,PIXEL00_1U,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL00_2,PIXEL10_6,PIXEL20_5,PIXEL21_1
+    DiffOrNot ebp-w8,ebp-w4,PIXEL00_1U,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL00_2,PIXEL10_6,PIXEL20_5,PIXEL21_1
     PIXEL01_1
     PIXEL02_1U
     PIXEL11
@@ -1985,7 +1987,7 @@ NEWSYM hq3x_16
     PIXEL22_1M
     jmp .loopx_end
 ..@flag221
-    DiffOrNot w6,w8,PIXEL02_1U,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL02_2,PIXEL12_6,PIXEL21_1,PIXEL22_5
+    DiffOrNot ebp-w6,ebp-w8,PIXEL02_1U,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL02_2,PIXEL12_6,PIXEL21_1,PIXEL22_5
     PIXEL00_1U
     PIXEL01_1
     PIXEL10_C
@@ -1993,7 +1995,7 @@ NEWSYM hq3x_16
     PIXEL20_1M
     jmp .loopx_end
 ..@flag207
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL02_1R,PIXEL10_C,PIXEL00_5,PIXEL01_6,PIXEL02_2,PIXEL10_1
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL02_1R,PIXEL10_C,PIXEL00_5,PIXEL01_6,PIXEL02_2,PIXEL10_1
     PIXEL11
     PIXEL12_1
     PIXEL20_1M
@@ -2001,7 +2003,7 @@ NEWSYM hq3x_16
     PIXEL22_1R
     jmp .loopx_end
 ..@flag238
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL22_1R,PIXEL10_1,PIXEL20_5,PIXEL21_6,PIXEL22_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL22_1R,PIXEL10_1,PIXEL20_5,PIXEL21_6,PIXEL22_2
     PIXEL00_1M
     PIXEL01_C
     PIXEL02_1R
@@ -2009,7 +2011,7 @@ NEWSYM hq3x_16
     PIXEL12_1
     jmp .loopx_end
 ..@flag190
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL22_1D,PIXEL01_1,PIXEL02_5,PIXEL12_6,PIXEL22_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL22_1D,PIXEL01_1,PIXEL02_5,PIXEL12_6,PIXEL22_2
     PIXEL00_1M
     PIXEL10_C
     PIXEL11
@@ -2017,7 +2019,7 @@ NEWSYM hq3x_16
     PIXEL21_1
     jmp .loopx_end
 ..@flag187
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL20_1D,PIXEL00_5,PIXEL01_1,PIXEL10_6,PIXEL20_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL20_1D,PIXEL00_5,PIXEL01_1,PIXEL10_6,PIXEL20_2
     PIXEL02_1M
     PIXEL11
     PIXEL12_C
@@ -2025,7 +2027,7 @@ NEWSYM hq3x_16
     PIXEL22_1D
     jmp .loopx_end
 ..@flag243
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL20_1L,PIXEL21_C,PIXEL22_C,PIXEL12_1,PIXEL20_2,PIXEL21_6,PIXEL22_5
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL20_1L,PIXEL21_C,PIXEL22_C,PIXEL12_1,PIXEL20_2,PIXEL21_6,PIXEL22_5
     PIXEL00_1L
     PIXEL01_C
     PIXEL02_1M
@@ -2033,7 +2035,7 @@ NEWSYM hq3x_16
     PIXEL11
     jmp .loopx_end
 ..@flag119
-    DiffOrNot w2,w6,PIXEL00_1L,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL00_2,PIXEL01_6,PIXEL02_5,PIXEL12_1
+    DiffOrNot ebp-w2,ebp-w6,PIXEL00_1L,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL00_2,PIXEL01_6,PIXEL02_5,PIXEL12_1
     PIXEL10_1
     PIXEL11
     PIXEL20_1L
@@ -2048,13 +2050,13 @@ NEWSYM hq3x_16
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL20_2
     PIXEL21_C
     PIXEL22_1R
     jmp .loopx_end
 ..@flag175
 ..@flag47
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL00_2
     PIXEL01_C
     PIXEL02_1R
     PIXEL10_C
@@ -2068,7 +2070,7 @@ NEWSYM hq3x_16
 ..@flag151
     PIXEL00_1L
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
@@ -2086,30 +2088,30 @@ NEWSYM hq3x_16
     PIXEL12_C
     PIXEL20_1L
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_C,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_C,PIXEL22_2
     jmp .loopx_end
 ..@flag250
     PIXEL00_1M
     PIXEL01_C
     PIXEL02_1M
     PIXEL11
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
     jmp .loopx_end
 ..@flag123
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
     PIXEL02_1M
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag95
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
     PIXEL11
     PIXEL20_1M
     PIXEL21_C
@@ -2117,12 +2119,12 @@ NEWSYM hq3x_16
     jmp .loopx_end
 ..@flag222
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
     PIXEL10_C
     PIXEL11
     PIXEL12_C
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag252
     PIXEL00_1M
@@ -2130,9 +2132,9 @@ NEWSYM hq3x_16
     PIXEL02_1U
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_C,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_C,PIXEL22_2
     jmp .loopx_end
 ..@flag249
     PIXEL00_1U
@@ -2140,34 +2142,34 @@ NEWSYM hq3x_16
     PIXEL02_1M
     PIXEL10_C
     PIXEL11
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
     jmp .loopx_end
 ..@flag235
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
     PIXEL02_1M
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL20_2
     PIXEL21_C
     PIXEL22_1R
     jmp .loopx_end
 ..@flag111
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL00_2
     PIXEL01_C
     PIXEL02_1R
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag63
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
     PIXEL10_C
     PIXEL11
     PIXEL20_1D
@@ -2175,9 +2177,9 @@ NEWSYM hq3x_16
     PIXEL22_1M
     jmp .loopx_end
 ..@flag159
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL02_2
     PIXEL11
     PIXEL12_C
     PIXEL20_1M
@@ -2187,29 +2189,29 @@ NEWSYM hq3x_16
 ..@flag215
     PIXEL00_1L
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag246
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1L
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_C,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_C,PIXEL22_2
     jmp .loopx_end
 ..@flag254
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
     PIXEL11
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_2
     jmp .loopx_end
 ..@flag253
     PIXEL00_1U
@@ -2218,39 +2220,39 @@ NEWSYM hq3x_16
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_C,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_C,PIXEL22_2
     jmp .loopx_end
 ..@flag251
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
     PIXEL02_1M
     PIXEL11
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_2,PIXEL21_3
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_2,PIXEL21_3
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
     jmp .loopx_end
 ..@flag239
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL00_2
     PIXEL01_C
     PIXEL02_1R
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL20_2
     PIXEL21_C
     PIXEL22_1R
     jmp .loopx_end
 ..@flag127
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_2,PIXEL01_3,PIXEL10_3
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_2,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
     PIXEL11
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag191
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL02_2
     PIXEL10_C
     PIXEL11
     PIXEL12_C
@@ -2259,33 +2261,33 @@ NEWSYM hq3x_16
     PIXEL22_1D
     jmp .loopx_end
 ..@flag223
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_2,PIXEL12_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_2,PIXEL12_3
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag247
     PIXEL00_1L
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1L
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_C,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_C,PIXEL22_2
     jmp .loopx_end
 ..@flag255
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL02_2
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_C,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_C,PIXEL22_2
     jmp .loopx_end
 
 ..@cross0
@@ -2303,7 +2305,7 @@ NEWSYM hq3x_16
     mov edx,eax
     shl eax,16
     or  eax,edx
-    mov ecx,[w2]
+    mov ecx,[ebp-w2]
     and edx,[zerolowbits]
     and ecx,[zerolowbits]
     add ecx,edx
@@ -2324,7 +2326,7 @@ NEWSYM hq3x_16
     mov edx,eax
     shl eax,16
     or  eax,edx
-    mov ecx,[w4]
+    mov ecx,[ebp-w4]
     and edx,[zerolowbits]
     and ecx,[zerolowbits]
     add ecx,edx
@@ -2344,7 +2346,7 @@ NEWSYM hq3x_16
     mov edx,eax
     shl eax,16
     or  eax,edx
-    mov ecx,[w6]
+    mov ecx,[ebp-w6]
     and edx,[zerolowbits]
     and ecx,[zerolowbits]
     add ecx,edx
@@ -2364,7 +2366,7 @@ NEWSYM hq3x_16
     mov edx,eax
     shl eax,16
     or  eax,edx
-    mov ecx,[w8]
+    mov ecx,[ebp-w8]
     and edx,[zerolowbits]
     and ecx,[zerolowbits]
     add ecx,edx
@@ -2385,75 +2387,75 @@ NEWSYM hq3x_16
 .loopx_end
     add     esi,2
     add     edi,6
-    dec     dword[xcounter]
+    dec     dword[ebp-xcounter]
     jle     .xres_2
     jmp     .loopx
 .xres_2
     ; x=Xres-2 - special case
     jl      .xres_1
-    mov     ebx,[prevline]
+    mov     ebx,[ebp-prevline]
     movq    mm5,[esi+ebx-4]
     movq    mm6,[esi-4]
-    mov     ebx,[nextline]
+    mov     ebx,[ebp-nextline]
     movq    mm7,[esi+ebx-4]
     psrlq   mm5,16
     psrlq   mm6,16
     psrlq   mm7,16
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w1],edx
+    mov     [ebp-w1],edx
     shr     eax,16
-    mov     [w2],eax
+    mov     [ebp-w2],eax
     psrlq   mm5,32
     movd    eax,mm5
-    mov     [w3],eax
+    mov     [ebp-w3],eax
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w4],edx
+    mov     [ebp-w4],edx
     shr     eax,16
-    mov     [w5],eax
+    mov     [ebp-w5],eax
     psrlq   mm6,32
     movd    eax,mm6
-    mov     [w6],eax
+    mov     [ebp-w6],eax
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w7],edx
+    mov     [ebp-w7],edx
     shr     eax,16
-    mov     [w8],eax
+    mov     [ebp-w8],eax
     psrlq   mm7,32
     movd    eax,mm7
-    mov     [w9],eax
+    mov     [ebp-w9],eax
     jmp     .flags
 .xres_1
-    cmp     dword[xcounter],-1
+    cmp     dword[ebp-xcounter],-1
     jl      .nexty
     ; x=Xres-1 - special case
-    mov     ebx,[prevline]
+    mov     ebx,[ebp-prevline]
     movq    mm5,[esi+ebx-6]
     movq    mm6,[esi-6]
-    mov     ebx,[nextline]
+    mov     ebx,[ebp-nextline]
     movq    mm7,[esi+ebx-6]
     psrlq   mm5,32
     psrlq   mm6,32
     psrlq   mm7,32
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w1],edx
+    mov     [ebp-w1],edx
     shr     eax,16
-    mov     [w2],eax
-    mov     [w3],eax
+    mov     [ebp-w2],eax
+    mov     [ebp-w3],eax
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w4],edx
+    mov     [ebp-w4],edx
     shr     eax,16
-    mov     [w5],eax
-    mov     [w6],eax
+    mov     [ebp-w5],eax
+    mov     [ebp-w6],eax
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w7],edx
+    mov     [ebp-w7],edx
     shr     eax,16
-    mov     [w8],eax
-    mov     [w9],eax
+    mov     [ebp-w8],eax
+    mov     [ebp-w9],eax
     jmp     .flags
 .nexty
     add		esi,[ebp+offset] ; added - move source pointer past end-of-line blanks
@@ -2465,19 +2467,19 @@ NEWSYM hq3x_16
     sub     edi, ebx
     sub     edi, ebx
     sub     edi, ebx
-    dec     dword[linesleft]
+    dec     dword[ebp-linesleft]
     jz      .fin
     add		ebx, [ebp+offset];
-    cmp     dword[linesleft],1
+    cmp     dword[ebp-linesleft],1
     je      .lastline
-    mov     dword[nextline],ebx
+    mov     dword[ebp-nextline],ebx
     neg     ebx
-    mov     dword[prevline],ebx
+    mov     dword[ebp-prevline],ebx
     jmp     .loopy
 .lastline
-    mov     dword[nextline],0
+    mov     dword[ebp-nextline],0
     neg     ebx
-    mov     dword[prevline],ebx
+    mov     dword[ebp-prevline],ebx
     jmp     .loopy
 .fin
     emms
diff --git a/src/hq/asm/hq3x_32.asm b/src/hq/asm/hq3x_32.asm
index ccbf825b..ff52eb20 100644
--- a/src/hq/asm/hq3x_32.asm
+++ b/src/hq/asm/hq3x_32.asm
@@ -1,4 +1,4 @@
-;hq3x filter
+;hq3x filter (thread-safe version)
 ;32bpp output
 ;----------------------------------------------------------
 ;Copyright (C) 2003 MaxSt ( maxst@hiend3d.com )
@@ -23,29 +23,6 @@
 EXTSYM LUT16to32,RGBtoYUV
 
 SECTION .bss
-linesleft resd 1
-xcounter  resd 1
-cross     resd 1
-nextline  resd 1
-prevline  resd 1
-w1        resd 1
-w2        resd 1
-w3        resd 1
-w4        resd 1
-w5        resd 1
-w6        resd 1
-w7        resd 1
-w8        resd 1
-w9        resd 1
-c1        resd 1
-c2        resd 1
-c3        resd 1
-c4        resd 1
-c5        resd 1
-c6        resd 1
-c7        resd 1
-c8        resd 1
-c9        resd 1
 
 SECTION .data
 
@@ -183,27 +160,27 @@ SECTION .text
 %endmacro
 
 %macro PIXEL00_1M 0
-    Interp1 [edi],eax,[c1]
+    Interp1 [edi],eax,dword[ebp-c1]
 %endmacro
 
 %macro PIXEL00_1U 0
-    Interp1 [edi],eax,[c2]
+    Interp1 [edi],eax,dword[ebp-c2]
 %endmacro
 
 %macro PIXEL00_1L 0
-    Interp1 [edi],eax,[c4]
+    Interp1 [edi],eax,dword[ebp-c4]
 %endmacro
 
 %macro PIXEL00_2 0
-    Interp2 [edi],eax,[c4],[c2]
+    Interp2 [edi],eax,dword[ebp-c4],dword[ebp-c2]
 %endmacro
 
 %macro PIXEL00_4 0
-    Interp4 [edi],[c4],[c2]
+    Interp4 [edi],dword[ebp-c4],dword[ebp-c2]
 %endmacro
 
 %macro PIXEL00_5 0
-    Interp5 [edi],[c4],[c2]
+    Interp5 [edi],dword[ebp-c4],dword[ebp-c2]
 %endmacro
 
 %macro PIXEL00_C 0
@@ -211,15 +188,15 @@ SECTION .text
 %endmacro
 
 %macro PIXEL01_1 0
-    Interp1 [edi+4],eax,[c2]
+    Interp1 [edi+4],eax,dword[ebp-c2]
 %endmacro
 
 %macro PIXEL01_3 0
-    Interp3 [edi+4],[c2]
+    Interp3 [edi+4],dword[ebp-c2]
 %endmacro
 
 %macro PIXEL01_6 0
-    Interp1 [edi+4],[c2],eax
+    Interp1 [edi+4],dword[ebp-c2],eax
 %endmacro
 
 %macro PIXEL01_C 0
@@ -227,27 +204,27 @@ SECTION .text
 %endmacro
 
 %macro PIXEL02_1M 0
-    Interp1 [edi+8],eax,[c3]
+    Interp1 [edi+8],eax,dword[ebp-c3]
 %endmacro
 
 %macro PIXEL02_1U 0
-    Interp1 [edi+8],eax,[c2]
+    Interp1 [edi+8],eax,dword[ebp-c2]
 %endmacro
 
 %macro PIXEL02_1R 0
-    Interp1 [edi+8],eax,[c6]
+    Interp1 [edi+8],eax,dword[ebp-c6]
 %endmacro
 
 %macro PIXEL02_2 0
-    Interp2 [edi+8],eax,[c2],[c6]
+    Interp2 [edi+8],eax,dword[ebp-c2],dword[ebp-c6]
 %endmacro
 
 %macro PIXEL02_4 0
-    Interp4 [edi+8],[c2],[c6]
+    Interp4 [edi+8],dword[ebp-c2],dword[ebp-c6]
 %endmacro
 
 %macro PIXEL02_5 0
-    Interp5 [edi+8],[c2],[c6]
+    Interp5 [edi+8],dword[ebp-c2],dword[ebp-c6]
 %endmacro
 
 %macro PIXEL02_C 0
@@ -255,15 +232,15 @@ SECTION .text
 %endmacro
 
 %macro PIXEL10_1 0
-    Interp1 [edi+ebx],eax,[c4]
+    Interp1 [edi+ebx],eax,dword[ebp-c4]
 %endmacro
 
 %macro PIXEL10_3 0
-    Interp3 [edi+ebx],[c4]
+    Interp3 [edi+ebx],dword[ebp-c4]
 %endmacro
 
 %macro PIXEL10_6 0
-    Interp1 [edi+ebx],[c4],eax
+    Interp1 [edi+ebx],dword[ebp-c4],eax
 %endmacro
 
 %macro PIXEL10_C 0
@@ -275,15 +252,15 @@ SECTION .text
 %endmacro
 
 %macro PIXEL12_1 0
-    Interp1 [edi+ebx+8],eax,[c6]
+    Interp1 [edi+ebx+8],eax,dword[ebp-c6]
 %endmacro
 
 %macro PIXEL12_3 0
-    Interp3 [edi+ebx+8],[c6]
+    Interp3 [edi+ebx+8],dword[ebp-c6]
 %endmacro
 
 %macro PIXEL12_6 0
-    Interp1 [edi+ebx+8],[c6],eax
+    Interp1 [edi+ebx+8],dword[ebp-c6],eax
 %endmacro
 
 %macro PIXEL12_C 0
@@ -291,27 +268,27 @@ SECTION .text
 %endmacro
 
 %macro PIXEL20_1M 0
-    Interp1 [edi+ebx*2],eax,[c7]
+    Interp1 [edi+ebx*2],eax,dword[ebp-c7]
 %endmacro
 
 %macro PIXEL20_1D 0
-    Interp1 [edi+ebx*2],eax,[c8]
+    Interp1 [edi+ebx*2],eax,dword[ebp-c8]
 %endmacro
 
 %macro PIXEL20_1L 0
-    Interp1 [edi+ebx*2],eax,[c4]
+    Interp1 [edi+ebx*2],eax,dword[ebp-c4]
 %endmacro
 
 %macro PIXEL20_2 0
-    Interp2 [edi+ebx*2],eax,[c8],[c4]
+    Interp2 [edi+ebx*2],eax,dword[ebp-c8],dword[ebp-c4]
 %endmacro
 
 %macro PIXEL20_4 0
-    Interp4 [edi+ebx*2],[c8],[c4]
+    Interp4 [edi+ebx*2],dword[ebp-c8],dword[ebp-c4]
 %endmacro
 
 %macro PIXEL20_5 0
-    Interp5 [edi+ebx*2],[c8],[c4]
+    Interp5 [edi+ebx*2],dword[ebp-c8],dword[ebp-c4]
 %endmacro
 
 %macro PIXEL20_C 0
@@ -319,15 +296,15 @@ SECTION .text
 %endmacro
 
 %macro PIXEL21_1 0
-    Interp1 [edi+ebx*2+4],eax,[c8]
+    Interp1 [edi+ebx*2+4],eax,dword[ebp-c8]
 %endmacro
 
 %macro PIXEL21_3 0
-    Interp3 [edi+ebx*2+4],[c8]
+    Interp3 [edi+ebx*2+4],dword[ebp-c8]
 %endmacro
 
 %macro PIXEL21_6 0
-    Interp1 [edi+ebx*2+4],[c8],eax
+    Interp1 [edi+ebx*2+4],dword[ebp-c8],eax
 %endmacro
 
 %macro PIXEL21_C 0
@@ -335,27 +312,27 @@ SECTION .text
 %endmacro
 
 %macro PIXEL22_1M 0
-    Interp1 [edi+ebx*2+8],eax,[c9]
+    Interp1 [edi+ebx*2+8],eax,dword[ebp-c9]
 %endmacro
 
 %macro PIXEL22_1D 0
-    Interp1 [edi+ebx*2+8],eax,[c8]
+    Interp1 [edi+ebx*2+8],eax,dword[ebp-c8]
 %endmacro
 
 %macro PIXEL22_1R 0
-    Interp1 [edi+ebx*2+8],eax,[c6]
+    Interp1 [edi+ebx*2+8],eax,dword[ebp-c6]
 %endmacro
 
 %macro PIXEL22_2 0
-    Interp2 [edi+ebx*2+8],eax,[c6],[c8]
+    Interp2 [edi+ebx*2+8],eax,dword[ebp-c6],dword[ebp-c8]
 %endmacro
 
 %macro PIXEL22_4 0
-    Interp4 [edi+ebx*2+8],[c6],[c8]
+    Interp4 [edi+ebx*2+8],dword[ebp-c6],dword[ebp-c8]
 %endmacro
 
 %macro PIXEL22_5 0
-    Interp5 [edi+ebx*2+8],[c6],[c8]
+    Interp5 [edi+ebx*2+8],dword[ebp-c6],dword[ebp-c8]
 %endmacro
 
 %macro PIXEL22_C 0
@@ -369,95 +346,120 @@ Yres         equ 20
 pitch        equ 24
 offset       equ 28
 
+linesleft equ 4
+xcounter  equ 8
+cross     equ 12
+nextline  equ 16
+prevline  equ 20
+w1        equ 24
+w2        equ 28
+w3        equ 32
+w4        equ 36
+w5        equ 40
+w6        equ 44
+w7        equ 48
+w8        equ 52
+w9        equ 56
+c1        equ 60
+c2        equ 64
+c3        equ 68
+c4        equ 72
+c5        equ 76
+c6        equ 80
+c7        equ 84
+c8        equ 88
+c9        equ 92
+localsize equ 92
+
 NEWSYM hq3x_32
     push ebp
     mov ebp,esp
+    sub esp, localsize
     pushad
 
     mov     esi,[ebp+inbuffer]
     mov     edi,[ebp+outbuffer]
     mov     edx,[ebp+Yres]
-    mov     [linesleft],edx
+    mov     [ebp-linesleft],edx
     mov     ebx,[ebp+Xres]
     shl     ebx,1
-    mov     dword[prevline],0
+    mov     dword[ebp-prevline],0
     mov     eax, ebx
     add     eax, [ebp+offset]
-    mov     dword[nextline],eax
-
+    mov     dword[ebp-nextline],eax
 .loopy
     mov     ecx,[ebp+Xres]
     sub     ecx,2                 ; x={Xres-2, Xres-1} are special cases.
-    mov     dword[xcounter],ecx
+    mov     dword[ebp-xcounter],ecx
     ; x=0 - special case
-    mov     ebx,[prevline]
+    mov     ebx,[ebp-prevline]
     movq    mm5,[esi+ebx]
     movq    mm6,[esi]
-    mov     ebx,[nextline]
+    mov     ebx,[ebp-nextline]
     movq    mm7,[esi+ebx]
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w1],edx
-    mov     [w2],edx
+    mov     [ebp-w1],edx
+    mov     [ebp-w2],edx
     shr     eax,16
-    mov     [w3],eax
+    mov     [ebp-w3],eax
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w4],edx
-    mov     [w5],edx
+    mov     [ebp-w4],edx
+    mov     [ebp-w5],edx
     shr     eax,16
-    mov     [w6],eax
+    mov     [ebp-w6],eax
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w7],edx
-    mov     [w8],edx
+    mov     [ebp-w7],edx
+    mov     [ebp-w8],edx
     shr     eax,16
-    mov     [w9],eax
+    mov     [ebp-w9],eax
     jmp     .flags
 .loopx
-    mov     ebx,[prevline]
+    mov     ebx,[ebp-prevline]
     movq    mm5,[esi+ebx-2]
     movq    mm6,[esi-2]
-    mov     ebx,[nextline]
+    mov     ebx,[ebp-nextline]
     movq    mm7,[esi+ebx-2]
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w1],edx
+    mov     [ebp-w1],edx
     shr     eax,16
-    mov     [w2],eax
+    mov     [ebp-w2],eax
     psrlq   mm5,32
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w3],edx
+    mov     [ebp-w3],edx
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w4],edx
+    mov     [ebp-w4],edx
     shr     eax,16
-    mov     [w5],eax
+    mov     [ebp-w5],eax
     psrlq   mm6,32
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w6],edx
+    mov     [ebp-w6],edx
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w7],edx
+    mov     [ebp-w7],edx
     shr     eax,16
-    mov     [w8],eax
+    mov     [ebp-w8],eax
     psrlq   mm7,32
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w9],edx
+    mov     [ebp-w9],edx
 .flags
     mov     ebx,RGBtoYUV
-    mov     eax,[w5]
+    mov     eax,[ebp-w5]
     xor     ecx,ecx
     movd    mm5,[ebx+eax*4]
-    mov     dword[cross],0
+    mov     dword[ebp-cross],0
 
-    mov     edx,[w2]
+    mov     edx,[ebp-w2]
     cmp     eax,edx
     je      .noflag2
-    or      dword[cross],1
+    or      dword[ebp-cross],1
     movq    mm1,mm5
     movd    mm2,[ebx+edx*4]
     psubusb mm1,mm2
@@ -469,10 +471,10 @@ NEWSYM hq3x_32
     jz      .noflag2
     or      ecx,2
 .noflag2
-    mov     edx,[w4]
+    mov     edx,[ebp-w4]
     cmp     eax,edx
     je      .noflag4
-    or      dword[cross],2
+    or      dword[ebp-cross],2
     movq    mm1,mm5
     movd    mm2,[ebx+edx*4]
     psubusb mm1,mm2
@@ -484,10 +486,10 @@ NEWSYM hq3x_32
     jz      .noflag4
     or      ecx,8
 .noflag4
-    mov     edx,[w6]
+    mov     edx,[ebp-w6]
     cmp     eax,edx
     je      .noflag6
-    or      dword[cross],4
+    or      dword[ebp-cross],4
     movq    mm1,mm5
     movd    mm2,[ebx+edx*4]
     psubusb mm1,mm2
@@ -499,10 +501,10 @@ NEWSYM hq3x_32
     jz      .noflag6
     or      ecx,16
 .noflag6
-    mov     edx,[w8]
+    mov     edx,[ebp-w8]
     cmp     eax,edx
     je      .noflag8
-    or      dword[cross],8
+    or      dword[ebp-cross],8
     movq    mm1,mm5
     movd    mm2,[ebx+edx*4]
     psubusb mm1,mm2
@@ -516,12 +518,12 @@ NEWSYM hq3x_32
 .noflag8
     test    ecx,ecx
     jnz     .testflag1
-    mov     ecx,[cross]
+    mov     ecx,[ebp-cross]
     mov     ebx,LUT16to32
     mov     eax,[ebx+eax*4]
     jmp     [FuncTable2+ecx*4]
 .testflag1
-    mov     edx,[w1]
+    mov     edx,[ebp-w1]
     cmp     eax,edx
     je      .noflag1
     movq    mm1,mm5
@@ -535,7 +537,7 @@ NEWSYM hq3x_32
     jz      .noflag1
     or      ecx,1
 .noflag1
-    mov     edx,[w3]
+    mov     edx,[ebp-w3]
     cmp     eax,edx
     je      .noflag3
     movq    mm1,mm5
@@ -549,7 +551,7 @@ NEWSYM hq3x_32
     jz      .noflag3
     or      ecx,4
 .noflag3
-    mov     edx,[w7]
+    mov     edx,[ebp-w7]
     cmp     eax,edx
     je      .noflag7
     movq    mm1,mm5
@@ -563,7 +565,7 @@ NEWSYM hq3x_32
     jz      .noflag7
     or      ecx,32
 .noflag7
-    mov     edx,[w9]
+    mov     edx,[ebp-w9]
     cmp     eax,edx
     je      .noflag9
     movq    mm1,mm5
@@ -579,32 +581,32 @@ NEWSYM hq3x_32
 .noflag9
     mov  ebx,LUT16to32
     mov  eax,[ebx+eax*4]
-    mov  edx,[w2]
+    mov  edx,[ebp-w2]
     mov  edx,[ebx+edx*4]
-    mov  [c2],edx
-    mov  edx,[w4]
+    mov  [ebp-c2],edx
+    mov  edx,[ebp-w4]
     mov  edx,[ebx+edx*4]
-    mov  [c4],edx
-    mov  edx,[w6]
+    mov  [ebp-c4],edx
+    mov  edx,[ebp-w6]
     mov  edx,[ebx+edx*4]
-    mov  [c6],edx
-    mov  edx,[w8]
+    mov  [ebp-c6],edx
+    mov  edx,[ebp-w8]
     mov  edx,[ebx+edx*4]
-    mov  [c8],edx
+    mov  [ebp-c8],edx
     test ecx,0x005A
     jz  .switch
-    mov  edx,[w1]
+    mov  edx,[ebp-w1]
     mov  edx,[ebx+edx*4]
-    mov  [c1],edx
-    mov  edx,[w3]
+    mov  [ebp-c1],edx
+    mov  edx,[ebp-w3]
     mov  edx,[ebx+edx*4]
-    mov  [c3],edx
-    mov  edx,[w7]
+    mov  [ebp-c3],edx
+    mov  edx,[ebp-w7]
     mov  edx,[ebx+edx*4]
-    mov  [c7],edx
-    mov  edx,[w9]
+    mov  [ebp-c7],edx
+    mov  edx,[ebp-w9]
     mov  edx,[ebx+edx*4]
-    mov  [c9],edx
+    mov  [ebp-c9],edx
 .switch
     mov  ebx,[ebp+pitch]
     jmp  [FuncTable+ecx*4]
@@ -638,41 +640,41 @@ NEWSYM hq3x_32
 ;   the same, only optimized
     mov ecx,eax
     shl ecx,1
-    add ecx,[c2]
+    add ecx,[ebp-c2]
     mov edx,ecx
-    add edx,[c4]
+    add edx,[ebp-c4]
     shr edx,2
     mov [edi],edx
     mov edx,ecx
     add edx,eax
     shr edx,2
     mov [edi+4],edx
-    add ecx,[c6]
+    add ecx,[ebp-c6]
     shr ecx,2
     mov [edi+8],ecx
     mov ecx,eax
     shl ecx,2
     sub ecx,eax
     mov edx,ecx
-    add edx,[c4]
+    add edx,[ebp-c4]
     shr edx,2
     mov [edi+ebx],edx
     mov [edi+ebx+4],eax
-    add ecx,[c6]
+    add ecx,[ebp-c6]
     shr ecx,2
     mov [edi+ebx+8],ecx
     mov ecx,eax
     shl ecx,1
-    add ecx,[c8]
+    add ecx,[ebp-c8]
     mov edx,ecx
-    add edx,[c4]
+    add edx,[ebp-c4]
     shr edx,2
     mov [edi+ebx*2],edx
     mov edx,ecx
     add edx,eax
     shr edx,2
     mov [edi+ebx*2+4],edx
-    add ecx,[c6]
+    add ecx,[ebp-c6]
     shr ecx,2
     mov [edi+ebx*2+8],ecx
     jmp .loopx_end
@@ -847,7 +849,7 @@ NEWSYM hq3x_32
 ..@flag18
 ..@flag50
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_1M,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_1M,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_1
     PIXEL11
     PIXEL20_2
@@ -862,7 +864,7 @@ NEWSYM hq3x_32
     PIXEL10_1
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_1M,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_1M,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag72
 ..@flag76
@@ -871,12 +873,12 @@ NEWSYM hq3x_32
     PIXEL02_2
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_1M,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_1M,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag10
 ..@flag138
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
@@ -961,7 +963,7 @@ NEWSYM hq3x_32
 ..@flag22
 ..@flag54
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_1
     PIXEL11
     PIXEL20_2
@@ -976,7 +978,7 @@ NEWSYM hq3x_32
     PIXEL10_1
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag104
 ..@flag108
@@ -985,12 +987,12 @@ NEWSYM hq3x_32
     PIXEL02_2
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag11
 ..@flag139
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
@@ -1000,7 +1002,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag19
 ..@flag51
-    DiffOrNot w2,w6,PIXEL00_1L,PIXEL01_C,PIXEL02_1M,PIXEL12_C,PIXEL00_2,PIXEL01_6,PIXEL02_5,PIXEL12_1
+    DiffOrNot ebp-w2,ebp-w6,PIXEL00_1L,PIXEL01_C,PIXEL02_1M,PIXEL12_C,PIXEL00_2,PIXEL01_6,PIXEL02_5,PIXEL12_1
     PIXEL10_1
     PIXEL11
     PIXEL20_2
@@ -1009,7 +1011,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag146
 ..@flag178
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_1M,PIXEL12_C,PIXEL22_1D,PIXEL01_1,PIXEL02_5,PIXEL12_6,PIXEL22_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_1M,PIXEL12_C,PIXEL22_1D,PIXEL01_1,PIXEL02_5,PIXEL12_6,PIXEL22_2
     PIXEL00_1M
     PIXEL10_1
     PIXEL11
@@ -1018,7 +1020,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag84
 ..@flag85
-    DiffOrNot w6,w8,PIXEL02_1U,PIXEL12_C,PIXEL21_C,PIXEL22_1M,PIXEL02_2,PIXEL12_6,PIXEL21_1,PIXEL22_5
+    DiffOrNot ebp-w6,ebp-w8,PIXEL02_1U,PIXEL12_C,PIXEL21_C,PIXEL22_1M,PIXEL02_2,PIXEL12_6,PIXEL21_1,PIXEL22_5
     PIXEL00_2
     PIXEL01_1
     PIXEL10_1
@@ -1027,7 +1029,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag112
 ..@flag113
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL20_1L,PIXEL21_C,PIXEL22_1M,PIXEL12_1,PIXEL20_2,PIXEL21_6,PIXEL22_5
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL20_1L,PIXEL21_C,PIXEL22_1M,PIXEL12_1,PIXEL20_2,PIXEL21_6,PIXEL22_5
     PIXEL00_2
     PIXEL01_1
     PIXEL02_1M
@@ -1036,7 +1038,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag200
 ..@flag204
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_1M,PIXEL21_C,PIXEL22_1R,PIXEL10_1,PIXEL20_5,PIXEL21_6,PIXEL22_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_1M,PIXEL21_C,PIXEL22_1R,PIXEL10_1,PIXEL20_5,PIXEL21_6,PIXEL22_2
     PIXEL00_1M
     PIXEL01_1
     PIXEL02_2
@@ -1045,7 +1047,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag73
 ..@flag77
-    DiffOrNot w8,w4,PIXEL00_1U,PIXEL10_C,PIXEL20_1M,PIXEL21_C,PIXEL00_2,PIXEL10_6,PIXEL20_5,PIXEL21_1
+    DiffOrNot ebp-w8,ebp-w4,PIXEL00_1U,PIXEL10_C,PIXEL20_1M,PIXEL21_C,PIXEL00_2,PIXEL10_6,PIXEL20_5,PIXEL21_1
     PIXEL01_1
     PIXEL02_2
     PIXEL11
@@ -1054,7 +1056,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag42
 ..@flag170
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL01_C,PIXEL10_C,PIXEL20_1D,PIXEL00_5,PIXEL01_1,PIXEL10_6,PIXEL20_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL01_C,PIXEL10_C,PIXEL20_1D,PIXEL00_5,PIXEL01_1,PIXEL10_6,PIXEL20_2
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
@@ -1063,7 +1065,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag14
 ..@flag142
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL01_C,PIXEL02_1R,PIXEL10_C,PIXEL00_5,PIXEL01_6,PIXEL02_2,PIXEL10_1
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL01_C,PIXEL02_1R,PIXEL10_C,PIXEL00_5,PIXEL01_6,PIXEL02_2,PIXEL10_1
     PIXEL11
     PIXEL12_1
     PIXEL20_1M
@@ -1160,9 +1162,9 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag26
 ..@flag31
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
     PIXEL11
     PIXEL20_1M
     PIXEL21_1
@@ -1171,12 +1173,12 @@ NEWSYM hq3x_32
 ..@flag82
 ..@flag214
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag88
 ..@flag248
@@ -1184,22 +1186,22 @@ NEWSYM hq3x_32
     PIXEL01_1
     PIXEL02_1M
     PIXEL11
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
     jmp .loopx_end
 ..@flag74
 ..@flag107
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
     PIXEL02_1M
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag27
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL12_C
@@ -1209,7 +1211,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag86
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_1
     PIXEL11
     PIXEL20_1M
@@ -1223,7 +1225,7 @@ NEWSYM hq3x_32
     PIXEL10_C
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag106
     PIXEL00_1M
@@ -1231,12 +1233,12 @@ NEWSYM hq3x_32
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag30
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_C
     PIXEL11
     PIXEL20_1M
@@ -1250,7 +1252,7 @@ NEWSYM hq3x_32
     PIXEL10_1
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag120
     PIXEL00_1M
@@ -1258,11 +1260,11 @@ NEWSYM hq3x_32
     PIXEL02_1M
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag75
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
@@ -1403,9 +1405,9 @@ NEWSYM hq3x_32
     PIXEL22_1D
     jmp .loopx_end
 ..@flag58
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_C
     PIXEL11
     PIXEL12_C
@@ -1416,13 +1418,13 @@ NEWSYM hq3x_32
 ..@flag83
     PIXEL00_1L
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1M
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag92
     PIXEL00_1M
@@ -1431,36 +1433,36 @@ NEWSYM hq3x_32
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag202
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
     PIXEL02_1M
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
     PIXEL22_1R
     jmp .loopx_end
 ..@flag78
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
     PIXEL02_1R
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
     PIXEL22_1M
     jmp .loopx_end
 ..@flag154
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_C
     PIXEL11
     PIXEL12_C
@@ -1471,13 +1473,13 @@ NEWSYM hq3x_32
 ..@flag114
     PIXEL00_1M
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1L
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag89
     PIXEL00_1U
@@ -1486,24 +1488,24 @@ NEWSYM hq3x_32
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag90
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag55
 ..@flag23
-    DiffOrNot w2,w6,PIXEL00_1L,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL00_2,PIXEL01_6,PIXEL02_5,PIXEL12_1
+    DiffOrNot ebp-w2,ebp-w6,PIXEL00_1L,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL00_2,PIXEL01_6,PIXEL02_5,PIXEL12_1
     PIXEL10_1
     PIXEL11
     PIXEL20_2
@@ -1512,7 +1514,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag182
 ..@flag150
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL22_1D,PIXEL01_1,PIXEL02_5,PIXEL12_6,PIXEL22_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL22_1D,PIXEL01_1,PIXEL02_5,PIXEL12_6,PIXEL22_2
     PIXEL00_1M
     PIXEL10_1
     PIXEL11
@@ -1521,7 +1523,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag213
 ..@flag212
-    DiffOrNot w6,w8,PIXEL02_1U,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL02_2,PIXEL12_6,PIXEL21_1,PIXEL22_5
+    DiffOrNot ebp-w6,ebp-w8,PIXEL02_1U,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL02_2,PIXEL12_6,PIXEL21_1,PIXEL22_5
     PIXEL00_2
     PIXEL01_1
     PIXEL10_1
@@ -1530,7 +1532,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag241
 ..@flag240
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL20_1L,PIXEL21_C,PIXEL22_C,PIXEL12_1,PIXEL20_2,PIXEL21_6,PIXEL22_5
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL20_1L,PIXEL21_C,PIXEL22_C,PIXEL12_1,PIXEL20_2,PIXEL21_6,PIXEL22_5
     PIXEL00_2
     PIXEL01_1
     PIXEL02_1M
@@ -1539,7 +1541,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag236
 ..@flag232
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL22_1R,PIXEL10_1,PIXEL20_5,PIXEL21_6,PIXEL22_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL22_1R,PIXEL10_1,PIXEL20_5,PIXEL21_6,PIXEL22_2
     PIXEL00_1M
     PIXEL01_1
     PIXEL02_2
@@ -1548,7 +1550,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag109
 ..@flag105
-    DiffOrNot w8,w4,PIXEL00_1U,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL00_2,PIXEL10_6,PIXEL20_5,PIXEL21_1
+    DiffOrNot ebp-w8,ebp-w4,PIXEL00_1U,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL00_2,PIXEL10_6,PIXEL20_5,PIXEL21_1
     PIXEL01_1
     PIXEL02_2
     PIXEL11
@@ -1557,7 +1559,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag171
 ..@flag43
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL20_1D,PIXEL00_5,PIXEL01_1,PIXEL10_6,PIXEL20_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL20_1D,PIXEL00_5,PIXEL01_1,PIXEL10_6,PIXEL20_2
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
@@ -1566,7 +1568,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag143
 ..@flag15
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL02_1R,PIXEL10_C,PIXEL00_5,PIXEL01_6,PIXEL02_2,PIXEL10_1
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL02_1R,PIXEL10_C,PIXEL00_5,PIXEL01_6,PIXEL02_2,PIXEL10_1
     PIXEL11
     PIXEL12_1
     PIXEL20_1M
@@ -1579,11 +1581,11 @@ NEWSYM hq3x_32
     PIXEL02_1U
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag203
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
@@ -1593,7 +1595,7 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag62
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_C
     PIXEL11
     PIXEL20_1D
@@ -1607,11 +1609,11 @@ NEWSYM hq3x_32
     PIXEL10_1
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag118
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_1
     PIXEL11
     PIXEL20_1L
@@ -1625,7 +1627,7 @@ NEWSYM hq3x_32
     PIXEL10_C
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag110
     PIXEL00_1M
@@ -1633,11 +1635,11 @@ NEWSYM hq3x_32
     PIXEL02_1R
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag155
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL12_C
@@ -1739,12 +1741,12 @@ NEWSYM hq3x_32
     PIXEL02_1U
     PIXEL10_C
     PIXEL11
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag158
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_C
     PIXEL11
     PIXEL20_1M
@@ -1752,26 +1754,26 @@ NEWSYM hq3x_32
     PIXEL22_1D
     jmp .loopx_end
 ..@flag234
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
     PIXEL02_1M
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1R
     jmp .loopx_end
 ..@flag242
     PIXEL00_1M
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL20_1L
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag59
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL11
     PIXEL12_C
     PIXEL20_1D
@@ -1784,62 +1786,62 @@ NEWSYM hq3x_32
     PIXEL02_1M
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag87
     PIXEL00_1L
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_1
     PIXEL11
     PIXEL20_1M
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag79
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1R
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
     PIXEL22_1M
     jmp .loopx_end
 ..@flag122
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag94
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL10_C
     PIXEL11
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag218
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_C
     PIXEL11
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag91
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag229
     PIXEL00_2
@@ -1886,9 +1888,9 @@ NEWSYM hq3x_32
     PIXEL22_1D
     jmp .loopx_end
 ..@flag186
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_C
     PIXEL11
     PIXEL12_C
@@ -1899,13 +1901,13 @@ NEWSYM hq3x_32
 ..@flag115
     PIXEL00_1L
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1L
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag93
     PIXEL00_1U
@@ -1914,18 +1916,18 @@ NEWSYM hq3x_32
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag206
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
     PIXEL02_1R
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
     PIXEL22_1R
     jmp .loopx_end
@@ -1937,13 +1939,13 @@ NEWSYM hq3x_32
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_1M,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_1M,PIXEL20_2
     PIXEL21_C
     PIXEL22_1R
     jmp .loopx_end
 ..@flag174
 ..@flag46
-    DiffOrNot w4,w2,PIXEL00_1M,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_1M,PIXEL00_2
     PIXEL01_C
     PIXEL02_1R
     PIXEL10_C
@@ -1957,7 +1959,7 @@ NEWSYM hq3x_32
 ..@flag147
     PIXEL00_1L
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_1M,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_1M,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
@@ -1975,7 +1977,7 @@ NEWSYM hq3x_32
     PIXEL12_C
     PIXEL20_1L
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_1M,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_1M,PIXEL22_2
     jmp .loopx_end
 ..@flag189
     PIXEL00_1U
@@ -2001,20 +2003,20 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag126
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_4,PIXEL12_3
     PIXEL11
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag219
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_4,PIXEL01_3,PIXEL10_3
     PIXEL02_1M
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag125
-    DiffOrNot w8,w4,PIXEL00_1U,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL00_2,PIXEL10_6,PIXEL20_5,PIXEL21_1
+    DiffOrNot ebp-w8,ebp-w4,PIXEL00_1U,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL00_2,PIXEL10_6,PIXEL20_5,PIXEL21_1
     PIXEL01_1
     PIXEL02_1U
     PIXEL11
@@ -2022,7 +2024,7 @@ NEWSYM hq3x_32
     PIXEL22_1M
     jmp .loopx_end
 ..@flag221
-    DiffOrNot w6,w8,PIXEL02_1U,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL02_2,PIXEL12_6,PIXEL21_1,PIXEL22_5
+    DiffOrNot ebp-w6,ebp-w8,PIXEL02_1U,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL02_2,PIXEL12_6,PIXEL21_1,PIXEL22_5
     PIXEL00_1U
     PIXEL01_1
     PIXEL10_C
@@ -2030,7 +2032,7 @@ NEWSYM hq3x_32
     PIXEL20_1M
     jmp .loopx_end
 ..@flag207
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL02_1R,PIXEL10_C,PIXEL00_5,PIXEL01_6,PIXEL02_2,PIXEL10_1
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL02_1R,PIXEL10_C,PIXEL00_5,PIXEL01_6,PIXEL02_2,PIXEL10_1
     PIXEL11
     PIXEL12_1
     PIXEL20_1M
@@ -2038,7 +2040,7 @@ NEWSYM hq3x_32
     PIXEL22_1R
     jmp .loopx_end
 ..@flag238
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL22_1R,PIXEL10_1,PIXEL20_5,PIXEL21_6,PIXEL22_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL22_1R,PIXEL10_1,PIXEL20_5,PIXEL21_6,PIXEL22_2
     PIXEL00_1M
     PIXEL01_C
     PIXEL02_1R
@@ -2046,7 +2048,7 @@ NEWSYM hq3x_32
     PIXEL12_1
     jmp .loopx_end
 ..@flag190
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL22_1D,PIXEL01_1,PIXEL02_5,PIXEL12_6,PIXEL22_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL22_1D,PIXEL01_1,PIXEL02_5,PIXEL12_6,PIXEL22_2
     PIXEL00_1M
     PIXEL10_C
     PIXEL11
@@ -2054,7 +2056,7 @@ NEWSYM hq3x_32
     PIXEL21_1
     jmp .loopx_end
 ..@flag187
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL20_1D,PIXEL00_5,PIXEL01_1,PIXEL10_6,PIXEL20_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL20_1D,PIXEL00_5,PIXEL01_1,PIXEL10_6,PIXEL20_2
     PIXEL02_1M
     PIXEL11
     PIXEL12_C
@@ -2062,7 +2064,7 @@ NEWSYM hq3x_32
     PIXEL22_1D
     jmp .loopx_end
 ..@flag243
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL20_1L,PIXEL21_C,PIXEL22_C,PIXEL12_1,PIXEL20_2,PIXEL21_6,PIXEL22_5
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL20_1L,PIXEL21_C,PIXEL22_C,PIXEL12_1,PIXEL20_2,PIXEL21_6,PIXEL22_5
     PIXEL00_1L
     PIXEL01_C
     PIXEL02_1M
@@ -2070,7 +2072,7 @@ NEWSYM hq3x_32
     PIXEL11
     jmp .loopx_end
 ..@flag119
-    DiffOrNot w2,w6,PIXEL00_1L,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL00_2,PIXEL01_6,PIXEL02_5,PIXEL12_1
+    DiffOrNot ebp-w2,ebp-w6,PIXEL00_1L,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL00_2,PIXEL01_6,PIXEL02_5,PIXEL12_1
     PIXEL10_1
     PIXEL11
     PIXEL20_1L
@@ -2085,13 +2087,13 @@ NEWSYM hq3x_32
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL20_2
     PIXEL21_C
     PIXEL22_1R
     jmp .loopx_end
 ..@flag175
 ..@flag47
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL00_2
     PIXEL01_C
     PIXEL02_1R
     PIXEL10_C
@@ -2105,7 +2107,7 @@ NEWSYM hq3x_32
 ..@flag151
     PIXEL00_1L
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
@@ -2123,30 +2125,30 @@ NEWSYM hq3x_32
     PIXEL12_C
     PIXEL20_1L
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_C,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_C,PIXEL22_2
     jmp .loopx_end
 ..@flag250
     PIXEL00_1M
     PIXEL01_C
     PIXEL02_1M
     PIXEL11
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
     jmp .loopx_end
 ..@flag123
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
     PIXEL02_1M
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag95
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
     PIXEL11
     PIXEL20_1M
     PIXEL21_C
@@ -2154,12 +2156,12 @@ NEWSYM hq3x_32
     jmp .loopx_end
 ..@flag222
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
     PIXEL10_C
     PIXEL11
     PIXEL12_C
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag252
     PIXEL00_1M
@@ -2167,9 +2169,9 @@ NEWSYM hq3x_32
     PIXEL02_1U
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_C,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_C,PIXEL22_2
     jmp .loopx_end
 ..@flag249
     PIXEL00_1U
@@ -2177,34 +2179,34 @@ NEWSYM hq3x_32
     PIXEL02_1M
     PIXEL10_C
     PIXEL11
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
     jmp .loopx_end
 ..@flag235
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
     PIXEL02_1M
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL20_2
     PIXEL21_C
     PIXEL22_1R
     jmp .loopx_end
 ..@flag111
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL00_2
     PIXEL01_C
     PIXEL02_1R
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag63
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
     PIXEL10_C
     PIXEL11
     PIXEL20_1D
@@ -2212,9 +2214,9 @@ NEWSYM hq3x_32
     PIXEL22_1M
     jmp .loopx_end
 ..@flag159
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL02_2
     PIXEL11
     PIXEL12_C
     PIXEL20_1M
@@ -2224,29 +2226,29 @@ NEWSYM hq3x_32
 ..@flag215
     PIXEL00_1L
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag246
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1L
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_C,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_C,PIXEL22_2
     jmp .loopx_end
 ..@flag254
     PIXEL00_1M
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL01_3,PIXEL02_4
     PIXEL11
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL10_3,PIXEL20_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL21_C,PIXEL22_C,PIXEL12_3,PIXEL21_3,PIXEL22_2
     jmp .loopx_end
 ..@flag253
     PIXEL00_1U
@@ -2255,39 +2257,39 @@ NEWSYM hq3x_32
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_C,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_C,PIXEL22_2
     jmp .loopx_end
 ..@flag251
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL00_4,PIXEL01_3
     PIXEL02_1M
     PIXEL11
-    DiffOrNot w8,w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_2,PIXEL21_3
-    DiffOrNot w6,w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
+    DiffOrNot ebp-w8,ebp-w4,PIXEL10_C,PIXEL20_C,PIXEL21_C,PIXEL10_3,PIXEL20_2,PIXEL21_3
+    DiffOrNot ebp-w6,ebp-w8,PIXEL12_C,PIXEL22_C,PIXEL12_3,PIXEL22_4
     jmp .loopx_end
 ..@flag239
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL00_2
     PIXEL01_C
     PIXEL02_1R
     PIXEL10_C
     PIXEL11
     PIXEL12_1
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL20_2
     PIXEL21_C
     PIXEL22_1R
     jmp .loopx_end
 ..@flag127
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_2,PIXEL01_3,PIXEL10_3
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL01_C,PIXEL10_C,PIXEL00_2,PIXEL01_3,PIXEL10_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL12_C,PIXEL02_4,PIXEL12_3
     PIXEL11
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL21_C,PIXEL20_4,PIXEL21_3
     PIXEL22_1M
     jmp .loopx_end
 ..@flag191
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL02_2
     PIXEL10_C
     PIXEL11
     PIXEL12_C
@@ -2296,33 +2298,33 @@ NEWSYM hq3x_32
     PIXEL22_1D
     jmp .loopx_end
 ..@flag223
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
-    DiffOrNot w2,w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_2,PIXEL12_3
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL10_C,PIXEL00_4,PIXEL10_3
+    DiffOrNot ebp-w2,ebp-w6,PIXEL01_C,PIXEL02_C,PIXEL12_C,PIXEL01_3,PIXEL02_2,PIXEL12_3
     PIXEL11
     PIXEL20_1M
-    DiffOrNot w6,w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
+    DiffOrNot ebp-w6,ebp-w8,PIXEL21_C,PIXEL22_C,PIXEL21_3,PIXEL22_4
     jmp .loopx_end
 ..@flag247
     PIXEL00_1L
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL02_2
     PIXEL10_1
     PIXEL11
     PIXEL12_C
     PIXEL20_1L
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_C,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_C,PIXEL22_2
     jmp .loopx_end
 ..@flag255
-    DiffOrNot w4,w2,PIXEL00_C,PIXEL00_2
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_C,PIXEL00_2
     PIXEL01_C
-    DiffOrNot w2,w6,PIXEL02_C,PIXEL02_2
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_C,PIXEL02_2
     PIXEL10_C
     PIXEL11
     PIXEL12_C
-    DiffOrNot w8,w4,PIXEL20_C,PIXEL20_2
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_C,PIXEL20_2
     PIXEL21_C
-    DiffOrNot w6,w8,PIXEL22_C,PIXEL22_2
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_C,PIXEL22_2
     jmp .loopx_end
 
 ..@cross0
@@ -2338,7 +2340,7 @@ NEWSYM hq3x_32
     mov     [edi+ebx*2+8],eax
     jmp     .loopx_end
 ..@cross1
-    mov     ecx,[w2]
+    mov     ecx,[ebp-w2]
     mov     edx,eax
     shl     edx,2
     add     edx,[ebx+ecx*4]
@@ -2356,7 +2358,7 @@ NEWSYM hq3x_32
     mov     [edi+ebx*2+8],eax
     jmp     .loopx_end
 ..@cross2
-    mov     ecx,[w4]
+    mov     ecx,[ebp-w4]
     mov     edx,eax
     shl     edx,2
     add     edx,[ebx+ecx*4]
@@ -2374,7 +2376,7 @@ NEWSYM hq3x_32
     mov     [edi+ebx*2+8],eax
     jmp     .loopx_end
 ..@cross4
-    mov     ecx,[w6]
+    mov     ecx,[ebp-w6]
     mov     edx,eax
     shl     edx,2
     add     edx,[ebx+ecx*4]
@@ -2392,7 +2394,7 @@ NEWSYM hq3x_32
     mov     [edi+ebx*2+8],edx
     jmp     .loopx_end
 ..@cross8
-    mov     ecx,[w8]
+    mov     ecx,[ebp-w8]
     mov     edx,eax
     shl     edx,2
     add     edx,[ebx+ecx*4]
@@ -2410,93 +2412,93 @@ NEWSYM hq3x_32
     mov     [edi+ebx*2+8],edx
     jmp     .loopx_end
 ..@crossN
-    mov     edx,[w2]
+    mov     edx,[ebp-w2]
     mov     ecx,[ebx+edx*4]
-    mov     [c2],ecx
-    mov     edx,[w4]
+    mov     [ebp-c2],ecx
+    mov     edx,[ebp-w4]
     mov     ecx,[ebx+edx*4]
-    mov     [c4],ecx
-    mov     edx,[w6]
+    mov     [ebp-c4],ecx
+    mov     edx,[ebp-w6]
     mov     ecx,[ebx+edx*4]
-    mov     [c6],ecx
-    mov     edx,[w8]
+    mov     [ebp-c6],ecx
+    mov     edx,[ebp-w8]
     mov     ecx,[ebx+edx*4]
-    mov     [c8],ecx
+    mov     [ebp-c8],ecx
     mov     ebx,[ebp+pitch]
     jmp     ..@flag0
 
 .loopx_end
     add     esi,2
     add     edi,12
-    dec     dword[xcounter]
+    dec     dword[ebp-xcounter]
     jle     .xres_2
     jmp     .loopx
 .xres_2
     ; x=Xres-2 - special case
     jl      .xres_1
-    mov     ebx,[prevline]
+    mov     ebx,[ebp-prevline]
     movq    mm5,[esi+ebx-4]
     movq    mm6,[esi-4]
-    mov     ebx,[nextline]
+    mov     ebx,[ebp-nextline]
     movq    mm7,[esi+ebx-4]
     psrlq   mm5,16
     psrlq   mm6,16
     psrlq   mm7,16
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w1],edx
+    mov     [ebp-w1],edx
     shr     eax,16
-    mov     [w2],eax
+    mov     [ebp-w2],eax
     psrlq   mm5,32
     movd    eax,mm5
-    mov     [w3],eax
+    mov     [ebp-w3],eax
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w4],edx
+    mov     [ebp-w4],edx
     shr     eax,16
-    mov     [w5],eax
+    mov     [ebp-w5],eax
     psrlq   mm6,32
     movd    eax,mm6
-    mov     [w6],eax
+    mov     [ebp-w6],eax
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w7],edx
+    mov     [ebp-w7],edx
     shr     eax,16
-    mov     [w8],eax
+    mov     [ebp-w8],eax
     psrlq   mm7,32
     movd    eax,mm7
-    mov     [w9],eax
+    mov     [ebp-w9],eax
     jmp     .flags
 .xres_1
-    cmp     dword[xcounter],-1
+    cmp     dword[ebp-xcounter],-1
     jl      .nexty
     ; x=Xres-1 - special case
-    mov     ebx,[prevline]
+    mov     ebx,[ebp-prevline]
     movq    mm5,[esi+ebx-6]
     movq    mm6,[esi-6]
-    mov     ebx,[nextline]
+    mov     ebx,[ebp-nextline]
     movq    mm7,[esi+ebx-6]
     psrlq   mm5,32
     psrlq   mm6,32
     psrlq   mm7,32
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w1],edx
+    mov     [ebp-w1],edx
     shr     eax,16
-    mov     [w2],eax
-    mov     [w3],eax
+    mov     [ebp-w2],eax
+    mov     [ebp-w3],eax
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w4],edx
+    mov     [ebp-w4],edx
     shr     eax,16
-    mov     [w5],eax
-    mov     [w6],eax
+    mov     [ebp-w5],eax
+    mov     [ebp-w6],eax
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w7],edx
+    mov     [ebp-w7],edx
     shr     eax,16
-    mov     [w8],eax
-    mov     [w9],eax
+    mov     [ebp-w8],eax
+    mov     [ebp-w9],eax
     jmp     .flags
 .nexty
     add		esi,[ebp+offset] ; added - move source pointer past end-of-line blanks
@@ -2509,19 +2511,19 @@ NEWSYM hq3x_32
     sub     edi, ebx
     sub     edi, ebx
 	shr     ebx, 1
-    dec     dword[linesleft]
+    dec     dword[ebp-linesleft]
     jz      .fin
     add		ebx, [ebp+offset];
-    cmp     dword[linesleft],1
+    cmp     dword[ebp-linesleft],1
     je      .lastline
-    mov     dword[nextline],ebx
+    mov     dword[ebp-nextline],ebx
     neg     ebx
-    mov     dword[prevline],ebx
+    mov     dword[ebp-prevline],ebx
     jmp     .loopy
 .lastline
-    mov     dword[nextline],0
+    mov     dword[ebp-nextline],0
     neg     ebx
-    mov     dword[prevline],ebx
+    mov     dword[ebp-prevline],ebx
     jmp     .loopy
 .fin
     emms
diff --git a/src/hq/asm/hq4x_16.asm b/src/hq/asm/hq4x_16.asm
index 6d4a6554..b7b26a51 100644
--- a/src/hq/asm/hq4x_16.asm
+++ b/src/hq/asm/hq4x_16.asm
@@ -1,4 +1,4 @@
-;hq4x filter
+;hq4x filter (thread-safe version)
 ;16bpp output
 ;----------------------------------------------------------
 ;Copyright (C) 2003 MaxSt ( maxst@hiend3d.com )
@@ -23,20 +23,6 @@
 EXTSYM LUT16to32,RGBtoYUV
 
 SECTION .bss
-linesleft resd 1
-xcounter  resd 1
-cross     resd 1
-nextline  resd 1
-prevline  resd 1
-w1        resd 1
-w2        resd 1
-w3        resd 1
-w4        resd 1
-w5        resd 1
-w6        resd 1
-w7        resd 1
-w8        resd 1
-w9        resd 1
 
 SECTION .data
 
@@ -284,579 +270,579 @@ SECTION .text
 %endmacro
 
 %macro PIXEL00_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [edi],ax
 %endmacro
 
 %macro PIXEL00_11 0
-    Interp1 [edi],[w5],[w4]
+    Interp1 [edi],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL00_12 0
-    Interp1 [edi],[w5],[w2]
+    Interp1 [edi],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL00_20 0
-    Interp2 [edi],[w5],[w2],[w4]
+    Interp2 [edi],dword[ebp-w5],dword[ebp-w2],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL00_50 0
-    Interp5 [edi],[w2],[w4]
+    Interp5 [edi],dword[ebp-w2],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL00_80 0
-    Interp8 [edi],[w5],[w1]
+    Interp8 [edi],dword[ebp-w5],dword[ebp-w1]
 %endmacro
 
 %macro PIXEL00_81 0
-    Interp8 [edi],[w5],[w4]
+    Interp8 [edi],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL00_82 0
-    Interp8 [edi],[w5],[w2]
+    Interp8 [edi],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL01_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [edi+2],ax
 %endmacro
 
 %macro PIXEL01_10 0
-    Interp1 [edi+2],[w5],[w1]
+    Interp1 [edi+2],dword[ebp-w5],dword[ebp-w1]
 %endmacro
 
 %macro PIXEL01_12 0
-    Interp1 [edi+2],[w5],[w2]
+    Interp1 [edi+2],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL01_14 0
-    Interp1 [edi+2],[w2],[w5]
+    Interp1 [edi+2],dword[ebp-w2],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL01_21 0
-    Interp2 [edi+2],[w2],[w5],[w4]
+    Interp2 [edi+2],dword[ebp-w2],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL01_31 0
-    Interp3 [edi+2],[w5],[w4]
+    Interp3 [edi+2],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL01_50 0
-    Interp5 [edi+2],[w2],[w5]
+    Interp5 [edi+2],dword[ebp-w2],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL01_60 0
-    Interp6 [edi+2],[w5],[w2],[w4]
+    Interp6 [edi+2],dword[ebp-w5],dword[ebp-w2],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL01_61 0
-    Interp6 [edi+2],[w5],[w2],[w1]
+    Interp6 [edi+2],dword[ebp-w5],dword[ebp-w2],dword[ebp-w1]
 %endmacro
 
 %macro PIXEL01_82 0
-    Interp8 [edi+2],[w5],[w2]
+    Interp8 [edi+2],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL01_83 0
-    Interp8 [edi+2],[w2],[w4]
+    Interp8 [edi+2],dword[ebp-w2],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL02_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [edi+4],ax
 %endmacro
 
 %macro PIXEL02_10 0
-    Interp1 [edi+4],[w5],[w3]
+    Interp1 [edi+4],dword[ebp-w5],dword[ebp-w3]
 %endmacro
 
 %macro PIXEL02_11 0
-    Interp1 [edi+4],[w5],[w2]
+    Interp1 [edi+4],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL02_13 0
-    Interp1 [edi+4],[w2],[w5]
+    Interp1 [edi+4],dword[ebp-w2],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL02_21 0
-    Interp2 [edi+4],[w2],[w5],[w6]
+    Interp2 [edi+4],dword[ebp-w2],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL02_32 0
-    Interp3 [edi+4],[w5],[w6]
+    Interp3 [edi+4],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL02_50 0
-    Interp5 [edi+4],[w2],[w5]
+    Interp5 [edi+4],dword[ebp-w2],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL02_60 0
-    Interp6 [edi+4],[w5],[w2],[w6]
+    Interp6 [edi+4],dword[ebp-w5],dword[ebp-w2],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL02_61 0
-    Interp6 [edi+4],[w5],[w2],[w3]
+    Interp6 [edi+4],dword[ebp-w5],dword[ebp-w2],dword[ebp-w3]
 %endmacro
 
 %macro PIXEL02_81 0
-    Interp8 [edi+4],[w5],[w2]
+    Interp8 [edi+4],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL02_83 0
-    Interp8 [edi+4],[w2],[w6]
+    Interp8 [edi+4],dword[ebp-w2],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL03_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [edi+6],ax
 %endmacro
 
 %macro PIXEL03_11 0
-    Interp1 [edi+6],[w5],[w2]
+    Interp1 [edi+6],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL03_12 0
-    Interp1 [edi+6],[w5],[w6]
+    Interp1 [edi+6],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL03_20 0
-    Interp2 [edi+6],[w5],[w2],[w6]
+    Interp2 [edi+6],dword[ebp-w5],dword[ebp-w2],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL03_50 0
-    Interp5 [edi+6],[w2],[w6]
+    Interp5 [edi+6],dword[ebp-w2],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL03_80 0
-    Interp8 [edi+6],[w5],[w3]
+    Interp8 [edi+6],dword[ebp-w5],dword[ebp-w3]
 %endmacro
 
 %macro PIXEL03_81 0
-    Interp8 [edi+6],[w5],[w2]
+    Interp8 [edi+6],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL03_82 0
-    Interp8 [edi+6],[w5],[w6]
+    Interp8 [edi+6],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL10_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [edi+ebx],ax
 %endmacro
 
 %macro PIXEL10_10 0
-    Interp1 [edi+ebx],[w5],[w1]
+    Interp1 [edi+ebx],dword[ebp-w5],dword[ebp-w1]
 %endmacro
 
 %macro PIXEL10_11 0
-    Interp1 [edi+ebx],[w5],[w4]
+    Interp1 [edi+ebx],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL10_13 0
-    Interp1 [edi+ebx],[w4],[w5]
+    Interp1 [edi+ebx],dword[ebp-w4],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL10_21 0
-    Interp2 [edi+ebx],[w4],[w5],[w2]
+    Interp2 [edi+ebx],dword[ebp-w4],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL10_32 0
-    Interp3 [edi+ebx],[w5],[w2]
+    Interp3 [edi+ebx],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL10_50 0
-    Interp5 [edi+ebx],[w4],[w5]
+    Interp5 [edi+ebx],dword[ebp-w4],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL10_60 0
-    Interp6 [edi+ebx],[w5],[w4],[w2]
+    Interp6 [edi+ebx],dword[ebp-w5],dword[ebp-w4],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL10_61 0
-    Interp6 [edi+ebx],[w5],[w4],[w1]
+    Interp6 [edi+ebx],dword[ebp-w5],dword[ebp-w4],dword[ebp-w1]
 %endmacro
 
 %macro PIXEL10_81 0
-    Interp8 [edi+ebx],[w5],[w4]
+    Interp8 [edi+ebx],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL10_83 0
-    Interp8 [edi+ebx],[w4],[w2]
+    Interp8 [edi+ebx],dword[ebp-w4],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL11_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [edi+ebx+2],ax
 %endmacro
 
 %macro PIXEL11_30 0
-    Interp3 [edi+ebx+2],[w5],[w1]
+    Interp3 [edi+ebx+2],dword[ebp-w5],dword[ebp-w1]
 %endmacro
 
 %macro PIXEL11_31 0
-    Interp3 [edi+ebx+2],[w5],[w4]
+    Interp3 [edi+ebx+2],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL11_32 0
-    Interp3 [edi+ebx+2],[w5],[w2]
+    Interp3 [edi+ebx+2],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL11_70 0
-    Interp7 [edi+ebx+2],[w5],[w4],[w2]
+    Interp7 [edi+ebx+2],dword[ebp-w5],dword[ebp-w4],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL12_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [edi+ebx+4],ax
 %endmacro
 
 %macro PIXEL12_30 0
-    Interp3 [edi+ebx+4],[w5],[w3]
+    Interp3 [edi+ebx+4],dword[ebp-w5],dword[ebp-w3]
 %endmacro
 
 %macro PIXEL12_31 0
-    Interp3 [edi+ebx+4],[w5],[w2]
+    Interp3 [edi+ebx+4],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL12_32 0
-    Interp3 [edi+ebx+4],[w5],[w6]
+    Interp3 [edi+ebx+4],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL12_70 0
-    Interp7 [edi+ebx+4],[w5],[w6],[w2]
+    Interp7 [edi+ebx+4],dword[ebp-w5],dword[ebp-w6],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL13_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [edi+ebx+6],ax
 %endmacro
 
 %macro PIXEL13_10 0
-    Interp1 [edi+ebx+6],[w5],[w3]
+    Interp1 [edi+ebx+6],dword[ebp-w5],dword[ebp-w3]
 %endmacro
 
 %macro PIXEL13_12 0
-    Interp1 [edi+ebx+6],[w5],[w6]
+    Interp1 [edi+ebx+6],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL13_14 0
-    Interp1 [edi+ebx+6],[w6],[w5]
+    Interp1 [edi+ebx+6],dword[ebp-w6],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL13_21 0
-    Interp2 [edi+ebx+6],[w6],[w5],[w2]
+    Interp2 [edi+ebx+6],dword[ebp-w6],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL13_31 0
-    Interp3 [edi+ebx+6],[w5],[w2]
+    Interp3 [edi+ebx+6],dword[ebp-w5],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL13_50 0
-    Interp5 [edi+ebx+6],[w6],[w5]
+    Interp5 [edi+ebx+6],dword[ebp-w6],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL13_60 0
-    Interp6 [edi+ebx+6],[w5],[w6],[w2]
+    Interp6 [edi+ebx+6],dword[ebp-w5],dword[ebp-w6],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL13_61 0
-    Interp6 [edi+ebx+6],[w5],[w6],[w3]
+    Interp6 [edi+ebx+6],dword[ebp-w5],dword[ebp-w6],dword[ebp-w3]
 %endmacro
 
 %macro PIXEL13_82 0
-    Interp8 [edi+ebx+6],[w5],[w6]
+    Interp8 [edi+ebx+6],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL13_83 0
-    Interp8 [edi+ebx+6],[w6],[w2]
+    Interp8 [edi+ebx+6],dword[ebp-w6],dword[ebp-w2]
 %endmacro
 
 %macro PIXEL20_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [ecx],ax
 %endmacro
 
 %macro PIXEL20_10 0
-    Interp1 [ecx],[w5],[w7]
+    Interp1 [ecx],dword[ebp-w5],dword[ebp-w7]
 %endmacro
 
 %macro PIXEL20_12 0
-    Interp1 [ecx],[w5],[w4]
+    Interp1 [ecx],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL20_14 0
-    Interp1 [ecx],[w4],[w5]
+    Interp1 [ecx],dword[ebp-w4],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL20_21 0
-    Interp2 [ecx],[w4],[w5],[w8]
+    Interp2 [ecx],dword[ebp-w4],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL20_31 0
-    Interp3 [ecx],[w5],[w8]
+    Interp3 [ecx],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL20_50 0
-    Interp5 [ecx],[w4],[w5]
+    Interp5 [ecx],dword[ebp-w4],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL20_60 0
-    Interp6 [ecx],[w5],[w4],[w8]
+    Interp6 [ecx],dword[ebp-w5],dword[ebp-w4],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL20_61 0
-    Interp6 [ecx],[w5],[w4],[w7]
+    Interp6 [ecx],dword[ebp-w5],dword[ebp-w4],dword[ebp-w7]
 %endmacro
 
 %macro PIXEL20_82 0
-    Interp8 [ecx],[w5],[w4]
+    Interp8 [ecx],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL20_83 0
-    Interp8 [ecx],[w4],[w8]
+    Interp8 [ecx],dword[ebp-w4],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL21_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [ecx+2],ax
 %endmacro
 
 %macro PIXEL21_30 0
-    Interp3 [ecx+2],[w5],[w7]
+    Interp3 [ecx+2],dword[ebp-w5],dword[ebp-w7]
 %endmacro
 
 %macro PIXEL21_31 0
-    Interp3 [ecx+2],[w5],[w8]
+    Interp3 [ecx+2],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL21_32 0
-    Interp3 [ecx+2],[w5],[w4]
+    Interp3 [ecx+2],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL21_70 0
-    Interp7 [ecx+2],[w5],[w4],[w8]
+    Interp7 [ecx+2],dword[ebp-w5],dword[ebp-w4],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL22_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [ecx+4],ax
 %endmacro
 
 %macro PIXEL22_30 0
-    Interp3 [ecx+4],[w5],[w9]
+    Interp3 [ecx+4],dword[ebp-w5],dword[ebp-w9]
 %endmacro
 
 %macro PIXEL22_31 0
-    Interp3 [ecx+4],[w5],[w6]
+    Interp3 [ecx+4],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL22_32 0
-    Interp3 [ecx+4],[w5],[w8]
+    Interp3 [ecx+4],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL22_70 0
-    Interp7 [ecx+4],[w5],[w6],[w8]
+    Interp7 [ecx+4],dword[ebp-w5],dword[ebp-w6],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL23_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [ecx+6],ax
 %endmacro
 
 %macro PIXEL23_10 0
-    Interp1 [ecx+6],[w5],[w9]
+    Interp1 [ecx+6],dword[ebp-w5],dword[ebp-w9]
 %endmacro
 
 %macro PIXEL23_11 0
-    Interp1 [ecx+6],[w5],[w6]
+    Interp1 [ecx+6],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL23_13 0
-    Interp1 [ecx+6],[w6],[w5]
+    Interp1 [ecx+6],dword[ebp-w6],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL23_21 0
-    Interp2 [ecx+6],[w6],[w5],[w8]
+    Interp2 [ecx+6],dword[ebp-w6],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL23_32 0
-    Interp3 [ecx+6],[w5],[w8]
+    Interp3 [ecx+6],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL23_50 0
-    Interp5 [ecx+6],[w6],[w5]
+    Interp5 [ecx+6],dword[ebp-w6],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL23_60 0
-    Interp6 [ecx+6],[w5],[w6],[w8]
+    Interp6 [ecx+6],dword[ebp-w5],dword[ebp-w6],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL23_61 0
-    Interp6 [ecx+6],[w5],[w6],[w9]
+    Interp6 [ecx+6],dword[ebp-w5],dword[ebp-w6],dword[ebp-w9]
 %endmacro
 
 %macro PIXEL23_81 0
-    Interp8 [ecx+6],[w5],[w6]
+    Interp8 [ecx+6],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL23_83 0
-    Interp8 [ecx+6],[w6],[w8]
+    Interp8 [ecx+6],dword[ebp-w6],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL30_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [ecx+ebx],ax
 %endmacro
 
 %macro PIXEL30_11 0
-    Interp1 [ecx+ebx],[w5],[w8]
+    Interp1 [ecx+ebx],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL30_12 0
-    Interp1 [ecx+ebx],[w5],[w4]
+    Interp1 [ecx+ebx],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL30_20 0
-    Interp2 [ecx+ebx],[w5],[w8],[w4]
+    Interp2 [ecx+ebx],dword[ebp-w5],dword[ebp-w8],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL30_50 0
-    Interp5 [ecx+ebx],[w8],[w4]
+    Interp5 [ecx+ebx],dword[ebp-w8],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL30_80 0
-    Interp8 [ecx+ebx],[w5],[w7]
+    Interp8 [ecx+ebx],dword[ebp-w5],dword[ebp-w7]
 %endmacro
 
 %macro PIXEL30_81 0
-    Interp8 [ecx+ebx],[w5],[w8]
+    Interp8 [ecx+ebx],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL30_82 0
-    Interp8 [ecx+ebx],[w5],[w4]
+    Interp8 [ecx+ebx],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL31_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [ecx+ebx+2],ax
 %endmacro
 
 %macro PIXEL31_10 0
-    Interp1 [ecx+ebx+2],[w5],[w7]
+    Interp1 [ecx+ebx+2],dword[ebp-w5],dword[ebp-w7]
 %endmacro
 
 %macro PIXEL31_11 0
-    Interp1 [ecx+ebx+2],[w5],[w8]
+    Interp1 [ecx+ebx+2],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL31_13 0
-    Interp1 [ecx+ebx+2],[w8],[w5]
+    Interp1 [ecx+ebx+2],dword[ebp-w8],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL31_21 0
-    Interp2 [ecx+ebx+2],[w8],[w5],[w4]
+    Interp2 [ecx+ebx+2],dword[ebp-w8],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL31_32 0
-    Interp3 [ecx+ebx+2],[w5],[w4]
+    Interp3 [ecx+ebx+2],dword[ebp-w5],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL31_50 0
-    Interp5 [ecx+ebx+2],[w8],[w5]
+    Interp5 [ecx+ebx+2],dword[ebp-w8],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL31_60 0
-    Interp6 [ecx+ebx+2],[w5],[w8],[w4]
+    Interp6 [ecx+ebx+2],dword[ebp-w5],dword[ebp-w8],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL31_61 0
-    Interp6 [ecx+ebx+2],[w5],[w8],[w7]
+    Interp6 [ecx+ebx+2],dword[ebp-w5],dword[ebp-w8],dword[ebp-w7]
 %endmacro
 
 %macro PIXEL31_81 0
-    Interp8 [ecx+ebx+2],[w5],[w8]
+    Interp8 [ecx+ebx+2],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL31_83 0
-    Interp8 [ecx+ebx+2],[w8],[w4]
+    Interp8 [ecx+ebx+2],dword[ebp-w8],dword[ebp-w4]
 %endmacro
 
 %macro PIXEL32_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [ecx+ebx+4],ax
 %endmacro
 
 %macro PIXEL32_10 0
-    Interp1 [ecx+ebx+4],[w5],[w9]
+    Interp1 [ecx+ebx+4],dword[ebp-w5],dword[ebp-w9]
 %endmacro
 
 %macro PIXEL32_12 0
-    Interp1 [ecx+ebx+4],[w5],[w8]
+    Interp1 [ecx+ebx+4],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL32_14 0
-    Interp1 [ecx+ebx+4],[w8],[w5]
+    Interp1 [ecx+ebx+4],dword[ebp-w8],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL32_21 0
-    Interp2 [ecx+ebx+4],[w8],[w5],[w6]
+    Interp2 [ecx+ebx+4],dword[ebp-w8],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL32_31 0
-    Interp3 [ecx+ebx+4],[w5],[w6]
+    Interp3 [ecx+ebx+4],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL32_50 0
-    Interp5 [ecx+ebx+4],[w8],[w5]
+    Interp5 [ecx+ebx+4],dword[ebp-w8],dword[ebp-w5]
 %endmacro
 
 %macro PIXEL32_60 0
-    Interp6 [ecx+ebx+4],[w5],[w8],[w6]
+    Interp6 [ecx+ebx+4],dword[ebp-w5],dword[ebp-w8],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL32_61 0
-    Interp6 [ecx+ebx+4],[w5],[w8],[w9]
+    Interp6 [ecx+ebx+4],dword[ebp-w5],dword[ebp-w8],dword[ebp-w9]
 %endmacro
 
 %macro PIXEL32_82 0
-    Interp8 [ecx+ebx+4],[w5],[w8]
+    Interp8 [ecx+ebx+4],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL32_83 0
-    Interp8 [ecx+ebx+4],[w8],[w6]
+    Interp8 [ecx+ebx+4],dword[ebp-w8],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL33_0 0
-    mov eax,[w5]
+    mov eax,dword[ebp-w5]
     mov [ecx+ebx+6],ax
 %endmacro
 
 %macro PIXEL33_11 0
-    Interp1 [ecx+ebx+6],[w5],[w6]
+    Interp1 [ecx+ebx+6],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL33_12 0
-    Interp1 [ecx+ebx+6],[w5],[w8]
+    Interp1 [ecx+ebx+6],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 %macro PIXEL33_20 0
-    Interp2 [ecx+ebx+6],[w5],[w8],[w6]
+    Interp2 [ecx+ebx+6],dword[ebp-w5],dword[ebp-w8],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL33_50 0
-    Interp5 [ecx+ebx+6],[w8],[w6]
+    Interp5 [ecx+ebx+6],dword[ebp-w8],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL33_80 0
-    Interp8 [ecx+ebx+6],[w5],[w9]
+    Interp8 [ecx+ebx+6],dword[ebp-w5],dword[ebp-w9]
 %endmacro
 
 %macro PIXEL33_81 0
-    Interp8 [ecx+ebx+6],[w5],[w6]
+    Interp8 [ecx+ebx+6],dword[ebp-w5],dword[ebp-w6]
 %endmacro
 
 %macro PIXEL33_82 0
-    Interp8 [ecx+ebx+6],[w5],[w8]
+    Interp8 [ecx+ebx+6],dword[ebp-w5],dword[ebp-w8]
 %endmacro
 
 inbuffer     equ 8
@@ -866,95 +852,111 @@ Yres         equ 20
 pitch        equ 24
 offset       equ 28
 
+linesleft equ 4
+xcounter  equ 8
+cross     equ 12
+nextline  equ 16
+prevline  equ 20
+w1        equ 24
+w2        equ 28
+w3        equ 32
+w4        equ 36
+w5        equ 40
+w6        equ 44
+w7        equ 48
+w8        equ 52
+w9        equ 56
+localsize equ 56
+
 NEWSYM hq4x_16
     push ebp
     mov ebp,esp
+    sub esp, localsize
     pushad
 
     mov     esi,[ebp+inbuffer]
     mov     edi,[ebp+outbuffer]
     mov     edx,[ebp+Yres]
-    mov     [linesleft],edx
+    mov     [ebp-linesleft],edx
     mov     ebx,[ebp+Xres]
     shl     ebx,1
-    mov     dword[prevline],0
-    mov     eax, ebx
-    add     eax, [ebp+offset]
-    mov     dword[nextline],eax
-
+    mov     dword[ebp-prevline],0
+    mov     eax, [ebp+offset]
+    add     eax, ebx
+    mov     dword[ebp-nextline],eax
 .loopy
     mov     ecx,[ebp+Xres]
     sub     ecx,2                 ; x={Xres-2, Xres-1} are special cases.
-    mov     dword[xcounter],ecx
+    mov     dword[ebp-xcounter],ecx
     ; x=0 - special case
-    mov     ebx,[prevline]
+    mov     ebx,[ebp-prevline]
     movq    mm5,[esi+ebx]
     movq    mm6,[esi]
-    mov     ebx,[nextline]
+    mov     ebx,[ebp-nextline]
     movq    mm7,[esi+ebx]
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w1],edx
-    mov     [w2],edx
+    mov     [ebp-w1],edx
+    mov     [ebp-w2],edx
     shr     eax,16
-    mov     [w3],eax
+    mov     [ebp-w3],eax
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w4],edx
-    mov     [w5],edx
+    mov     [ebp-w4],edx
+    mov     [ebp-w5],edx
     shr     eax,16
-    mov     [w6],eax
+    mov     [ebp-w6],eax
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w7],edx
-    mov     [w8],edx
+    mov     [ebp-w7],edx
+    mov     [ebp-w8],edx
     shr     eax,16
-    mov     [w9],eax
+    mov     [ebp-w9],eax
     jmp     .flags
 .loopx
-    mov     ebx,[prevline]
+    mov     ebx,[ebp-prevline]
     movq    mm5,[esi+ebx-2]
     movq    mm6,[esi-2]
-    mov     ebx,[nextline]
+    mov     ebx,[ebp-nextline]
     movq    mm7,[esi+ebx-2]
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w1],edx
+    mov     [ebp-w1],edx
     shr     eax,16
-    mov     [w2],eax
+    mov     [ebp-w2],eax
     psrlq   mm5,32
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w3],edx
+    mov     [ebp-w3],edx
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w4],edx
+    mov     [ebp-w4],edx
     shr     eax,16
-    mov     [w5],eax
+    mov     [ebp-w5],eax
     psrlq   mm6,32
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w6],edx
+    mov     [ebp-w6],edx
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w7],edx
+    mov     [ebp-w7],edx
     shr     eax,16
-    mov     [w8],eax
+    mov     [ebp-w8],eax
     psrlq   mm7,32
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w9],edx
+    mov     [ebp-w9],edx
 .flags
     mov     ebx,RGBtoYUV
-    mov     eax,[w5]
+    mov     eax,[ebp-w5]
     xor     ecx,ecx
     movd    mm5,[ebx+eax*4]
-    mov     dword[cross],0
+    mov     dword[ebp-cross],0
 
-    mov     edx,[w2]
+    mov     edx,[ebp-w2]
     cmp     eax,edx
     je      .noflag2
-    or      dword[cross],1
+    or      dword[ebp-cross],1
     movq    mm1,mm5
     movd    mm2,[ebx+edx*4]
     psubusb mm1,mm2
@@ -966,10 +968,10 @@ NEWSYM hq4x_16
     jz      .noflag2
     or      ecx,2
 .noflag2
-    mov     edx,[w4]
+    mov     edx,[ebp-w4]
     cmp     eax,edx
     je      .noflag4
-    or      dword[cross],2
+    or      dword[ebp-cross],2
     movq    mm1,mm5
     movd    mm2,[ebx+edx*4]
     psubusb mm1,mm2
@@ -981,10 +983,10 @@ NEWSYM hq4x_16
     jz      .noflag4
     or      ecx,8
 .noflag4
-    mov     edx,[w6]
+    mov     edx,[ebp-w6]
     cmp     eax,edx
     je      .noflag6
-    or      dword[cross],4
+    or      dword[ebp-cross],4
     movq    mm1,mm5
     movd    mm2,[ebx+edx*4]
     psubusb mm1,mm2
@@ -996,10 +998,10 @@ NEWSYM hq4x_16
     jz      .noflag6
     or      ecx,16
 .noflag6
-    mov     edx,[w8]
+    mov     edx,[ebp-w8]
     cmp     eax,edx
     je      .noflag8
-    or      dword[cross],8
+    or      dword[ebp-cross],8
     movq    mm1,mm5
     movd    mm2,[ebx+edx*4]
     psubusb mm1,mm2
@@ -1011,7 +1013,7 @@ NEWSYM hq4x_16
     jz      .noflag8
     or      ecx,64
 .noflag8
-    cmp     dword[cross],0
+    cmp     dword[ebp-cross],0
     jnz     .testflag1
     mov     ebx,[ebp+pitch]
     mov     edx,eax
@@ -1028,7 +1030,7 @@ NEWSYM hq4x_16
     mov     [ecx+ebx+4],eax
     jmp     .loopx_end
 .testflag1
-    mov     edx,[w1]
+    mov     edx,[ebp-w1]
     cmp     eax,edx
     je      .noflag1
     movq    mm1,mm5
@@ -1042,7 +1044,7 @@ NEWSYM hq4x_16
     jz      .noflag1
     or      ecx,1
 .noflag1
-    mov     edx,[w3]
+    mov     edx,[ebp-w3]
     cmp     eax,edx
     je      .noflag3
     movq    mm1,mm5
@@ -1056,7 +1058,7 @@ NEWSYM hq4x_16
     jz      .noflag3
     or      ecx,4
 .noflag3
-    mov     edx,[w7]
+    mov     edx,[ebp-w7]
     cmp     eax,edx
     je      .noflag7
     movq    mm1,mm5
@@ -1070,7 +1072,7 @@ NEWSYM hq4x_16
     jz      .noflag7
     or      ecx,32
 .noflag7
-    mov     edx,[w9]
+    mov     edx,[ebp-w9]
     cmp     eax,edx
     je      .noflag9
     movq    mm1,mm5
@@ -1390,7 +1392,7 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_50,PIXEL03_50,PIXEL12_0,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_50,PIXEL03_50,PIXEL12_0,PIXEL13_50
     PIXEL10_61
     PIXEL11_30
     PIXEL20_60
@@ -1415,7 +1417,7 @@ NEWSYM hq4x_16
     PIXEL13_10
     PIXEL20_61
     PIXEL21_30
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     PIXEL30_80
     PIXEL31_10
     jmp .loopx_end
@@ -1430,7 +1432,7 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_70
     PIXEL13_60
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_50,PIXEL21_0,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_50,PIXEL21_0,PIXEL30_50,PIXEL31_50
     PIXEL22_30
     PIXEL23_61
     PIXEL32_10
@@ -1439,7 +1441,7 @@ NEWSYM hq4x_16
 ..@flag10
 ..@flag138
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_50,PIXEL01_50,PIXEL10_50,PIXEL11_0
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_50,PIXEL01_50,PIXEL10_50,PIXEL11_0
     PIXEL02_10
     PIXEL03_80
     PIXEL12_30
@@ -1580,7 +1582,7 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL10_61
     PIXEL11_30
     PIXEL12_0
@@ -1607,7 +1609,7 @@ NEWSYM hq4x_16
     PIXEL20_61
     PIXEL21_30
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     PIXEL30_80
     PIXEL31_10
     jmp .loopx_end
@@ -1622,7 +1624,7 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_70
     PIXEL13_60
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_30
     PIXEL23_61
@@ -1632,7 +1634,7 @@ NEWSYM hq4x_16
 ..@flag11
 ..@flag139
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
     PIXEL02_10
     PIXEL03_80
     PIXEL11_0
@@ -1650,7 +1652,7 @@ NEWSYM hq4x_16
 ..@flag19
 ..@flag51
     AUXADDRESS
-    DiffOrNot w2,w6,PIXEL00_81,PIXEL01_31,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL00_12,PIXEL01_14,PIXEL02_83,PIXEL03_50,PIXEL12_70,PIXEL13_21
+    DiffOrNot ebp-w2,ebp-w6,PIXEL00_81,PIXEL01_31,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL00_12,PIXEL01_14,PIXEL02_83,PIXEL03_50,PIXEL12_70,PIXEL13_21
     PIXEL10_81
     PIXEL11_31
     PIXEL20_60
@@ -1667,7 +1669,7 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL23_32,PIXEL33_82,PIXEL02_21,PIXEL03_50,PIXEL12_70,PIXEL13_83,PIXEL23_13,PIXEL33_11
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL23_32,PIXEL33_82,PIXEL02_21,PIXEL03_50,PIXEL12_70,PIXEL13_83,PIXEL23_13,PIXEL33_11
     PIXEL10_61
     PIXEL11_30
     PIXEL20_60
@@ -1683,7 +1685,7 @@ NEWSYM hq4x_16
     PIXEL00_20
     PIXEL01_60
     PIXEL02_81
-    DiffOrNot w6,w8,PIXEL03_81,PIXEL13_31,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL03_12,PIXEL13_14,PIXEL22_70,PIXEL23_83,PIXEL32_21,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL03_81,PIXEL13_31,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL03_12,PIXEL13_14,PIXEL22_70,PIXEL23_83,PIXEL32_21,PIXEL33_50
     PIXEL10_60
     PIXEL11_70
     PIXEL12_31
@@ -1705,7 +1707,7 @@ NEWSYM hq4x_16
     PIXEL13_10
     PIXEL20_82
     PIXEL21_32
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL30_82,PIXEL31_32,PIXEL32_10,PIXEL33_80,PIXEL22_70,PIXEL23_21,PIXEL30_11,PIXEL31_13,PIXEL32_83,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL30_82,PIXEL31_32,PIXEL32_10,PIXEL33_80,PIXEL22_70,PIXEL23_21,PIXEL30_11,PIXEL31_13,PIXEL32_83,PIXEL33_50
     jmp .loopx_end
 ..@flag200
 ..@flag204
@@ -1718,14 +1720,14 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_70
     PIXEL13_60
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL32_31,PIXEL33_81,PIXEL20_21,PIXEL21_70,PIXEL30_50,PIXEL31_83,PIXEL32_14,PIXEL33_12
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL32_31,PIXEL33_81,PIXEL20_21,PIXEL21_70,PIXEL30_50,PIXEL31_83,PIXEL32_14,PIXEL33_12
     PIXEL22_31
     PIXEL23_81
     jmp .loopx_end
 ..@flag73
 ..@flag77
     AUXADDRESS
-    DiffOrNot w8,w4,PIXEL00_82,PIXEL10_32,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL00_11,PIXEL10_13,PIXEL20_83,PIXEL21_70,PIXEL30_50,PIXEL31_21
+    DiffOrNot ebp-w8,ebp-w4,PIXEL00_82,PIXEL10_32,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL00_11,PIXEL10_13,PIXEL20_83,PIXEL21_70,PIXEL30_50,PIXEL31_21
     PIXEL01_82
     PIXEL02_60
     PIXEL03_20
@@ -1740,7 +1742,7 @@ NEWSYM hq4x_16
 ..@flag42
 ..@flag170
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL20_31,PIXEL30_81,PIXEL00_50,PIXEL01_21,PIXEL10_83,PIXEL11_70,PIXEL20_14,PIXEL30_12
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL20_31,PIXEL30_81,PIXEL00_50,PIXEL01_21,PIXEL10_83,PIXEL11_70,PIXEL20_14,PIXEL30_12
     PIXEL02_10
     PIXEL03_80
     PIXEL12_30
@@ -1755,7 +1757,7 @@ NEWSYM hq4x_16
 ..@flag14
 ..@flag142
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL02_32,PIXEL03_82,PIXEL10_10,PIXEL11_30,PIXEL00_50,PIXEL01_83,PIXEL02_13,PIXEL03_11,PIXEL10_21,PIXEL11_70
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL02_32,PIXEL03_82,PIXEL10_10,PIXEL11_30,PIXEL00_50,PIXEL01_83,PIXEL02_13,PIXEL03_11,PIXEL10_21,PIXEL11_70
     PIXEL12_32
     PIXEL13_82
     PIXEL20_10
@@ -1922,8 +1924,8 @@ NEWSYM hq4x_16
 ..@flag26
 ..@flag31
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL11_0
     PIXEL12_0
     PIXEL20_10
@@ -1940,14 +1942,14 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL10_61
     PIXEL11_30
     PIXEL12_0
     PIXEL20_61
     PIXEL21_30
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     PIXEL30_80
     PIXEL31_10
     jmp .loopx_end
@@ -1962,21 +1964,21 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_30
     PIXEL13_10
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     jmp .loopx_end
 ..@flag74
 ..@flag107
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
     PIXEL02_10
     PIXEL03_80
     PIXEL11_0
     PIXEL12_30
     PIXEL13_61
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_30
     PIXEL23_61
@@ -1985,7 +1987,7 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag27
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
     PIXEL02_10
     PIXEL03_80
     PIXEL11_0
@@ -2004,7 +2006,7 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL10_61
     PIXEL11_30
     PIXEL12_0
@@ -2030,7 +2032,7 @@ NEWSYM hq4x_16
     PIXEL20_10
     PIXEL21_30
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     PIXEL30_80
     PIXEL31_10
     jmp .loopx_end
@@ -2044,7 +2046,7 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_30
     PIXEL13_61
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_30
     PIXEL23_61
@@ -2055,7 +2057,7 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL10_10
     PIXEL11_30
     PIXEL12_0
@@ -2081,7 +2083,7 @@ NEWSYM hq4x_16
     PIXEL20_61
     PIXEL21_30
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     PIXEL30_80
     PIXEL31_10
     jmp .loopx_end
@@ -2095,7 +2097,7 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_30
     PIXEL13_10
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_30
     PIXEL23_10
@@ -2104,7 +2106,7 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag75
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
     PIXEL02_10
     PIXEL03_80
     PIXEL11_0
@@ -2349,8 +2351,8 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag58
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
     PIXEL20_31
     PIXEL21_31
     PIXEL22_30
@@ -2364,12 +2366,12 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_81
     PIXEL01_31
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
     PIXEL10_81
     PIXEL11_31
     PIXEL20_61
     PIXEL21_30
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
     PIXEL30_80
     PIXEL31_10
     jmp .loopx_end
@@ -2383,17 +2385,17 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_31
     PIXEL13_31
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
     jmp .loopx_end
 ..@flag202
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
     PIXEL02_10
     PIXEL03_80
     PIXEL12_30
     PIXEL13_61
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
     PIXEL22_31
     PIXEL23_81
     PIXEL32_31
@@ -2401,12 +2403,12 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag78
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
     PIXEL02_32
     PIXEL03_82
     PIXEL12_32
     PIXEL13_82
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
     PIXEL22_30
     PIXEL23_61
     PIXEL32_10
@@ -2414,8 +2416,8 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag154
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
     PIXEL20_10
     PIXEL21_30
     PIXEL22_32
@@ -2429,12 +2431,12 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
     PIXEL10_61
     PIXEL11_30
     PIXEL20_82
     PIXEL21_32
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
     PIXEL30_82
     PIXEL31_32
     jmp .loopx_end
@@ -2448,20 +2450,20 @@ NEWSYM hq4x_16
     PIXEL11_32
     PIXEL12_30
     PIXEL13_10
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
     jmp .loopx_end
 ..@flag90
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
     jmp .loopx_end
 ..@flag55
 ..@flag23
     AUXADDRESS
-    DiffOrNot w2,w6,PIXEL00_81,PIXEL01_31,PIXEL02_0,PIXEL03_0,PIXEL12_0,PIXEL13_0,PIXEL00_12,PIXEL01_14,PIXEL02_83,PIXEL03_50,PIXEL12_70,PIXEL13_21
+    DiffOrNot ebp-w2,ebp-w6,PIXEL00_81,PIXEL01_31,PIXEL02_0,PIXEL03_0,PIXEL12_0,PIXEL13_0,PIXEL00_12,PIXEL01_14,PIXEL02_83,PIXEL03_50,PIXEL12_70,PIXEL13_21
     PIXEL10_81
     PIXEL11_31
     PIXEL20_60
@@ -2478,7 +2480,7 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL12_0,PIXEL13_0,PIXEL23_32,PIXEL33_82,PIXEL02_21,PIXEL03_50,PIXEL12_70,PIXEL13_83,PIXEL23_13,PIXEL33_11
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL12_0,PIXEL13_0,PIXEL23_32,PIXEL33_82,PIXEL02_21,PIXEL03_50,PIXEL12_70,PIXEL13_83,PIXEL23_13,PIXEL33_11
     PIXEL10_61
     PIXEL11_30
     PIXEL20_60
@@ -2494,7 +2496,7 @@ NEWSYM hq4x_16
     PIXEL00_20
     PIXEL01_60
     PIXEL02_81
-    DiffOrNot w6,w8,PIXEL03_81,PIXEL13_31,PIXEL22_0,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL03_12,PIXEL13_14,PIXEL22_70,PIXEL23_83,PIXEL32_21,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL03_81,PIXEL13_31,PIXEL22_0,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL03_12,PIXEL13_14,PIXEL22_70,PIXEL23_83,PIXEL32_21,PIXEL33_50
     PIXEL10_60
     PIXEL11_70
     PIXEL12_31
@@ -2516,7 +2518,7 @@ NEWSYM hq4x_16
     PIXEL13_10
     PIXEL20_82
     PIXEL21_32
-    DiffOrNot w6,w8,PIXEL22_0,PIXEL23_0,PIXEL30_82,PIXEL31_32,PIXEL32_0,PIXEL33_0,PIXEL22_70,PIXEL23_21,PIXEL30_11,PIXEL31_13,PIXEL32_83,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_0,PIXEL23_0,PIXEL30_82,PIXEL31_32,PIXEL32_0,PIXEL33_0,PIXEL22_70,PIXEL23_21,PIXEL30_11,PIXEL31_13,PIXEL32_83,PIXEL33_50
     jmp .loopx_end
 ..@flag236
 ..@flag232
@@ -2529,14 +2531,14 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_70
     PIXEL13_60
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL21_0,PIXEL30_0,PIXEL31_0,PIXEL32_31,PIXEL33_81,PIXEL20_21,PIXEL21_70,PIXEL30_50,PIXEL31_83,PIXEL32_14,PIXEL33_12
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL21_0,PIXEL30_0,PIXEL31_0,PIXEL32_31,PIXEL33_81,PIXEL20_21,PIXEL21_70,PIXEL30_50,PIXEL31_83,PIXEL32_14,PIXEL33_12
     PIXEL22_31
     PIXEL23_81
     jmp .loopx_end
 ..@flag109
 ..@flag105
     AUXADDRESS
-    DiffOrNot w8,w4,PIXEL00_82,PIXEL10_32,PIXEL20_0,PIXEL21_0,PIXEL30_0,PIXEL31_0,PIXEL00_11,PIXEL10_13,PIXEL20_83,PIXEL21_70,PIXEL30_50,PIXEL31_21
+    DiffOrNot ebp-w8,ebp-w4,PIXEL00_82,PIXEL10_32,PIXEL20_0,PIXEL21_0,PIXEL30_0,PIXEL31_0,PIXEL00_11,PIXEL10_13,PIXEL20_83,PIXEL21_70,PIXEL30_50,PIXEL31_21
     PIXEL01_82
     PIXEL02_60
     PIXEL03_20
@@ -2551,7 +2553,7 @@ NEWSYM hq4x_16
 ..@flag171
 ..@flag43
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL11_0,PIXEL20_31,PIXEL30_81,PIXEL00_50,PIXEL01_21,PIXEL10_83,PIXEL11_70,PIXEL20_14,PIXEL30_12
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL11_0,PIXEL20_31,PIXEL30_81,PIXEL00_50,PIXEL01_21,PIXEL10_83,PIXEL11_70,PIXEL20_14,PIXEL30_12
     PIXEL02_10
     PIXEL03_80
     PIXEL12_30
@@ -2566,7 +2568,7 @@ NEWSYM hq4x_16
 ..@flag143
 ..@flag15
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL02_32,PIXEL03_82,PIXEL10_0,PIXEL11_0,PIXEL00_50,PIXEL01_83,PIXEL02_13,PIXEL03_11,PIXEL10_21,PIXEL11_70
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL02_32,PIXEL03_82,PIXEL10_0,PIXEL11_0,PIXEL00_50,PIXEL01_83,PIXEL02_13,PIXEL03_11,PIXEL10_21,PIXEL11_70
     PIXEL12_32
     PIXEL13_82
     PIXEL20_10
@@ -2588,7 +2590,7 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_31
     PIXEL13_31
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_30
     PIXEL23_10
@@ -2597,7 +2599,7 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag203
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
     PIXEL02_10
     PIXEL03_80
     PIXEL11_0
@@ -2616,7 +2618,7 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL10_10
     PIXEL11_30
     PIXEL12_0
@@ -2642,7 +2644,7 @@ NEWSYM hq4x_16
     PIXEL20_61
     PIXEL21_30
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     PIXEL30_80
     PIXEL31_10
     jmp .loopx_end
@@ -2650,7 +2652,7 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL10_61
     PIXEL11_30
     PIXEL12_0
@@ -2676,7 +2678,7 @@ NEWSYM hq4x_16
     PIXEL20_10
     PIXEL21_30
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     PIXEL30_80
     PIXEL31_10
     jmp .loopx_end
@@ -2690,7 +2692,7 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_32
     PIXEL13_82
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_30
     PIXEL23_61
@@ -2699,7 +2701,7 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag155
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
     PIXEL02_10
     PIXEL03_80
     PIXEL11_0
@@ -2876,14 +2878,14 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_31
     PIXEL13_31
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     jmp .loopx_end
 ..@flag158
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL12_0
     PIXEL20_10
     PIXEL21_30
@@ -2896,12 +2898,12 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag234
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
     PIXEL02_10
     PIXEL03_80
     PIXEL12_30
     PIXEL13_61
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_31
     PIXEL23_81
@@ -2912,20 +2914,20 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
     PIXEL10_61
     PIXEL11_30
     PIXEL20_82
     PIXEL21_32
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     PIXEL30_82
     PIXEL31_32
     jmp .loopx_end
 ..@flag59
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
     PIXEL11_0
     PIXEL20_31
     PIXEL21_31
@@ -2946,33 +2948,33 @@ NEWSYM hq4x_16
     PIXEL11_32
     PIXEL12_30
     PIXEL13_10
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
     jmp .loopx_end
 ..@flag87
     AUXADDRESS
     PIXEL00_81
     PIXEL01_31
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL10_81
     PIXEL11_31
     PIXEL12_0
     PIXEL20_61
     PIXEL21_30
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
     PIXEL30_80
     PIXEL31_10
     jmp .loopx_end
 ..@flag79
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
     PIXEL02_32
     PIXEL03_82
     PIXEL11_0
     PIXEL12_32
     PIXEL13_82
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
     PIXEL22_30
     PIXEL23_61
     PIXEL32_10
@@ -2980,35 +2982,35 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag122
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
     jmp .loopx_end
 ..@flag94
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL12_0
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
     jmp .loopx_end
 ..@flag218
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     jmp .loopx_end
 ..@flag91
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
     PIXEL11_0
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
     jmp .loopx_end
 ..@flag229
     AUXADDRESS
@@ -3088,8 +3090,8 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag186
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
     PIXEL20_31
     PIXEL21_31
     PIXEL22_32
@@ -3103,12 +3105,12 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_81
     PIXEL01_31
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
     PIXEL10_81
     PIXEL11_31
     PIXEL20_82
     PIXEL21_32
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
     PIXEL30_82
     PIXEL31_32
     jmp .loopx_end
@@ -3122,17 +3124,17 @@ NEWSYM hq4x_16
     PIXEL11_32
     PIXEL12_31
     PIXEL13_31
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
     jmp .loopx_end
 ..@flag206
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
     PIXEL02_32
     PIXEL03_82
     PIXEL12_32
     PIXEL13_82
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
     PIXEL22_31
     PIXEL23_81
     PIXEL32_31
@@ -3149,7 +3151,7 @@ NEWSYM hq4x_16
     PIXEL11_32
     PIXEL12_70
     PIXEL13_60
-    DiffOrNot w8,w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_10,PIXEL21_30,PIXEL30_80,PIXEL31_10,PIXEL20_12,PIXEL21_0,PIXEL30_20,PIXEL31_11
     PIXEL22_31
     PIXEL23_81
     PIXEL32_31
@@ -3158,7 +3160,7 @@ NEWSYM hq4x_16
 ..@flag174
 ..@flag46
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_80,PIXEL01_10,PIXEL10_10,PIXEL11_30,PIXEL00_20,PIXEL01_12,PIXEL10_11,PIXEL11_0
     PIXEL02_32
     PIXEL03_82
     PIXEL12_32
@@ -3177,7 +3179,7 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_81
     PIXEL01_31
-    DiffOrNot w2,w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_10,PIXEL03_80,PIXEL12_30,PIXEL13_10,PIXEL02_11,PIXEL03_20,PIXEL12_0,PIXEL13_12
     PIXEL10_81
     PIXEL11_31
     PIXEL20_60
@@ -3202,7 +3204,7 @@ NEWSYM hq4x_16
     PIXEL13_31
     PIXEL20_82
     PIXEL21_32
-    DiffOrNot w6,w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_30,PIXEL23_10,PIXEL32_10,PIXEL33_80,PIXEL22_0,PIXEL23_11,PIXEL32_12,PIXEL33_20
     PIXEL30_82
     PIXEL31_32
     jmp .loopx_end
@@ -3248,11 +3250,11 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL10_10
     PIXEL11_30
     PIXEL12_0
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_30
     PIXEL23_10
@@ -3261,7 +3263,7 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag219
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
     PIXEL02_10
     PIXEL03_80
     PIXEL11_0
@@ -3270,13 +3272,13 @@ NEWSYM hq4x_16
     PIXEL20_10
     PIXEL21_30
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     PIXEL30_80
     PIXEL31_10
     jmp .loopx_end
 ..@flag125
     AUXADDRESS
-    DiffOrNot w8,w4,PIXEL00_82,PIXEL10_32,PIXEL20_0,PIXEL21_0,PIXEL30_0,PIXEL31_0,PIXEL00_11,PIXEL10_13,PIXEL20_83,PIXEL21_70,PIXEL30_50,PIXEL31_21
+    DiffOrNot ebp-w8,ebp-w4,PIXEL00_82,PIXEL10_32,PIXEL20_0,PIXEL21_0,PIXEL30_0,PIXEL31_0,PIXEL00_11,PIXEL10_13,PIXEL20_83,PIXEL21_70,PIXEL30_50,PIXEL31_21
     PIXEL01_82
     PIXEL02_81
     PIXEL03_81
@@ -3293,7 +3295,7 @@ NEWSYM hq4x_16
     PIXEL00_82
     PIXEL01_82
     PIXEL02_81
-    DiffOrNot w6,w8,PIXEL03_81,PIXEL13_31,PIXEL22_0,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL03_12,PIXEL13_14,PIXEL22_70,PIXEL23_83,PIXEL32_21,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL03_81,PIXEL13_31,PIXEL22_0,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL03_12,PIXEL13_14,PIXEL22_70,PIXEL23_83,PIXEL32_21,PIXEL33_50
     PIXEL10_32
     PIXEL11_32
     PIXEL12_31
@@ -3304,7 +3306,7 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag207
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL02_32,PIXEL03_82,PIXEL10_0,PIXEL11_0,PIXEL00_50,PIXEL01_83,PIXEL02_13,PIXEL03_11,PIXEL10_21,PIXEL11_70
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL02_32,PIXEL03_82,PIXEL10_0,PIXEL11_0,PIXEL00_50,PIXEL01_83,PIXEL02_13,PIXEL03_11,PIXEL10_21,PIXEL11_70
     PIXEL12_32
     PIXEL13_82
     PIXEL20_10
@@ -3326,7 +3328,7 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_32
     PIXEL13_82
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL21_0,PIXEL30_0,PIXEL31_0,PIXEL32_31,PIXEL33_81,PIXEL20_21,PIXEL21_70,PIXEL30_50,PIXEL31_83,PIXEL32_14,PIXEL33_12
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL21_0,PIXEL30_0,PIXEL31_0,PIXEL32_31,PIXEL33_81,PIXEL20_21,PIXEL21_70,PIXEL30_50,PIXEL31_83,PIXEL32_14,PIXEL33_12
     PIXEL22_31
     PIXEL23_81
     jmp .loopx_end
@@ -3334,7 +3336,7 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL12_0,PIXEL13_0,PIXEL23_32,PIXEL33_82,PIXEL02_21,PIXEL03_50,PIXEL12_70,PIXEL13_83,PIXEL23_13,PIXEL33_11
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL12_0,PIXEL13_0,PIXEL23_32,PIXEL33_82,PIXEL02_21,PIXEL03_50,PIXEL12_70,PIXEL13_83,PIXEL23_13,PIXEL33_11
     PIXEL10_10
     PIXEL11_30
     PIXEL20_31
@@ -3346,7 +3348,7 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag187
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL11_0,PIXEL20_31,PIXEL30_81,PIXEL00_50,PIXEL01_21,PIXEL10_83,PIXEL11_70,PIXEL20_14,PIXEL30_12
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL11_0,PIXEL20_31,PIXEL30_81,PIXEL00_50,PIXEL01_21,PIXEL10_83,PIXEL11_70,PIXEL20_14,PIXEL30_12
     PIXEL02_10
     PIXEL03_80
     PIXEL12_30
@@ -3370,11 +3372,11 @@ NEWSYM hq4x_16
     PIXEL13_10
     PIXEL20_82
     PIXEL21_32
-    DiffOrNot w6,w8,PIXEL22_0,PIXEL23_0,PIXEL30_82,PIXEL31_32,PIXEL32_0,PIXEL33_0,PIXEL22_70,PIXEL23_21,PIXEL30_11,PIXEL31_13,PIXEL32_83,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL22_0,PIXEL23_0,PIXEL30_82,PIXEL31_32,PIXEL32_0,PIXEL33_0,PIXEL22_70,PIXEL23_21,PIXEL30_11,PIXEL31_13,PIXEL32_83,PIXEL33_50
     jmp .loopx_end
 ..@flag119
     AUXADDRESS
-    DiffOrNot w2,w6,PIXEL00_81,PIXEL01_31,PIXEL02_0,PIXEL03_0,PIXEL12_0,PIXEL13_0,PIXEL00_12,PIXEL01_14,PIXEL02_83,PIXEL03_50,PIXEL12_70,PIXEL13_21
+    DiffOrNot ebp-w2,ebp-w6,PIXEL00_81,PIXEL01_31,PIXEL02_0,PIXEL03_0,PIXEL12_0,PIXEL13_0,PIXEL00_12,PIXEL01_14,PIXEL02_83,PIXEL03_50,PIXEL12_70,PIXEL13_21
     PIXEL10_81
     PIXEL11_31
     PIXEL20_82
@@ -3401,7 +3403,7 @@ NEWSYM hq4x_16
     PIXEL21_0
     PIXEL22_31
     PIXEL23_81
-    DiffOrNot w8,w4,PIXEL30_0,PIXEL30_20
+    DiffOrNot ebp-w8,ebp-w4,PIXEL30_0,PIXEL30_20
     PIXEL31_0
     PIXEL32_31
     PIXEL33_81
@@ -3409,7 +3411,7 @@ NEWSYM hq4x_16
 ..@flag175
 ..@flag47
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL00_20
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL00_20
     PIXEL01_0
     PIXEL02_32
     PIXEL03_82
@@ -3432,7 +3434,7 @@ NEWSYM hq4x_16
     PIXEL00_81
     PIXEL01_31
     PIXEL02_0
-    DiffOrNot w2,w6,PIXEL03_0,PIXEL03_20
+    DiffOrNot ebp-w2,ebp-w6,PIXEL03_0,PIXEL03_20
     PIXEL10_81
     PIXEL11_31
     PIXEL12_0
@@ -3464,7 +3466,7 @@ NEWSYM hq4x_16
     PIXEL30_82
     PIXEL31_32
     PIXEL32_0
-    DiffOrNot w6,w8,PIXEL33_0,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL33_0,PIXEL33_20
     jmp .loopx_end
 ..@flag250
     AUXADDRESS
@@ -3476,20 +3478,20 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_30
     PIXEL13_10
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     jmp .loopx_end
 ..@flag123
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
     PIXEL02_10
     PIXEL03_80
     PIXEL11_0
     PIXEL12_30
     PIXEL13_10
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_30
     PIXEL23_10
@@ -3498,8 +3500,8 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag95
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL11_0
     PIXEL12_0
     PIXEL20_10
@@ -3515,14 +3517,14 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL10_10
     PIXEL11_30
     PIXEL12_0
     PIXEL20_10
     PIXEL21_30
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     PIXEL30_80
     PIXEL31_10
     jmp .loopx_end
@@ -3536,12 +3538,12 @@ NEWSYM hq4x_16
     PIXEL11_30
     PIXEL12_31
     PIXEL13_31
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_0
     PIXEL23_0
     PIXEL32_0
-    DiffOrNot w6,w8,PIXEL33_0,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL33_0,PIXEL33_20
     jmp .loopx_end
 ..@flag249
     AUXADDRESS
@@ -3556,13 +3558,13 @@ NEWSYM hq4x_16
     PIXEL20_0
     PIXEL21_0
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
-    DiffOrNot w8,w4,PIXEL30_0,PIXEL30_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL30_0,PIXEL30_20
     PIXEL31_0
     jmp .loopx_end
 ..@flag235
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
     PIXEL02_10
     PIXEL03_80
     PIXEL11_0
@@ -3572,14 +3574,14 @@ NEWSYM hq4x_16
     PIXEL21_0
     PIXEL22_31
     PIXEL23_81
-    DiffOrNot w8,w4,PIXEL30_0,PIXEL30_20
+    DiffOrNot ebp-w8,ebp-w4,PIXEL30_0,PIXEL30_20
     PIXEL31_0
     PIXEL32_31
     PIXEL33_81
     jmp .loopx_end
 ..@flag111
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL00_20
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL00_20
     PIXEL01_0
     PIXEL02_32
     PIXEL03_82
@@ -3587,7 +3589,7 @@ NEWSYM hq4x_16
     PIXEL11_0
     PIXEL12_32
     PIXEL13_82
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_30
     PIXEL23_61
@@ -3596,9 +3598,9 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag63
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL00_20
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL00_20
     PIXEL01_0
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL10_0
     PIXEL11_0
     PIXEL12_0
@@ -3613,9 +3615,9 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag159
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
     PIXEL02_0
-    DiffOrNot w2,w6,PIXEL03_0,PIXEL03_20
+    DiffOrNot ebp-w2,ebp-w6,PIXEL03_0,PIXEL03_20
     PIXEL11_0
     PIXEL12_0
     PIXEL13_0
@@ -3633,7 +3635,7 @@ NEWSYM hq4x_16
     PIXEL00_81
     PIXEL01_31
     PIXEL02_0
-    DiffOrNot w2,w6,PIXEL03_0,PIXEL03_20
+    DiffOrNot ebp-w2,ebp-w6,PIXEL03_0,PIXEL03_20
     PIXEL10_81
     PIXEL11_31
     PIXEL12_0
@@ -3641,7 +3643,7 @@ NEWSYM hq4x_16
     PIXEL20_61
     PIXEL21_30
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     PIXEL30_80
     PIXEL31_10
     jmp .loopx_end
@@ -3649,7 +3651,7 @@ NEWSYM hq4x_16
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL10_61
     PIXEL11_30
     PIXEL12_0
@@ -3660,22 +3662,22 @@ NEWSYM hq4x_16
     PIXEL30_82
     PIXEL31_32
     PIXEL32_0
-    DiffOrNot w6,w8,PIXEL33_0,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL33_0,PIXEL33_20
     jmp .loopx_end
 ..@flag254
     AUXADDRESS
     PIXEL00_80
     PIXEL01_10
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL10_10
     PIXEL11_30
     PIXEL12_0
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_0
     PIXEL23_0
     PIXEL32_0
-    DiffOrNot w6,w8,PIXEL33_0,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL33_0,PIXEL33_20
     jmp .loopx_end
 ..@flag253
     AUXADDRESS
@@ -3691,14 +3693,14 @@ NEWSYM hq4x_16
     PIXEL21_0
     PIXEL22_0
     PIXEL23_0
-    DiffOrNot w8,w4,PIXEL30_0,PIXEL30_20
+    DiffOrNot ebp-w8,ebp-w4,PIXEL30_0,PIXEL30_20
     PIXEL31_0
     PIXEL32_0
-    DiffOrNot w6,w8,PIXEL33_0,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL33_0,PIXEL33_20
     jmp .loopx_end
 ..@flag251
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
     PIXEL02_10
     PIXEL03_80
     PIXEL11_0
@@ -3707,13 +3709,13 @@ NEWSYM hq4x_16
     PIXEL20_0
     PIXEL21_0
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
-    DiffOrNot w8,w4,PIXEL30_0,PIXEL30_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL30_0,PIXEL30_20
     PIXEL31_0
     jmp .loopx_end
 ..@flag239
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL00_20
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL00_20
     PIXEL01_0
     PIXEL02_32
     PIXEL03_82
@@ -3725,20 +3727,20 @@ NEWSYM hq4x_16
     PIXEL21_0
     PIXEL22_31
     PIXEL23_81
-    DiffOrNot w8,w4,PIXEL30_0,PIXEL30_20
+    DiffOrNot ebp-w8,ebp-w4,PIXEL30_0,PIXEL30_20
     PIXEL31_0
     PIXEL32_31
     PIXEL33_81
     jmp .loopx_end
 ..@flag127
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL00_20
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL00_20
     PIXEL01_0
-    DiffOrNot w2,w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
+    DiffOrNot ebp-w2,ebp-w6,PIXEL02_0,PIXEL03_0,PIXEL13_0,PIXEL02_50,PIXEL03_50,PIXEL13_50
     PIXEL10_0
     PIXEL11_0
     PIXEL12_0
-    DiffOrNot w8,w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
+    DiffOrNot ebp-w8,ebp-w4,PIXEL20_0,PIXEL30_0,PIXEL31_0,PIXEL20_50,PIXEL30_50,PIXEL31_50
     PIXEL21_0
     PIXEL22_30
     PIXEL23_10
@@ -3747,10 +3749,10 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag191
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL00_20
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL00_20
     PIXEL01_0
     PIXEL02_0
-    DiffOrNot w2,w6,PIXEL03_0,PIXEL03_20
+    DiffOrNot ebp-w2,ebp-w6,PIXEL03_0,PIXEL03_20
     PIXEL10_0
     PIXEL11_0
     PIXEL12_0
@@ -3766,16 +3768,16 @@ NEWSYM hq4x_16
     jmp .loopx_end
 ..@flag223
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL01_0,PIXEL10_0,PIXEL00_50,PIXEL01_50,PIXEL10_50
     PIXEL02_0
-    DiffOrNot w2,w6,PIXEL03_0,PIXEL03_20
+    DiffOrNot ebp-w2,ebp-w6,PIXEL03_0,PIXEL03_20
     PIXEL11_0
     PIXEL12_0
     PIXEL13_0
     PIXEL20_10
     PIXEL21_30
     PIXEL22_0
-    DiffOrNot w6,w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
+    DiffOrNot ebp-w6,ebp-w8,PIXEL23_0,PIXEL32_0,PIXEL33_0,PIXEL23_50,PIXEL32_50,PIXEL33_50
     PIXEL30_80
     PIXEL31_10
     jmp .loopx_end
@@ -3784,7 +3786,7 @@ NEWSYM hq4x_16
     PIXEL00_81
     PIXEL01_31
     PIXEL02_0
-    DiffOrNot w2,w6,PIXEL03_0,PIXEL03_20
+    DiffOrNot ebp-w2,ebp-w6,PIXEL03_0,PIXEL03_20
     PIXEL10_81
     PIXEL11_31
     PIXEL12_0
@@ -3796,14 +3798,14 @@ NEWSYM hq4x_16
     PIXEL30_82
     PIXEL31_32
     PIXEL32_0
-    DiffOrNot w6,w8,PIXEL33_0,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL33_0,PIXEL33_20
     jmp .loopx_end
 ..@flag255
     AUXADDRESS
-    DiffOrNot w4,w2,PIXEL00_0,PIXEL00_20
+    DiffOrNot ebp-w4,ebp-w2,PIXEL00_0,PIXEL00_20
     PIXEL01_0
     PIXEL02_0
-    DiffOrNot w2,w6,PIXEL03_0,PIXEL03_20
+    DiffOrNot ebp-w2,ebp-w6,PIXEL03_0,PIXEL03_20
     PIXEL10_0
     PIXEL11_0
     PIXEL12_0
@@ -3812,84 +3814,84 @@ NEWSYM hq4x_16
     PIXEL21_0
     PIXEL22_0
     PIXEL23_0
-    DiffOrNot w8,w4,PIXEL30_0,PIXEL30_20
+    DiffOrNot ebp-w8,ebp-w4,PIXEL30_0,PIXEL30_20
     PIXEL31_0
     PIXEL32_0
-    DiffOrNot w6,w8,PIXEL33_0,PIXEL33_20
+    DiffOrNot ebp-w6,ebp-w8,PIXEL33_0,PIXEL33_20
     jmp .loopx_end
 
 .loopx_end
     add     esi,2
     add     edi,8
-    dec     dword[xcounter]
+    dec     dword[ebp-xcounter]
     jle     .xres_2
     jmp     .loopx
 .xres_2
     ; x=Xres-2 - special case
     jl      .xres_1
-    mov     ebx,[prevline]
+    mov     ebx,[ebp-prevline]
     movq    mm5,[esi+ebx-4]
     movq    mm6,[esi-4]
-    mov     ebx,[nextline]
+    mov     ebx,[ebp-nextline]
     movq    mm7,[esi+ebx-4]
     psrlq   mm5,16
     psrlq   mm6,16
     psrlq   mm7,16
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w1],edx
+    mov     [ebp-w1],edx
     shr     eax,16
-    mov     [w2],eax
+    mov     [ebp-w2],eax
     psrlq   mm5,32
     movd    eax,mm5
-    mov     [w3],eax
+    mov     [ebp-w3],eax
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w4],edx
+    mov     [ebp-w4],edx
     shr     eax,16
-    mov     [w5],eax
+    mov     [ebp-w5],eax
     psrlq   mm6,32
     movd    eax,mm6
-    mov     [w6],eax
+    mov     [ebp-w6],eax
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w7],edx
+    mov     [ebp-w7],edx
     shr     eax,16
-    mov     [w8],eax
+    mov     [ebp-w8],eax
     psrlq   mm7,32
     movd    eax,mm7
-    mov     [w9],eax
+    mov     [ebp-w9],eax
     jmp     .flags
 .xres_1
-    cmp     dword[xcounter],-1
+    cmp     dword[ebp-xcounter],-1
     jl      .nexty
     ; x=Xres-1 - special case
-    mov     ebx,[prevline]
+    mov     ebx,[ebp-prevline]
     movq    mm5,[esi+ebx-6]
     movq    mm6,[esi-6]
-    mov     ebx,[nextline]
+    mov     ebx,[ebp-nextline]
     movq    mm7,[esi+ebx-6]
     psrlq   mm5,32
     psrlq   mm6,32
     psrlq   mm7,32
     movd    eax,mm5
     movzx   edx,ax
-    mov     [w1],edx
+    mov     [ebp-w1],edx
     shr     eax,16
-    mov     [w2],eax
-    mov     [w3],eax
+    mov     [ebp-w2],eax
+    mov     [ebp-w3],eax
     movd    eax,mm6
     movzx   edx,ax
-    mov     [w4],edx
+    mov     [ebp-w4],edx
     shr     eax,16
-    mov     [w5],eax
-    mov     [w6],eax
+    mov     [ebp-w5],eax
+    mov     [ebp-w6],eax
     movd    eax,mm7
     movzx   edx,ax
-    mov     [w7],edx
+    mov     [ebp-w7],edx
     shr     eax,16
-    mov     [w8],eax
-    mov     [w9],eax
+    mov     [ebp-w8],eax
+    mov     [ebp-w9],eax
     jmp     .flags
 .nexty
     add		esi,[ebp+offset] ; added - move source pointer past end-of-line blanks
@@ -3901,19 +3903,19 @@ NEWSYM hq4x_16
     shl     ebx, 3
     sub     edi, ebx
     shr     ebx, 2
-    dec     dword[linesleft]
+    dec     dword[ebp-linesleft]
     jz      .fin
-    add		ebx, [ebp+offset];
-    cmp     dword[linesleft],1
+     add		ebx, [ebp+offset];
+    cmp     dword[ebp-linesleft],1
     je      .lastline
-    mov     dword[nextline],ebx
+    mov     dword[ebp-nextline],ebx
     neg     ebx
-    mov     dword[prevline],ebx
+    mov     dword[ebp-prevline],ebx
     jmp     .loopy
 .lastline
-    mov     dword[nextline],0
+    mov     dword[ebp-nextline],0
     neg     ebx
-    mov     dword[prevline],ebx
+    mov     dword[ebp-prevline],ebx
     jmp     .loopy
 .fin
     emms
diff --git a/src/win32/Direct3D.cpp b/src/win32/Direct3D.cpp
index 785d4a9f..10ee3d36 100644
--- a/src/win32/Direct3D.cpp
+++ b/src/win32/Direct3D.cpp
@@ -21,7 +21,7 @@
 #ifndef NO_D3D
 
 // The number of pixel-filter threads to be created
-#define NTHREADS 4
+#define NTHREADS ( theApp.maxCpuCores )
 
 #pragma comment( lib, "d3d9" )
 #pragma comment( lib, "d3dx9" )
@@ -454,7 +454,7 @@ void Direct3DDisplay::render()
 			u32 dst_bytes_per_thread = lr.Pitch * dst_height_per_thread;
 
 			// Use Multi Threading
-			assert( NTHREADS > 0 );
+			assert( ( NTHREADS > 0 ) && ( NTHREADS < MAXIMUM_PROCESSORS ) );
 			for( int i = ( NTHREADS - 1 ) ; i > -1 ; i-- ) {
 				// create last thread first because it could have more work than the others (for eg. if NTHREADS = 3)
 				// (last thread has to process the remaining lines if (height / NTHREADS) is not an integer)
@@ -819,7 +819,7 @@ void Direct3DDisplay::calculateDestRect()
 		rectangleFillsScreen = true; // no clear() necessary
 		destRect.left = 0;
 		destRect.top = 0;
-		destRect.right = dpp.BackBufferWidth;   // for some reason there'l be a black
+		destRect.right = dpp.BackBufferWidth;   // for some reason there will be a black
 		destRect.bottom = dpp.BackBufferHeight; // border line when using -1 at the end
 	} else {
 		// use aspect ratio
diff --git a/src/win32/VBA.cpp b/src/win32/VBA.cpp
index 1acc4fea..674a7f59 100644
--- a/src/win32/VBA.cpp
+++ b/src/win32/VBA.cpp
@@ -24,6 +24,7 @@
 #endif
 
 #include "stdafx.h"
+#include <intrin.h>
 
 #include "AVIWrite.h"
 #include "LangSelect.h"
@@ -207,16 +208,6 @@ void directXMessage(const char *msg)
 /////////////////////////////////////////////////////////////////////////////
 // VBA
 
-BEGIN_MESSAGE_MAP(VBA, CWinApp)
-  //{{AFX_MSG_MAP(VBA)
-  // NOTE - the ClassWizard will add and remove mapping macros here.
-  //    DO NOT EDIT what you see in these blocks of generated code!
-  //}}AFX_MSG_MAP
-  END_MESSAGE_MAP()
-
-  /////////////////////////////////////////////////////////////////////////////
-// VBA construction
-
 VBA::VBA()
 {
   // COINIT_MULTITHREADED is not supported by SHBrowseForFolder with BIF_USENEWUI
@@ -229,6 +220,7 @@ VBA::VBA()
   mode800Available = false;
   mode1024Available = false;
   mode1280Available = false;
+  maxCpuCores = 1;
   windowPositionX = 0;
   windowPositionY = 0;
   filterFunction = NULL;
@@ -1486,6 +1478,14 @@ void VBA::loadSettings()
   if(windowPositionY < 0)
     windowPositionY = 0;
 
+  maxCpuCores = regQueryDwordValue("maxCpuCores", 0);
+  if(maxCpuCores < 0) {
+	  maxCpuCores = 0;
+  }
+  if(maxCpuCores == 0) {
+	  maxCpuCores = detectCpuCores();
+  }
+
   useBiosFileGBA = ( regQueryDwordValue("useBiosGBA", 0) == 1 ) ? true : false;
 
   useBiosFileGB = ( regQueryDwordValue("useBiosGB", 0) == 1 ) ? true : false;
@@ -2474,6 +2474,8 @@ void VBA::saveSettings()
   regSetDwordValue("windowX", windowPositionX);
   regSetDwordValue("windowY", windowPositionY);
 
+  regSetDwordValue("maxCpuCores", maxCpuCores);
+
   regSetDwordValue("useBiosGBA", useBiosFileGBA);
 
   regSetDwordValue("useBiosGB", useBiosFileGB);
@@ -2587,6 +2589,17 @@ void VBA::saveSettings()
 #endif
 }
 
+int VBA::detectCpuCores()
+{
+	int CPUInfo[4];
+
+	__cpuid( CPUInfo, 1 );
+
+	int processor_count = ( CPUInfo[1] & 0x00FF0000 ) >> 16;
+
+	return processor_count;
+}
+
 void winSignal(int, int)
 {
 }
diff --git a/src/win32/VBA.h b/src/win32/VBA.h
index 9fb7f6fa..9a6b333b 100644
--- a/src/win32/VBA.h
+++ b/src/win32/VBA.h
@@ -79,6 +79,7 @@ class VBA : public CWinApp
   bool mode800Available;
   bool mode1024Available;
   bool mode1280Available;
+  int maxCpuCores; // maximum number of CPU cores VBA should use, 0 means auto-detect
   int windowPositionX;
   int windowPositionY;
   void (*filterFunction)(u8*,u32,u8*,u8*,u32,int,int);
@@ -262,13 +263,10 @@ class VBA : public CWinApp
   void updateFrameSkip();
   void loadSettings();
   void addRecentFile(CString file);
-  //{{AFX_MSG(VBA)
-  afx_msg void OnAppAbout();
-  // NOTE - the ClassWizard will add and remove member functions here.
-  //    DO NOT EDIT what you see in these blocks of generated code !
-  //}}AFX_MSG
-  DECLARE_MESSAGE_MAP()
-    };
+
+  private:
+  int detectCpuCores();
+};
 
     extern VBA theApp;
 	extern int emulating;