diff --git a/Makefile b/Makefile
index 2c22352e..b188a47b 100644
--- a/Makefile
+++ b/Makefile
@@ -1,9 +1,9 @@
 CC=gcc
 CPPC=g++
-CFLAGS=-W -Wall -Wno-unused -O3 -DHAVE_NETINET_IN_H -DHAVE_ARPA_INET_H -DFINAL_VERSION -DBKPT_SUPPORT -DC_CORE -DSDL -DSYSCONFDIR="home" -DUSE_OPENGL
+CFLAGS=-W -Wall -Wno-unused -O3 -DHAVE_NETINET_IN_H -DHAVE_ARPA_INET_H -DFINAL_VERSION -DBKPT_SUPPORT -DSDL -DSYSCONFDIR="home" -DUSE_OPENGL
 CXXFLAGS=${CFLAGS}
 ASM=nasm
-ASMFLAGS=-w-orphan-labels -f elf -DELF -O1 -Isrc/
+ASMFLAGS=-w-orphan-labels -f elf -DELF -O1 -Isrc/hq/asm/
 LFLAGS=-lz -lpng -lGL `sdl-config --libs`
 STRIP=strip -s
 DEL=rm -f
@@ -11,7 +11,7 @@ OE=.o
 OUT=vba
 
 ifeq ($(PLATFORM),win)
-  ASMFLAGS=-w-orphan-labels -f win32 -O1 -Isrc/
+  ASMFLAGS=-w-orphan-labels -f win32 -O1 -Isrc/hq/asm/
   LFLAGS=-lz -lpng -lSDL -lwsock32 -lopengl32
   DELETECOMMAND = del
   OE=.obj
@@ -21,7 +21,7 @@ endif
 ifeq ($(PLATFORM),win-cross)
   CC=i586-mingw32-gcc
   CPPC=i586-mingw32-g++
-  ASMFLAGS=-w-orphan-labels -f win32 -O1 -Isrc/
+  ASMFLAGS=-w-orphan-labels -f win32 -O1 -Isrc/hq/asm/
   LFLAGS=-lz -lpng -lSDL -lwsock32 -lopengl32
   STRIP=i586-mingw32-strip -s
   OE=.obj
@@ -33,16 +33,17 @@ SDLDIR=src/sdl
 DMGDIR=src/gb
 GBAPUDIR=src/gb/gb_apu
 FEXDIR=../dependencies/File_Extractor-0.4.3
+HQCDIR=src/hq/c
+HQASMDIR=src/hq/asm
 
 
-ASMOBJ=${MAINDIR}/hq3x_16${OE} ${MAINDIR}/hq3x_32${OE} ${MAINDIR}/hq4x_16${OE} ${MAINDIR}/hq4x_32${OE} \
-${MAINDIR}/hq3x32${OE}
+ASMOBJ=${HQASMDIR}/hq3x_16${OE} ${HQASMDIR}/hq3x_32${OE} ${HQASMDIR}/hq4x_16${OE} \
+${HQASMDIR}/hq4x_32${OE} ${HQASMDIR}/hq3x32${OE}
 
 GBAPUOBJ=${GBAPUDIR}/Blip_Buffer${OE} ${GBAPUDIR}/Effects_Buffer${OE} ${GBAPUDIR}/Gb_Apu${OE} \
 ${GBAPUDIR}/Gb_Apu_State${OE} ${GBAPUDIR}/Gb_Oscs${OE} ${GBAPUDIR}/Multi_Buffer${OE}
 
-CALTERNOBJ=${MAINDIR}/hq3x16c${OE} ${MAINDIR}/hq3x32c${OE} ${MAINDIR}/hq4x16c${OE} ${MAINDIR}/hq4x32c${OE} \
-${MAINDIR}/hq_shared32${OE}
+CALTERNOBJ=${HQCDIR}/hq_implementation${OE}
 
 MAINOBJ=${MAINDIR}/2xSaI${OE} ${MAINDIR}/admame${OE} ${MAINDIR}/agbprint${OE} ${MAINDIR}/armdis${OE} \
 ${MAINDIR}/bilinear${OE} ${MAINDIR}/bios${OE} ${MAINDIR}/Cheats${OE} ${MAINDIR}/CheatSearch${OE} \
@@ -68,6 +69,7 @@ ifeq ($(USEASM),yes)
 OBJECTS+=${ASMOBJ}
 else
 OBJECTS+=${CALTERNOBJ}
+CFLAGS+=-DC_CORE
 endif
 
 ifeq ($(USEFEX),yes)
diff --git a/src/GBA-arm.cpp b/src/GBA-arm.cpp
index 82e4d459..500c3858 100644
--- a/src/GBA-arm.cpp
+++ b/src/GBA-arm.cpp
@@ -321,10 +321,10 @@ static void count(u32 opcode, int cond_res)
  #define EMIT1(op,arg)        #op" "arg"; "
  #define EMIT2(op,src,dest)   #op" "src", "dest"; "
  #define CONST(val)           "$"#val
- #define VAR(var)             #var
- #define VARL(var)            #var
- #define REGREF1(index)       "reg("index")"
- #define REGREF2(index,scale) "reg(,"index","#scale")"
+ #define VAR(var)             "_"#var
+ #define VARL(var)            "_"#var
+ #define REGREF1(index)       "_reg("index")"
+ #define REGREF2(index,scale) "_reg(,"index","#scale")"
  #define LABEL(n)             #n": "
  #define LABELREF(n,dir)      #n#dir
  #define al "%%al"
@@ -697,7 +697,7 @@ static void count(u32 opcode, int cond_res)
         : "0" (offset), "c" (shift));
 
 #define RRX_OFFSET \
-    asm("btl $0, C_FLAG;"               \
+    asm("btl $0, _C_FLAG;"               \
         "rcr $1, %0"                    \
         : "=r" (offset)                 \
         : "0" (offset));
diff --git a/src/GBA-thumb.cpp b/src/GBA-thumb.cpp
index 53047993..aa1534d6 100644
--- a/src/GBA-thumb.cpp
+++ b/src/GBA-thumb.cpp
@@ -324,95 +324,95 @@ static INSN_REGPARM void thumbBreakpoint(u32 opcode)
   #define ADD_RN_O8(d) \
      asm ("andl $0xFF, %%eax;"\
           "addl %%eax, %0;"\
-          "setsb N_FLAG;"\
-          "setzb Z_FLAG;"\
-          "setcb C_FLAG;"\
-          "setob V_FLAG;"\
+          "setsb _N_FLAG;"\
+          "setzb _Z_FLAG;"\
+          "setcb _C_FLAG;"\
+          "setob _V_FLAG;"\
           : "=m" (reg[(d)].I));
   #define CMN_RD_RS \
      asm ("add %0, %1;"\
-          "setsb N_FLAG;"\
-          "setzb Z_FLAG;"\
-          "setcb C_FLAG;"\
-          "setob V_FLAG;"\
+          "setsb _N_FLAG;"\
+          "setzb _Z_FLAG;"\
+          "setcb _C_FLAG;"\
+          "setob _V_FLAG;"\
           : \
           : "r" (value), "r" (reg[dest].I):"1");
   #define ADC_RD_RS \
-     asm ("bt $0, C_FLAG;"\
+     asm ("bt $0, _C_FLAG;"\
           "adc %1, %%ebx;"\
-          "setsb N_FLAG;"\
-          "setzb Z_FLAG;"\
-          "setcb C_FLAG;"\
-          "setob V_FLAG;"\
+          "setsb _N_FLAG;"\
+          "setzb _Z_FLAG;"\
+          "setcb _C_FLAG;"\
+          "setob _V_FLAG;"\
           : "=b" (reg[dest].I)\
           : "r" (value), "b" (reg[dest].I));
   #define SUB_RN_O8(d) \
      asm ("andl $0xFF, %%eax;"\
           "subl %%eax, %0;"\
-          "setsb N_FLAG;"\
-          "setzb Z_FLAG;"\
-          "setncb C_FLAG;"\
-          "setob V_FLAG;"\
+          "setsb _N_FLAG;"\
+          "setzb _Z_FLAG;"\
+          "setncb _C_FLAG;"\
+          "setob _V_FLAG;"\
           : "=m" (reg[(d)].I));
   #define MOV_RN_O8(d) \
      asm ("andl $0xFF, %%eax;"\
-          "movb $0, N_FLAG;"\
+          "movb $0, _N_FLAG;"\
           "movl %%eax, %0;"\
-          "setzb Z_FLAG;"\
+          "setzb _Z_FLAG;"\
           : "=m" (reg[(d)].I));
   #define CMP_RN_O8(d) \
      asm ("andl $0xFF, %%eax;"\
           "cmpl %%eax, %0;"\
-          "setsb N_FLAG;"\
-          "setzb Z_FLAG;"\
-          "setncb C_FLAG;"\
-          "setob V_FLAG;"\
+          "setsb _N_FLAG;"\
+          "setzb _Z_FLAG;"\
+          "setncb _C_FLAG;"\
+          "setob _V_FLAG;"\
           : \
           : "m" (reg[(d)].I));
   #define SBC_RD_RS \
-     asm volatile ("bt $0, C_FLAG;"\
+     asm volatile ("bt $0, _C_FLAG;"\
                    "cmc;"\
                    "sbb %1, %%ebx;"\
-                   "setsb N_FLAG;"\
-                   "setzb Z_FLAG;"\
-                   "setncb C_FLAG;"\
-                   "setob V_FLAG;"\
+                   "setsb _N_FLAG;"\
+                   "setzb _Z_FLAG;"\
+                   "setncb _C_FLAG;"\
+                   "setob _V_FLAG;"\
                    : "=b" (reg[dest].I)\
                    : "r" (value), "b" (reg[dest].I) : "cc", "memory");
   #define LSL_RD_RS \
          asm ("shl %%cl, %%eax;"\
-              "setcb C_FLAG;"\
+              "setcb _C_FLAG;"\
               : "=a" (value)\
               : "a" (reg[dest].I), "c" (value));
   #define LSR_RD_RS \
          asm ("shr %%cl, %%eax;"\
-              "setcb C_FLAG;"\
+              "setcb _C_FLAG;"\
               : "=a" (value)\
               : "a" (reg[dest].I), "c" (value));
   #define ASR_RD_RS \
          asm ("sar %%cl, %%eax;"\
-              "setcb C_FLAG;"\
+              "setcb _C_FLAG;"\
               : "=a" (value)\
               : "a" (reg[dest].I), "c" (value));
   #define ROR_RD_RS \
          asm ("ror %%cl, %%eax;"\
-              "setcb C_FLAG;"\
+              "setcb _C_FLAG;"\
               : "=a" (value)\
               : "a" (reg[dest].I), "c" (value));
   #define NEG_RD_RS \
      asm ("neg %%ebx;"\
-          "setsb N_FLAG;"\
-          "setzb Z_FLAG;"\
-          "setncb C_FLAG;"\
-          "setob V_FLAG;"\
+          "setsb _N_FLAG;"\
+          "setzb _Z_FLAG;"\
+          "setncb _C_FLAG;"\
+          "setob _V_FLAG;"\
           : "=b" (reg[dest].I)\
           : "b" (reg[source].I));
   #define CMP_RD_RS \
      asm ("sub %0, %1;"\
-          "setsb N_FLAG;"\
-          "setzb Z_FLAG;"\
-          "setncb C_FLAG;"\
-          "setob V_FLAG;"\
+          "setsb _N_FLAG;"\
+          "setzb _Z_FLAG;"\
+          "setncb _C_FLAG;"\
+          "setob _V_FLAG;"\
           : \
           : "r" (value), "r" (reg[dest].I):"1");
   #define IMM5_INSN(OP,N) \
@@ -420,75 +420,75 @@ static INSN_REGPARM void thumbBreakpoint(u32 opcode)
          "shrl $1,%%eax;"            \
          "andl $7,%%ecx;"            \
          "andl $0x1C,%%eax;"         \
-         "movl reg(%%eax),%%edx;"    \
+         "movl _reg(%%eax),%%edx;"    \
          OP                          \
-         "setsb N_FLAG;"             \
-         "setzb Z_FLAG;"             \
-         "movl %%edx,reg(,%%ecx,4);" \
+         "setsb _N_FLAG;"             \
+         "setzb _Z_FLAG;"             \
+         "movl %%edx,_reg(,%%ecx,4);" \
          : : "i" (N))
   #define IMM5_INSN_0(OP)            \
      asm("movl %%eax,%%ecx;"         \
          "shrl $1,%%eax;"            \
          "andl $7,%%ecx;"            \
          "andl $0x1C,%%eax;"         \
-         "movl reg(%%eax),%%edx;"    \
+         "movl _reg(%%eax),%%edx;"    \
          OP                          \
-         "setsb N_FLAG;"             \
-         "setzb Z_FLAG;"             \
-         "movl %%edx,reg(,%%ecx,4);" \
+         "setsb _N_FLAG;"             \
+         "setzb _Z_FLAG;"             \
+         "movl %%edx,_reg(,%%ecx,4);" \
          : : )
   #define IMM5_LSL \
          "shll %0,%%edx;"\
-         "setcb C_FLAG;"
+         "setcb _C_FLAG;"
   #define IMM5_LSL_0 \
          "testl %%edx,%%edx;"
   #define IMM5_LSR \
          "shrl %0,%%edx;"\
-         "setcb C_FLAG;"
+         "setcb _C_FLAG;"
   #define IMM5_LSR_0 \
          "testl %%edx,%%edx;"\
-         "setsb C_FLAG;"\
+         "setsb _C_FLAG;"\
          "xorl %%edx,%%edx;"
   #define IMM5_ASR \
          "sarl %0,%%edx;"\
-         "setcb C_FLAG;"
+         "setcb _C_FLAG;"
   #define IMM5_ASR_0 \
          "sarl $31,%%edx;"\
-         "setsb C_FLAG;"
+         "setsb _C_FLAG;"
   #define THREEARG_INSN(OP,N) \
      asm("movl %%eax,%%edx;"       \
          "shrl $1,%%edx;"          \
          "andl $0x1C,%%edx;"       \
          "andl $7,%%eax;"          \
-         "movl reg(%%edx),%%ecx;"  \
+         "movl _reg(%%edx),%%ecx;"  \
          OP(N)                     \
-         "setsb N_FLAG;"           \
-         "setzb Z_FLAG;"           \
-         "movl %%ecx,reg(,%%eax,4)"::)
+         "setsb _N_FLAG;"           \
+         "setzb _Z_FLAG;"           \
+         "movl %%ecx,_reg(,%%eax,4)"::)
   #define ADD_RD_RS_RN(N)          \
-         "add (reg+"#N"*4),%%ecx;" \
-         "setcb C_FLAG;"           \
-         "setob V_FLAG;"
+         "add (_reg+"#N"*4),%%ecx;" \
+         "setcb _C_FLAG;"           \
+         "setob _V_FLAG;"
   #define ADD_RD_RS_O3(N)          \
          "add $"#N",%%ecx;"        \
-         "setcb C_FLAG;"           \
-         "setob V_FLAG;"
+         "setcb _C_FLAG;"           \
+         "setob _V_FLAG;"
   #define ADD_RD_RS_O3_0(N)        \
-         "movb $0,C_FLAG;"         \
+         "movb $0,_C_FLAG;"         \
          "add $0,%%ecx;"           \
-         "movb $0,V_FLAG;"
+         "movb $0,_V_FLAG;"
   #define SUB_RD_RS_RN(N) \
-         "sub (reg+"#N"*4),%%ecx;" \
-         "setncb C_FLAG;"          \
-         "setob V_FLAG;"
+         "sub (_reg+"#N"*4),%%ecx;" \
+         "setncb _C_FLAG;"          \
+         "setob _V_FLAG;"
   #define SUB_RD_RS_O3(N) \
          "sub $"#N",%%ecx;"        \
-         "setncb C_FLAG;"          \
-         "setob V_FLAG;"
+         "setncb _C_FLAG;"          \
+         "setob _V_FLAG;"
   #define SUB_RD_RS_O3_0(N)        \
-         "movb $1,C_FLAG;"         \
+         "movb $1,_C_FLAG;"         \
          "sub $0,%%ecx;"           \
-         "movb $0,V_FLAG;"
+         "movb $0,_V_FLAG;"
 #endif
 #else // !__GNUC__
   #define ADD_RD_RS_RN(N) \
diff --git a/src/Gfx.h b/src/Gfx.h
index 20c2f054..8a04d430 100644
--- a/src/Gfx.h
+++ b/src/Gfx.h
@@ -297,6 +297,11 @@ static inline void gfxDrawRotScreen(u16 control,
     break;
   }
 
+  int maskX = sizeX-1;
+  int maskY = sizeY-1;
+
+  int yshift = ((control >> 14) & 3)+4;
+
   int dx = pa & 0x7FFF;
   if(pa & 0x8000)
     dx |= 0xFFFF8000;
@@ -343,12 +348,8 @@ static inline void gfxDrawRotScreen(u16 control,
   int yyy = (realY >> 8);
 
   if(control & 0x2000) {
-    xxx %= sizeX;
-    yyy %= sizeY;
-    if(xxx < 0)
-      xxx += sizeX;
-    if(yyy < 0)
-      yyy += sizeY;
+    xxx &= maskX;
+    yyy &= maskY;
   }
 
   if(control & 0x80) {
@@ -359,7 +360,7 @@ static inline void gfxDrawRotScreen(u16 control,
          yyy >= sizeY) {
         line[x] = 0x80000000;
       } else {
-        int tile = screenBase[(xxx>>3) + (yyy>>3)*(sizeX>>3)];
+        int tile = screenBase[(xxx>>3) + ((yyy>>3)<<yshift)];
 
         int tileX = (xxx & 7);
         int tileY = yyy & 7;
@@ -375,12 +376,8 @@ static inline void gfxDrawRotScreen(u16 control,
       yyy = (realY >> 8);
 
       if(control & 0x2000) {
-        xxx %= sizeX;
-        yyy %= sizeY;
-        if(xxx < 0)
-          xxx += sizeX;
-        if(yyy < 0)
-          yyy += sizeY;
+        xxx &= maskX;
+        yyy &= maskY;
       }
     }
   } else {
@@ -391,7 +388,7 @@ static inline void gfxDrawRotScreen(u16 control,
          yyy >= sizeY) {
         line[x] = 0x80000000;
       } else {
-        int tile = screenBase[(xxx>>3) + (yyy>>3)*(sizeX>>3)];
+        int tile = screenBase[(xxx>>3) + ((yyy>>3)<<yshift)];
 
         int tileX = (xxx & 7);
         int tileY = yyy & 7;
@@ -407,12 +404,8 @@ static inline void gfxDrawRotScreen(u16 control,
       yyy = (realY >> 8);
 
       if(control & 0x2000) {
-        xxx %= sizeX;
-        yyy %= sizeY;
-        if(xxx < 0)
-          xxx += sizeX;
-        if(yyy < 0)
-          yyy += sizeY;
+        xxx &= maskX;
+        yyy &= maskY;
       }
     }
   }