diff --git a/genctx.py b/genctx.py
index 105eba435..a172347be 100644
--- a/genctx.py
+++ b/genctx.py
@@ -25,28 +25,32 @@
 def search_directories(*patterns):
     return it.chain.from_iterable(glob.iglob(pattern,recursive=True) for pattern in patterns)
 
-for filename in search_directories('./src/**/*.h', './include/**/*.h'):
+for filename in search_directories('./include/**/*.h'):
     with open(filename) as header:
         header_name = "/".join(filename.split('/')[2:])
-        
-        # not needed for decomp work
-        if header_name.startswith("platform"):
-            continue
 
         if not header_name in depends_on:            
             depends_on[header_name] = set()
-            if not header_name.startswith("gba/") and header_name != "global.h" and header_name != "functions.h":
-                depends_on[header_name].add("global.h")
-            
-            if header_name.startswith("gba/") and not header_name.endswith("types.h") and not header_name.endswith("defines.h"):
-                depends_on[header_name].add("gba/types.h")
+
+            if not header_name.startswith("gba/"):
+                if header_name != "global.h"      \
+                and header_name != "config.h"     \
+                and header_name != "functions.h":
+                    depends_on[header_name].add("global.h")
             
-            if header_name.startswith("gba/") and not header_name.endswith("multiboot.h") and not header_name.endswith("types.h"):
-                depends_on[header_name].add("gba/multiboot.h")
+            else: # header_name.startswith("gba/")
+                if  not header_name.endswith("types.h") \
+                and not header_name.endswith("defines.h"):
+                    depends_on[header_name].add("gba/types.h")
+                elif not header_name.endswith("multiboot.h") \
+                and not header_name.endswith("types.h"):
+                    depends_on[header_name].add("gba/multiboot.h")
                 
         data[header_name] = ""
         for line in header.readlines():
-            if "#include" in line and not line.startswith("//") and '<' not in line:
+            if "#include" in line         \
+            and not line.startswith("//") \
+            and '<' not in line:
                 requires = line.split('"')[1]
                 depends_on[header_name].add(requires)
                 continue
@@ -58,7 +62,6 @@ def search_directories(*patterns):
 print_order = tuple(ts.static_order())
 
 with open('ctx.c', 'w') as context:
-    context.write("#define PLATFORM_GBA 1\n")
     for header in print_order:
         exclude = False
         for e in excluded:
diff --git a/include/config.h b/include/config.h
index afadc0557..ee95e6bad 100644
--- a/include/config.h
+++ b/include/config.h
@@ -3,13 +3,23 @@
 
 /* TODO: Move config.h into a different location? */
 
-#define GAME_SA1 1
-#define GAME_SA2 2
-#define GAME_SA3 3
+#define GAME_SA1   1
+#define GAME_SA2   2
+#define GAME_SA3   3
+#define GAME_KATAM 4
 
-// TODO: Define this in Makefile through a compiler macro!
+#define ENGINE_1 1
+#define ENGINE_2 2
+#define ENGINE_3 3
+#define ENGINE_4 4
+
+// TODO: Define this in Makefile through a compiler macro?
 #define GAME GAME_SA2
 
+// TODO: Do SA1 and SA2 use the same engine ver?
+// TODO: Do SA3 and KATAM use the same engine ver?
+#define ENGINE GAME
+
 // TODO: Put somewhere else?
 #if PLATFORM_GBA
 #define USE_NEW_DMA 0
diff --git a/include/core.h b/include/core.h
index ca6fa4e34..16e80022b 100644
--- a/include/core.h
+++ b/include/core.h
@@ -368,7 +368,7 @@ extern struct MultiBootParam gMultiBootParam;
 
 extern const struct SpriteTables *gRefSpriteTables;
 
-void GameInit(void);
-void GameLoop(void);
+void EngineInit(void);
+void EngineMainLoop(void);
 
 #endif
diff --git a/include/game/game.h b/include/game/game.h
index 7cc1af7f5..1d5e6edc6 100644
--- a/include/game/game.h
+++ b/include/game/game.h
@@ -3,6 +3,6 @@
 
 #include "global.h"
 
-void GameStart(void);
+void GameInit(void);
 
 #endif // GUARD_GAME_H
diff --git a/include/game/parameters/characters.h b/include/game/parameters/characters.h
index c4900bf24..265989455 100644
--- a/include/game/parameters/characters.h
+++ b/include/game/parameters/characters.h
@@ -3,6 +3,8 @@
 
 #include "constants/zones.h"
 
+// TODO: Be consistent about whether these are Q() or I() values!
+
 /*** Common ***/
 #define PLAYER_FLYING_END_GRAVITY  (0.033) // = (8. / 256.)
 #define PLAYER_GRAVITY             42.0 / 256.0
diff --git a/include/sprite.h b/include/sprite.h
index 0649d23a5..207e98ba3 100644
--- a/include/sprite.h
+++ b/include/sprite.h
@@ -93,11 +93,18 @@ typedef struct {
 } Background; /* size = 0x40 */
 
 typedef struct {
+#if (ENGINE >= ENGINE_3)
+    // In SA3 flip-bits are integrated into the oamIndex.
+    // X-Flip: Bit 14
+    // Y-Flip: Bit 15
+    /* 0x00 */ u16 oamIndex;
+#else
     /* 0x00 */ u8 flip;
 
     // every animation has an associated oamData pointer, oamIndex starts at
     // 0 for every new animation and ends at variantCount-1
     /* 0x01 */ u8 oamIndex;
+#endif
 
     // some sprite frames consist of multiple images (of the same size
     // as GBA's Object Attribute Memory, e.g. 8x8, 8x32, 32x64, ...)
diff --git a/ldscript.txt b/ldscript.txt
index 9464d07c2..38fd76a22 100644
--- a/ldscript.txt
+++ b/ldscript.txt
@@ -39,6 +39,7 @@ SECTIONS
 
         /* engine */
         src/core.o(.data);
+        src/input_recorder.o(.data);
 
         /* "SA1 Leftovers" */
         . = ALIGN(32); 
diff --git a/src/core.c b/src/core.c
index dfe1ce94f..ff3e64a39 100644
--- a/src/core.c
+++ b/src/core.c
@@ -35,7 +35,10 @@ u8 gNumHBlankIntrs = 0;
 struct BlendRegs gBldRegs ALIGNED(8) = {};
 u8 gOamFreeIndex = 0;
 struct Task gEmptyTask ALIGNED(16) = {};
-BgAffineReg gBgAffineRegs[NUM_AFFINE_BACKGROUNDS] ALIGNED(8) = {};
+
+// NOTE: gNextFreeAffineIndex introduced in SA3, unused before.
+u8 gNextFreeAffineIndex = 0;
+BgAffineReg gBgAffineRegs[NUM_AFFINE_BACKGROUNDS] = {};
 void *gVramHeapStartAddr = NULL;
 u16 gUnknown_03001944 ALIGNED(4) = 0;
 u8 gUnknown_03001948 ALIGNED(4) = 0;
@@ -124,8 +127,6 @@ FuncType_030053A0 gUnknown_030053A0[] ALIGNED(16) = {};
 const u8 *gInputPlaybackData = NULL;
 bool8 gExecSoundMain ALIGNED(4) = FALSE;
 s32 gPseudoRandom = 0;
-struct InputRecorder gInputRecorder ALIGNED(8) = {};
-u16 *gInputRecorderTapeBuffer = NULL;
 
 static void UpdateScreenDma(void);
 static void UpdateScreenCpuSet(void);
@@ -179,7 +180,7 @@ static VBlankFunc const sVblankFuncs[] = {
     sub_8002B20,
 };
 
-void GameInit(void)
+void EngineInit(void)
 {
     s16 i;
     u16 errorIdentifying;
@@ -252,6 +253,9 @@ void GameInit(void)
     gBgAffineRegs[1].x = 0;
     gBgAffineRegs[1].y = 0;
 
+#if (ENGINE >= ENGINE_3)
+    gNextFreeAffineIndex = 0;
+#endif
     gUnknown_03001944 = 0;
     gUnknown_030017F0 = 0x100;
     gUnknown_03005394 = 0x100;
@@ -358,7 +362,7 @@ void GameInit(void)
     MultiSioInit(0);
 }
 
-void GameLoop(void)
+void EngineMainLoop(void)
 {
     while (TRUE) {
         gExecSoundMain = FALSE;
@@ -379,6 +383,10 @@ void GameLoop(void)
         gFlagsPreVBlank = gFlags;
         VBlankIntrWait();
 
+#if (ENGINE >= ENGINE_3)
+        gNextFreeAffineIndex = 0;
+#endif
+
         if (gFlags & FLAGS_4000) {
             UpdateScreenCpuSet();
 
diff --git a/src/game/game.c b/src/game/game.c
index fa0bfd94c..9646bb539 100644
--- a/src/game/game.c
+++ b/src/game/game.c
@@ -30,7 +30,7 @@
 
 #include "data/sprite_tables.h"
 
-void GameStart(void)
+void GameInit(void)
 {
     u32 i;
     bool32 hasProfile = FALSE;
@@ -77,7 +77,7 @@ void GameStart(void)
         hasProfile = TRUE;
     }
 
-    // This flag is only set in GameInit
+    // This flag is only set in EngineInit
     if (gFlags & FLAGS_200) {
         ShowSinglePakResults();
         return;
diff --git a/src/game/math.c b/src/game/math.c
index b00d30a8a..0123f7224 100644
--- a/src/game/math.c
+++ b/src/game/math.c
@@ -27,6 +27,15 @@ u32 gRngValue = 0;
 
 const u16 gUnknown_080E0290[] = { 0x0AAA, 0x02AA };
 
+typedef struct {
+    void *unk0;
+    void *start;
+    void *next;
+    void *unkC;
+} UNK_8085DEC;
+
+bool8 sub_8085D98(UNK_8085DEC *thing, UNK_8085DEC *target);
+
 #define RAND_CONST 0x37119371;
 
 NONMATCH("asm/non_matching/game/math/unused_sub_80832FC.inc", void sub_80832FC()) { }
@@ -609,15 +618,6 @@ void sub_8085D14(UNK_8085D14 *p1, u32 p2, u32 p3, u32 p4)
     p1->unk12 = p4;
 }
 
-typedef struct {
-    void *unk0;
-    void *start;
-    void *next;
-    void *unkC;
-} UNK_8085DEC;
-
-bool8 sub_8085D98(UNK_8085DEC *thing, UNK_8085DEC *target);
-
 void sub_8085D44(UNK_8085DEC *thing)
 {
     thing->unk0 = NULL;
diff --git a/src/game/title_screen.c b/src/game/title_screen.c
index e9be28940..a7f935fba 100644
--- a/src/game/title_screen.c
+++ b/src/game/title_screen.c
@@ -553,7 +553,13 @@ static void InitTitleScreenUI(TitleScreen *titleScreen)
     s->graphics.anim = SA2_ANIM_TITLE_COPYRIGHT;
     s->variant = SA2_ANIM_VARIANT_COPYRIGHT_2003;
     s->prevVariant = -1;
+#if PLATFORM_GBA
     s->x = 0;
+#else
+    // Only display the COPYRIGHT message, not the "Licensed by ***" msg
+    // TODO: Split the graphics properly.
+    s->x = DISPLAY_WIDTH - 136;
+#endif
     s->y = DISPLAY_HEIGHT - 30; // set to the screen's bottom
     s->graphics.size = 0;
     s->oamFlags = SPRITE_OAM_ORDER(4);
@@ -1892,9 +1898,7 @@ static void Task_StartTitleScreenDemo(void)
     gSelectedCharacter = CHARACTER_SONIC;
     gCurrentLevel = sDemoLevels[0];
 
-    gDemoPlayCounter++;
-    // Don't count higher than 3
-    gDemoPlayCounter &= 3;
+    gDemoPlayCounter = (gDemoPlayCounter + 1) % 4u;
 
     gGameMode = GAME_MODE_SINGLE_PLAYER;
 
diff --git a/src/input_recorder.c b/src/input_recorder.c
index 0c41fb2a0..e47954b64 100644
--- a/src/input_recorder.c
+++ b/src/input_recorder.c
@@ -11,6 +11,9 @@
 #define TAPE_LENGTH 0x800
 #endif
 
+struct InputRecorder gInputRecorder ALIGNED(8) = { 0 };
+u16 *gInputRecorderTapeBuffer = NULL;
+
 static void Task_InputRecorder(void);
 static void InputRecorderEject(struct Task *);
 
diff --git a/src/main.c b/src/main.c
index 547787dd1..4c79283d0 100644
--- a/src/main.c
+++ b/src/main.c
@@ -3,7 +3,7 @@
 
 void AgbMain(void)
 {
+    EngineInit();
     GameInit();
-    GameStart();
-    GameLoop();
+    EngineMainLoop();
 }
diff --git a/src/multi_sio.c b/src/multi_sio.c
index c580b3265..ceb29fdd8 100644
--- a/src/multi_sio.c
+++ b/src/multi_sio.c
@@ -1,4 +1,5 @@
 #include "global.h"
+#include "game/game.h"
 #include "multi_sio.h"
 #include "sio32_multi_load.h"
 
@@ -10,7 +11,9 @@ u32 gMultiSioIntrFuncBuf[0x120 / 4] = {}; // Interrupt Routine RAM Execution Buf
 
 struct MultiSioArea gMultiSioArea = {};
 
+#if (GAME <= GAME_SA2)
 UNUSED u32 gUnusedMultiSioSpace[2] = {};
+#endif
 
 #ifdef MULTI_SIO_DI_FUNC_FAST
 u32 gMultiSioRecvFuncBuf[0x40 / 4] = {}; // Receive Data/Check Buffer Change Routine RAM Execution Buffer
diff --git a/src/platform/pret_sdl/sdl2.c b/src/platform/pret_sdl/sdl2.c
index 138bed81a..4b6f385db 100644
--- a/src/platform/pret_sdl/sdl2.c
+++ b/src/platform/pret_sdl/sdl2.c
@@ -243,7 +243,7 @@ int main(int argc, char **argv)
 #if ENABLE_VRAM_VIEW
     VramDraw(vramTexture);
 #endif
-    // Prevent the multiplayer screen from being drawn ( see core.c:GameInit() )
+    // Prevent the multiplayer screen from being drawn ( see core.c:EngineInit() )
     REG_RCNT = 0x8000;
     REG_KEYINPUT = 0x3FF;
 
diff --git a/src/platform/win32/win32.c b/src/platform/win32/win32.c
index d52d2cdbe..47d9dbf7d 100644
--- a/src/platform/win32/win32.c
+++ b/src/platform/win32/win32.c
@@ -4,9 +4,7 @@
 #include "global.h"
 #include "core.h"
 
-extern void GameInit(void);
-extern void GameStart(void);
-extern void GameLoop(void);
+void GameInit(void);
 
 DWORD WINAPI GameThread(void *pThreadParam);
 
@@ -25,8 +23,8 @@ int WINAPI WinMain(HINSTANCE instance, HINSTANCE prevInstance, LPSTR lpCmdLine,
     REG_RCNT = 0x8000;
     REG_KEYINPUT = 0x3FF;
 
+    EngineInit();
     GameInit();
-    GameStart();
 
     DWORD threadId;
     HANDLE GameThreadHandle = CreateThread(NULL, 0, GameThread, NULL, 0, &threadId);
@@ -39,7 +37,7 @@ DWORD WINAPI GameThread(void *pThreadParam)
     REG_KEYINPUT &= ~START_BUTTON;
     while (1) {
         gFlags |= 0x4000;
-        GameLoop();
+        EngineMainLoop();
         REG_KEYINPUT ^= (A_BUTTON | START_BUTTON);
         REG_KEYINPUT |= (DPAD_RIGHT);
     }
diff --git a/src/sprite.c b/src/sprite.c
index cd9caa152..205579b59 100644
--- a/src/sprite.c
+++ b/src/sprite.c
@@ -629,12 +629,12 @@ void DisplaySprite(Sprite *sprite)
     }
 }
 
-void sub_081569A0(Sprite *sprite, u16 *sp08, u8 sp0C)
+void DisplaySprites(Sprite *sprite, Vec2_16 *positions, u8 numPositions)
 {
     vs32 x, y;
     s32 sprWidth, sprHeight;
     u8 subframe, i;
-    u32 x1, y1, sp24, sp28;
+    s32 x1, y1, centerOffsetX, centerOffsetY;
 
     if (sprite->dimensions != (void *)-1) {
         const SpriteOffset *sprDims = sprite->dimensions;
@@ -671,8 +671,8 @@ void sub_081569A0(Sprite *sprite, u16 *sp08, u8 sp0C)
             }
         }
 
-        sp24 = x - sprite->x;
-        sp28 = y - sprite->y;
+        centerOffsetX = x - sprite->x;
+        centerOffsetY = y - sprite->y;
         if (x + sprWidth >= 0 && x <= DISPLAY_WIDTH && y + sprHeight >= 0 && y <= DISPLAY_HEIGHT) {
             for (subframe = 0; subframe < sprDims->numSubframes; ++subframe) {
                 const u16 *oamData = gRefSpriteTables->oamData[sprite->graphics.anim];
@@ -728,7 +728,7 @@ void sub_081569A0(Sprite *sprite, u16 *sp08, u8 sp0C)
                 }
                 oam->all.attr2 += GET_TILE_NUM(sprite->graphics.dest);
 
-                for (i = 0; i < sp0C; ++i) {
+                for (i = 0; i < numPositions; ++i) {
                     OamData *r5 = OamMalloc(GET_SPRITE_OAM_ORDER(sprite));
 
                     if (iwram_end == oam)
@@ -736,8 +736,8 @@ void sub_081569A0(Sprite *sprite, u16 *sp08, u8 sp0C)
                     DmaCopy16(3, oam, r5, sizeof(OamDataShort));
                     r5->all.attr1 &= 0xFE00;
                     r5->all.attr0 &= 0xFF00;
-                    r5->all.attr0 += (sp08[2 * i + 1] + sp28 + y1) & 0xFF;
-                    r5->all.attr1 += (sp08[2 * i + 0] + sp24 + x1) & 0x1FF;
+                    r5->all.attr0 += (positions[i].y + centerOffsetY + y1) & 0xFF;
+                    r5->all.attr1 += (positions[i].x + centerOffsetX + x1) & 0x1FF;
                 }
             }
         }