diff --git a/Makefile b/Makefile
index e3dc527..a85d3a3 100644
--- a/Makefile
+++ b/Makefile
@@ -41,7 +41,7 @@ CPU_FLAGS = -I src -I /usr/local/include/SDL2 -I/usr/include/libpng16 -I/opt/pop
 IPU_FLAGS = -I src/ipu \
 			--target=ipu2 \
 			-Wall -Werror -Wno-unused-variable \
-			-O2
+			-O2 -g
 
 
 all: build build/doom build/ipu_rt.gp
diff --git a/src/ipu/am_map.c b/src/ipu/am_map.c
index d197952..5bc5ca3 100644
--- a/src/ipu/am_map.c
+++ b/src/ipu/am_map.c
@@ -241,6 +241,7 @@ static boolean stopped = true;
 //
 //
 //
+__SUPER__
 void AM_activateNewScale(void) {
   m_x += m_w / 2;
   m_y += m_h / 2;
@@ -255,6 +256,7 @@ void AM_activateNewScale(void) {
 //
 //
 //
+__SUPER__
 void AM_saveScaleAndLoc(void) {
   old_m_x = m_x;
   old_m_y = m_y;
@@ -265,6 +267,7 @@ void AM_saveScaleAndLoc(void) {
 //
 //
 //
+__SUPER__
 void AM_restoreScaleAndLoc(void) {
 
   m_w = old_m_w;
@@ -287,6 +290,7 @@ void AM_restoreScaleAndLoc(void) {
 //
 // adds a marker at the current location
 //
+__SUPER__
 void AM_addMark(void) {
   markpoints[markpointnum].x = m_x + m_w / 2;
   markpoints[markpointnum].y = m_y + m_h / 2;
@@ -297,6 +301,7 @@ void AM_addMark(void) {
 // Determines bounding box of all vertices,
 // sets global variables controlling zoom range.
 //
+__SUPER__
 void AM_findMinMaxBoundaries(void) {
   int i;
   fixed_t a;
@@ -333,6 +338,7 @@ void AM_findMinMaxBoundaries(void) {
 //
 //
 //
+__SUPER__
 void AM_changeWindowLoc(void) {
   if (m_paninc.x || m_paninc.y) {
     followplayer = 0;
@@ -359,6 +365,7 @@ void AM_changeWindowLoc(void) {
 //
 //
 //
+__SUPER__
 void AM_initVariables(void) {
   int pnum;
   // static event_t st_notify = {ev_keyup, AM_MSGENTERED, 0, 0}; // LATER
@@ -409,6 +416,7 @@ void AM_initVariables(void) {
 }
 
 
+__SUPER__
 void AM_clearMarks(void) {
   int i;
 
@@ -421,6 +429,7 @@ void AM_clearMarks(void) {
 // should be called at the start of every level
 // right now, i figure it out myself
 //
+__SUPER__
 void AM_LevelInit(void) {
   leveljuststarted = 0;
 
@@ -441,6 +450,7 @@ void AM_LevelInit(void) {
 //
 //
 //
+__SUPER__
 void AM_Stop(void) {
   // static event_t st_notify = {0, ev_keyup, AM_MSGEXITED, 0}; // LATER
 
@@ -453,6 +463,7 @@ void AM_Stop(void) {
 //
 //
 //
+__SUPER__
 void AM_Start(void) {
   static int lastlevel = -1, lastepisode = -1;
 
@@ -471,6 +482,7 @@ void AM_Start(void) {
 //
 // set the window scale to the maximum size
 //
+__SUPER__
 void AM_minOutWindowScale(void) {
   scale_mtof = min_scale_mtof;
   scale_ftom = FixedDiv(FRACUNIT, scale_mtof);
@@ -480,6 +492,7 @@ void AM_minOutWindowScale(void) {
 //
 // set the window scale to the minimum size
 //
+__SUPER__
 void AM_maxOutWindowScale(void) {
   scale_mtof = max_scale_mtof;
   scale_ftom = FixedDiv(FRACUNIT, scale_mtof);
@@ -489,6 +502,7 @@ void AM_maxOutWindowScale(void) {
 //
 // Handle events (user inputs) in automap mode
 //
+__SUPER__
 boolean AM_Responder(event_t *ev) {
 
   int rc;
@@ -637,6 +651,7 @@ boolean AM_Responder(event_t *ev) {
 //
 // Zooming
 //
+__SUPER__
 void AM_changeWindowScale(void) {
 
   // Change the scaling multipliers
@@ -654,6 +669,7 @@ void AM_changeWindowScale(void) {
 //
 //
 //
+__SUPER__
 void AM_doFollowPlayer(void) {
 
   if (f_oldloc.x != am_playerpos.x || f_oldloc.y != am_playerpos.y) {
@@ -674,6 +690,7 @@ void AM_doFollowPlayer(void) {
 //
 // Updates on Game Tick
 //
+__SUPER__
 void AM_Ticker(void) {
   if (!automapactive) 
     return;
@@ -698,6 +715,7 @@ void AM_Ticker(void) {
 //
 // Clear automap frame buffer.
 //
+__SUPER__
 void AM_clearFB(int color) { memset(fb, color, f_w * f_h * sizeof(*fb)); }
 
 //
@@ -707,6 +725,7 @@ void AM_clearFB(int color) { memset(fb, color, f_w * f_h * sizeof(*fb)); }
 // faster reject and precalculated slopes.  If the speed is needed,
 // use a hash algorithm to handle  the common cases.
 //
+__SUPER__
 boolean AM_clipMline(mline_t *ml, fline_t *fl) {
   enum { LEFT = 1, RIGHT = 2, BOTTOM = 4, TOP = 8 };
 
@@ -821,6 +840,7 @@ boolean AM_clipMline(mline_t *ml, fline_t *fl) {
 //
 // Classic Bresenham w/ whatever optimizations needed for speed
 //
+__SUPER__
 void AM_drawFline(fline_t *fl, int color) {
   register int x;
   register int y;
@@ -886,6 +906,7 @@ void AM_drawFline(fline_t *fl, int color) {
 //
 // Clip lines, draw visible part sof lines.
 //
+__SUPER__
 void AM_drawMline(mline_t *ml, int color) {
   static fline_t fl;
 
@@ -896,6 +917,7 @@ void AM_drawMline(mline_t *ml, int color) {
 //
 // Draws flat (floor/ceiling tile) aligned grid lines.
 //
+__SUPER__
 void AM_drawGrid(int color) {
   fixed_t x, y;
   fixed_t start, end;
@@ -938,6 +960,7 @@ void AM_drawGrid(int color) {
 // Determines visible lines, draws them.
 // This is LineDef based, not LineSeg based.
 //
+__SUPER__
 void AM_drawWalls(void) {
   int i;
   static mline_t l;
@@ -986,18 +1009,20 @@ void AM_drawWalls(void) {
 // Rotation in 2D.
 // Used to rotate player arrow line character.
 //
+__SUPER__
 void AM_rotate(fixed_t *x, fixed_t *y, angle_t a) {
   fixed_t tmpx;
 
   tmpx = FixedMul(*x, finecosine[a >> ANGLETOFINESHIFT]) -
-         FixedMul(*y, IPU_finesine(a >> ANGLETOFINESHIFT));
+         FixedMul(*y, finesine[a >> ANGLETOFINESHIFT]);
 
-  *y = FixedMul(*x, IPU_finesine(a >> ANGLETOFINESHIFT)) +
+  *y = FixedMul(*x, finesine[a >> ANGLETOFINESHIFT]) +
        FixedMul(*y, finecosine[a >> ANGLETOFINESHIFT]);
 
   *x = tmpx;
 }
 
+__SUPER__
 void AM_drawLineCharacter(mline_t *lineguy, int lineguylines, fixed_t scale,
                           angle_t angle, int color, fixed_t x, fixed_t y) {
   int i;
@@ -1036,6 +1061,7 @@ void AM_drawLineCharacter(mline_t *lineguy, int lineguylines, fixed_t scale,
   }
 }
 
+__SUPER__
 void AM_drawPlayers(void) {
   int i;
   player_t *p;
@@ -1074,6 +1100,7 @@ void AM_drawPlayers(void) {
   */
 }
 
+__SUPER__
 void AM_drawMarks(void) {
   int i, fx, fy, w, h;
 
@@ -1092,10 +1119,12 @@ void AM_drawMarks(void) {
   }
 }
 
+__SUPER__
 void AM_drawCrosshair(int color) {
   fb[(f_w * (f_h + 1)) / 2] = color; // single point for now
 }
 
+__SUPER__
 void AM_drawIPUWatermark(int color) {
   const int wmx = 283, wmy = 160;
   for (int i = 0; i < arrlen(ipu_watermark_lines); ++i) {
@@ -1106,6 +1135,7 @@ void AM_drawIPUWatermark(int color) {
   }
 }
 
+__SUPER__
 void AM_Drawer(pixel_t* fb_tensor) {
   fb = fb_tensor; // JOSEF
   V_UseBuffer(fb);
diff --git a/src/ipu/am_map.h b/src/ipu/am_map.h
index 665bd5b..5376525 100644
--- a/src/ipu/am_map.h
+++ b/src/ipu/am_map.h
@@ -22,6 +22,7 @@
 #include "d_event.h"
 #include "doomtype.h"
 #include "m_cheat.h"
+#include "ipu_utils.h"
 
 // Used by ST StatusBar stuff.
 #define AM_MSGHEADER (('a' << 24) + ('m' << 16))
@@ -29,18 +30,18 @@
 #define AM_MSGEXITED (AM_MSGHEADER | ('x' << 8))
 
 // Called by main loop.
-boolean AM_Responder(event_t *ev);
+__SUPER__ boolean AM_Responder(event_t *ev);
 
 // Called by main loop.
-void AM_Ticker(void);
+__SUPER__ void AM_Ticker(void);
 
 // Called by main loop,
 // called instead of view drawer if automap active.
-void AM_Drawer(void);
+__SUPER__ void AM_Drawer(void);
 
 // Called to force the automap to quit
 // if the level is completed while it is up.
-void AM_Stop(void);
+__SUPER__ void AM_Stop(void);
 
 extern cheatseq_t cheat_amap;
 
diff --git a/src/ipu/g_game.c b/src/ipu/g_game.c
index c3c18f7..1615ac8 100644
--- a/src/ipu/g_game.c
+++ b/src/ipu/g_game.c
@@ -69,6 +69,7 @@ int testcontrols_mousespeed;
 //
 // G_DoLoadLevel
 //
+__SUPER__ 
 void G_DoLoadLevel(void) {
   int i;
 
@@ -136,6 +137,7 @@ void G_DoLoadLevel(void) {
 // G_Ticker
 // Make ticcmd_ts for the players.
 //
+__SUPER__ 
 void G_Ticker(void) {
   int i;
   int buf;
@@ -295,6 +297,7 @@ void G_Ticker(void) {
 // G_Responder
 // Get info needed to make ticcmd_ts for the players.
 //
+__SUPER__ 
 boolean G_Responder(event_t *ev) {
 
   /* LATER
diff --git a/src/ipu/g_game.h b/src/ipu/g_game.h
index 528779d..ed48bf0 100644
--- a/src/ipu/g_game.h
+++ b/src/ipu/g_game.h
@@ -24,8 +24,11 @@
 #include "d_ticcmd.h"
 #include "doomtype.h"
 
+#include "ipu_utils.h"
+
 
 // JOSEF TMP
+__SUPER__ 
 void G_DoLoadLevel(void);
 
 //
@@ -69,8 +72,8 @@ void G_WorldDone(void);
 
 void G_BuildTiccmd(ticcmd_t *cmd, int maketic);
 
-void G_Ticker(void);
-boolean G_Responder(event_t *ev);
+__SUPER__ void G_Ticker(void);
+__SUPER__ boolean G_Responder(event_t *ev);
 
 void G_ScreenShot(void);
 
diff --git a/src/ipu/g_game_codelets.cpp b/src/ipu/g_game_codelets.cpp
index 0f51805..4102f53 100644
--- a/src/ipu/g_game_codelets.cpp
+++ b/src/ipu/g_game_codelets.cpp
@@ -4,16 +4,16 @@
 #include "ipu_transfer.h"
 
 extern "C" {
-    void G_DoLoadLevel(void);
-    void G_Ticker(void);
-    boolean G_Responder(event_t *ev);
+    __SUPER__ void G_DoLoadLevel(void);
+    __SUPER__ void G_Ticker(void);
+    __SUPER__ boolean G_Responder(event_t *ev);
 };
 
 
-class G_DoLoadLevel_Vertex : public poplar::Vertex {
+class G_DoLoadLevel_Vertex : public poplar::SupervisorVertex {
  public:
     poplar::Input<poplar::Vector<unsigned char>> miscValues;
-  bool compute() {
+  __SUPER__ bool compute() {
     IPU_G_LoadLevel_UnpackMiscValues((G_LoadLevel_MiscValues_t*)&miscValues[0]);
     G_DoLoadLevel();
     return true;
@@ -21,20 +21,20 @@ class G_DoLoadLevel_Vertex : public poplar::Vertex {
 };
 
 
-class G_Ticker_Vertex : public poplar::Vertex {
+class G_Ticker_Vertex : public poplar::SupervisorVertex {
  public:
   poplar::Input<poplar::Vector<unsigned char>> miscValues;
-  bool compute() {
+  __SUPER__ bool compute() {
     IPU_G_Ticker_UnpackMiscValues((G_Ticker_MiscValues_t*)&miscValues[0]);
     G_Ticker();
     return true;
   }
 };
 
-class G_Responder_Vertex : public poplar::Vertex {
+class G_Responder_Vertex : public poplar::SupervisorVertex {
  public:
   poplar::Input<poplar::Vector<unsigned char>> miscValues;
-  bool compute() {
+  __SUPER__ bool compute() {
     // Respond to each of the buffered events in turn
     G_Responder_MiscValues_t* ev_buf = (G_Responder_MiscValues_t*)&miscValues[0];
     for (int i = 0; i < ev_buf->num_ev; ++i) {
diff --git a/src/ipu/ipu_malloc.c b/src/ipu/ipu_malloc.c
index 9fb36c0..8d70ece 100644
--- a/src/ipu/ipu_malloc.c
+++ b/src/ipu/ipu_malloc.c
@@ -21,7 +21,7 @@ static unsigned char PU_STATIC_pool[PU_STATIC_max_size];
 static unsigned char PU_LEVEL_pool[PU_LEVEL_max_size];
 static unsigned char PU_TMP_pool[PU_TMP_max_size];
 
-
+__SUPER__
 void* IPU_level_malloc(int size, const char* name) {
     void* ret = (void*)(&PU_LEVEL_pool[PU_LEVEL_size]);
     PU_LEVEL_size = ALIGN32(PU_LEVEL_size + size);
@@ -39,10 +39,12 @@ void* IPU_level_malloc(int size, const char* name) {
     return ret;
 }
 
+__SUPER__
 void IPU_level_free() {
     PU_LEVEL_size = 0;
 }
 
+__SUPER__
 void* IPU_static_malloc(int size, const char* name) {
     void* ret = (void*)(&PU_STATIC_pool[PU_STATIC_size]);
     PU_STATIC_size = ALIGN32(PU_STATIC_size + size);
@@ -62,6 +64,7 @@ void* IPU_static_malloc(int size, const char* name) {
 
 
 static unsigned PU_TMP_count = 0;
+__SUPER__
 void* IPU_tmp_malloc(int size, const char* name) {
     void* ret = (void*)(&PU_TMP_pool[PU_TMP_size]);
     PU_TMP_size = ALIGN32(PU_TMP_size + size);
@@ -81,6 +84,7 @@ void* IPU_tmp_malloc(int size, const char* name) {
 }
 
 
+__SUPER__
 void IPU_tmp_free(void* ptr) {
     (void) ptr;
     PU_TMP_count -= 1;
@@ -90,6 +94,7 @@ void IPU_tmp_free(void* ptr) {
 }
 
 
+__SUPER__
 void IPU_summarise_malloc() {
     if (tileID != 0) 
         return;
diff --git a/src/ipu/ipu_malloc.h b/src/ipu/ipu_malloc.h
index 663c704..9ff2695 100644
--- a/src/ipu/ipu_malloc.h
+++ b/src/ipu/ipu_malloc.h
@@ -4,17 +4,19 @@
 extern "C" {
 #endif
 
+#include "ipu_utils.h"
 
-void* IPU_static_malloc(int size, const char* name);
+
+__SUPER__ void* IPU_static_malloc(int size, const char* name);
 // There is no IPU_static_free()  :D
 
-void* IPU_level_malloc(int size, const char* name);
-void IPU_level_free(void);
+__SUPER__ void* IPU_level_malloc(int size, const char* name);
+__SUPER__ void IPU_level_free(void);
 
-void* IPU_tmp_malloc(int size, const char* name);
-void IPU_tmp_free(void* ptr);
+__SUPER__ void* IPU_tmp_malloc(int size, const char* name);
+__SUPER__ void IPU_tmp_free(void* ptr);
 
-void IPU_summarise_malloc(void);
+__SUPER__ void IPU_summarise_malloc(void);
 
 
 #ifdef __cplusplus
diff --git a/src/ipu/ipu_texturetiles.cpp b/src/ipu/ipu_texturetiles.cpp
index 7d9287a..106a88b 100644
--- a/src/ipu/ipu_texturetiles.cpp
+++ b/src/ipu/ipu_texturetiles.cpp
@@ -3,7 +3,6 @@
 
 #include "ipu_utils.h"
 #include "ipu_texturetiles.h"
-// #include "../../xcom.hpp"
 
 
 #define NUMCACHECOLS (20)
@@ -11,8 +10,7 @@
 static byte columnCache[NUMCACHECOLS][CACHECOLSIZE];
 
 
-extern "C" 
-void IPU_R_InitColumnRequester(void) {
+extern "C" __SUPER__ void IPU_R_InitColumnRequester(void) {
     // TMP colours
     for (int i = 0; i < NUMCACHECOLS; ++i) {
         unsigned* col = (unsigned*) columnCache[i];
@@ -28,6 +26,6 @@ void IPU_R_InitColumnRequester(void) {
 }
 
 
-extern "C" byte* IPU_R_RequestColumn(int texture, int column) {
+extern "C" __SUPER__ byte* IPU_R_RequestColumn(int texture, int column) {
     return columnCache[texture % NUMCACHECOLS];
 }
\ No newline at end of file
diff --git a/src/ipu/ipu_texturetiles.h b/src/ipu/ipu_texturetiles.h
index e88dc94..1554824 100644
--- a/src/ipu/ipu_texturetiles.h
+++ b/src/ipu/ipu_texturetiles.h
@@ -12,9 +12,9 @@ extern "C" {
 
 
 
-void IPU_R_InitColumnRequester(void);
+__SUPER__ void IPU_R_InitColumnRequester(void);
 
-byte* IPU_R_RequestColumn(int texture, int column);
+__SUPER__ byte* IPU_R_RequestColumn(int texture, int column);
 
 
 
diff --git a/src/ipu/ipu_transfer.c b/src/ipu/ipu_transfer.c
index 24db13b..97b493b 100644
--- a/src/ipu/ipu_transfer.c
+++ b/src/ipu/ipu_transfer.c
@@ -14,6 +14,7 @@ extern int setdetail; // r_main.c
 
 int gamelumpnum;
 
+__SUPER__
 void IPU_G_LoadLevel_UnpackMiscValues(G_LoadLevel_MiscValues_t* pack) {
   gameepisode = pack->gameepisode;
   gamemap = pack->gamemap;
@@ -25,6 +26,7 @@ void IPU_G_LoadLevel_UnpackMiscValues(G_LoadLevel_MiscValues_t* pack) {
   }
 }
 
+__SUPER__
 void IPU_G_Ticker_UnpackMiscValues(G_Ticker_MiscValues_t* pack) {
   gamestate = pack->gamestate;
   if (gamestate != GS_LEVEL)
@@ -52,6 +54,7 @@ void IPU_G_Ticker_UnpackMiscValues(G_Ticker_MiscValues_t* pack) {
 
 }
 
+__SUPER__
 void IPU_Setup_UnpackMarkNums(const unsigned char* buf) {
   short* offsets = (short*) buf;
   const int offsetssize = 10 * sizeof(short);
@@ -62,11 +65,12 @@ void IPU_Setup_UnpackMarkNums(const unsigned char* buf) {
 }
 
 
+__SUPER__ 
 void IPU_R_RenderPlayerView_UnpackMiscValues(R_RenderPlayerView_MiscValues_t* pack) {
   // Nothing to unpack
 }
 
-
+__SUPER__
 void IPU_R_ExecuteSetViewSize_UnpackMiscValues(R_ExecuteSetViewSize_MiscValues_t* pack) {
   setblocks = pack->setblocks;
   setdetail = pack->setdetail;
diff --git a/src/ipu/ipu_transfer.h b/src/ipu/ipu_transfer.h
index 4a948ad..4b0204f 100644
--- a/src/ipu/ipu_transfer.h
+++ b/src/ipu/ipu_transfer.h
@@ -5,6 +5,7 @@ extern "C" {
 #endif
 
 #include "v_patch.h"
+#include "ipu_utils.h"
 #include "ipu_interface.h"
 
 
@@ -13,10 +14,10 @@ extern IPUTransfer_playerstate_t am_playerpos;
 extern patch_t* marknums[10];
 extern unsigned char markbuf[IPUAMMARKBUFSIZE];
 
-void IPU_G_LoadLevel_UnpackMiscValues(G_LoadLevel_MiscValues_t* pack);
-void IPU_G_Ticker_UnpackMiscValues(G_Ticker_MiscValues_t* pack);
-void IPU_R_RenderPlayerView_UnpackMiscValues(R_RenderPlayerView_MiscValues_t* pack);
-void IPU_R_ExecuteSetViewSize_UnpackMiscValues(R_ExecuteSetViewSize_MiscValues_t* pack);
+__SUPER__ void IPU_G_LoadLevel_UnpackMiscValues(G_LoadLevel_MiscValues_t* pack);
+__SUPER__ void IPU_G_Ticker_UnpackMiscValues(G_Ticker_MiscValues_t* pack);
+__SUPER__ void IPU_R_RenderPlayerView_UnpackMiscValues(R_RenderPlayerView_MiscValues_t* pack);
+__SUPER__ void IPU_R_ExecuteSetViewSize_UnpackMiscValues(R_ExecuteSetViewSize_MiscValues_t* pack);
 
 #ifdef __cplusplus
 }
diff --git a/src/ipu/ipu_utils.h b/src/ipu/ipu_utils.h
index 9fd8a27..6cef057 100644
--- a/src/ipu/ipu_utils.h
+++ b/src/ipu/ipu_utils.h
@@ -1,7 +1,12 @@
 #ifndef __IPU_UTILS__
 #define __IPU_UTILS__
 
-// #define __SUPER__ __attribute__((target("supervisor")))
+
+#ifdef __IPU__
+#define __SUPER__ __attribute__((target("supervisor")))
+#else
+#define __SUPER__
+#endif
 
 extern int tileID;
 
diff --git a/src/ipu/ipu_vertices.cpp b/src/ipu/ipu_vertices.cpp
index e141647..81e3743 100644
--- a/src/ipu/ipu_vertices.cpp
+++ b/src/ipu/ipu_vertices.cpp
@@ -13,29 +13,28 @@ typedef uint8_t pixel_t;
 
 
 extern "C" {
-    void AM_LevelInit(void);
-    void AM_Drawer(pixel_t*);
-    void IPU_Setup_UnpackMarkNums(const unsigned char* buf);
+  __SUPER__ void AM_LevelInit(void);
+  __SUPER__ void AM_Drawer(pixel_t*);
+  __SUPER__ void IPU_Setup_UnpackMarkNums(const unsigned char* buf);
 };
 
 
-class AM_LevelInit_Vertex : public poplar::Vertex {
+class AM_LevelInit_Vertex : public poplar::SupervisorVertex {
  public:
-  bool compute() {
+  __SUPER__ bool compute() {
     AM_LevelInit(); 
     return true;
   }
 };
 
 
-class AM_Drawer_Vertex : public poplar::Vertex {
+class AM_Drawer_Vertex : public poplar::SupervisorVertex {
  public:
   poplar::InOut<poplar::Vector<unsigned char>> frame;
 
-  bool compute() {
+  __SUPER__ void compute() {
     assert(&frame[0] == I_VideoBuffer);
     AM_Drawer(&frame[0]);
-    return true;
   }
 };
 
@@ -44,8 +43,7 @@ class AM_Drawer_Vertex : public poplar::Vertex {
 
 struct IPU_Init_Vertex : public poplar::SupervisorVertex {
 
-  __attribute__((target("supervisor")))
-  void compute() {
+  __SUPER__ void compute() {
 
     // Deduce logical tile ID
     int physical = __builtin_ipu_get_tile_id();
@@ -60,9 +58,6 @@ struct IPU_Init_Vertex : public poplar::SupervisorVertex {
     logical += 2 * row;
     tileID = logical;
 
-    
-    // IPU_R_InitColumnRequester(physical);
-
   }
 };
 
@@ -72,18 +67,17 @@ struct IPU_Init_Vertex : public poplar::SupervisorVertex {
 struct IPU_MiscSetup_Vertex : public poplar::SupervisorVertex {
   poplar::InOut<poplar::Vector<unsigned char>> frame;
 
-  __attribute__((target("supervisor")))
-  void compute() {
+  __SUPER__ void compute() {
 
     I_VideoBuffer = &frame[0];
 
   }
 };
 
-struct IPU_UnpackMarknumSprites_Vertex : public poplar::Vertex {
+struct IPU_UnpackMarknumSprites_Vertex : public poplar::SupervisorVertex {
   poplar::Input<poplar::Vector<unsigned char>> buf;
 
-  void compute() {
+  __SUPER__ void compute() {
     IPU_Setup_UnpackMarkNums(&buf[0]);
     return;
   }
diff --git a/src/ipu/m_bbox.c b/src/ipu/m_bbox.c
index 889b70c..7088a7c 100644
--- a/src/ipu/m_bbox.c
+++ b/src/ipu/m_bbox.c
@@ -26,7 +26,7 @@
 #include <limits.h>
 
 
-
+__SUPER__ 
 void M_ClearBox (fixed_t *box)
 {
     box[BOXTOP] = box[BOXRIGHT] = INT_MIN;
@@ -34,6 +34,7 @@ void M_ClearBox (fixed_t *box)
 }
 
 
+__SUPER__ 
 void
 M_AddToBox
 ( fixed_t*	box,
diff --git a/src/ipu/m_bbox.h b/src/ipu/m_bbox.h
index 958e21f..b1bbc65 100644
--- a/src/ipu/m_bbox.h
+++ b/src/ipu/m_bbox.h
@@ -33,13 +33,8 @@ enum
 };	// bbox coordinates
 
 // Bounding box functions.
-void M_ClearBox (fixed_t*	box);
-
-void
-M_AddToBox
-( fixed_t*	box,
-  fixed_t	x,
-  fixed_t	y );
+__SUPER__ void M_ClearBox (fixed_t*	box);
+__SUPER__ void M_AddToBox(fixed_t* box, fixed_t x, fixed_t y);
 
 
 #endif
diff --git a/src/ipu/m_fixed.c b/src/ipu/m_fixed.cpp
similarity index 50%
rename from src/ipu/m_fixed.c
rename to src/ipu/m_fixed.cpp
index 6d37d17..4ee799d 100644
--- a/src/ipu/m_fixed.c
+++ b/src/ipu/m_fixed.cpp
@@ -18,23 +18,28 @@
 
 
 
+extern "C" {
 #include <limits.h>
 #include <stdint.h>
 
 #include "m_fixed.h"
+}
+#include <poplar/Vertex.hpp>
+
+#include "print.h"
 
 
-
+extern "C"
+__SUPER__ 
 int abs(int x) {
     return (x < 0) ? -x : x;
 }
 
 // Fixme. __USE_C_FIXED__ or something.
 
-fixed_t
-FixedMul
-( fixed_t	a,
-  fixed_t	b )
+extern "C"
+__SUPER__ 
+fixed_t FixedMul(fixed_t a, fixed_t b)
 {
     return ((int64_t) a * (int64_t) b) >> FRACBITS;
     // JOSEF: What about this? fewer inst.
@@ -47,23 +52,43 @@ FixedMul
 
 
 
+struct DivWorker : public poplar::Vertex {
+    fixed_t a, b;
+    void compute() {
+        // Using int64_ caused some weird errors. So use double instead...?
+        // a = (((int64_t) a) << FRACBITS) / b;
+        a = ((double) a * FRACUNIT) / (double) b; 
+    }
+};
+
 //
 // FixedDiv, C version.
 //
-
+extern "C"
+__SUPER__ 
 fixed_t FixedDiv(fixed_t a, fixed_t b)
 {
-    if ((abs(a) >> 14) >= abs(b))
-    {
-	return (a^b) < 0 ? INT_MIN : INT_MAX;
-    }
-    else
-    {
-	int64_t result;
+    if ((abs(a) >> 14) >= abs(b)) {
+    	return (a^b) < 0 ? INT_MIN : INT_MAX;
+    } else {
+        // int64_t result; // JOSEF
+        // result = ((int64_t) a << FRACBITS) / b;
+        // return (fixed_t) result;
 
-	result = ((int64_t) a << FRACBITS) / b;
-
-	return (fixed_t) result;
+        DivWorker workerArgs;
+        unsigned workerArgsPtr = (unsigned)&workerArgs - TMEM_BASE_ADDR;
+        workerArgs.a = a;
+        workerArgs.b = b;
+        void* workerFn;
+        asm volatile("setzi %0, __runCodelet_DivWorker" : [workerFn] "+r"(workerFn));
+        asm volatile(
+            "run %[workerFn], %[workerArgs], 0\n"
+            "sync %[zone]\n"
+            : [workerFn] "+r"(workerFn), 
+              [workerArgs] "+r"(workerArgsPtr)
+            : [zone] "i"(TEXCH_SYNCZONE_LOCAL)
+        );
+        return workerArgs.a;
     }
 }
 
diff --git a/src/ipu/m_fixed.h b/src/ipu/m_fixed.h
index 3a66c16..e152332 100644
--- a/src/ipu/m_fixed.h
+++ b/src/ipu/m_fixed.h
@@ -20,7 +20,7 @@
 #ifndef __M_FIXED__
 #define __M_FIXED__
 
-
+#include "ipu_utils.h"
 
 
 //
@@ -32,8 +32,8 @@
 typedef int fixed_t;
 
 
-fixed_t FixedMul	(fixed_t a, fixed_t b);
-fixed_t FixedDiv	(fixed_t a, fixed_t b);
+__SUPER__ fixed_t FixedMul	(fixed_t a, fixed_t b);
+__SUPER__ fixed_t FixedDiv	(fixed_t a, fixed_t b);
 
 
 
diff --git a/src/ipu/p_mobj.c b/src/ipu/p_mobj.c
index dd92fc8..257f796 100644
--- a/src/ipu/p_mobj.c
+++ b/src/ipu/p_mobj.c
@@ -448,6 +448,7 @@ void P_MobjThinker(mobj_t *mobj) {
 //
 // P_SpawnMobj
 //
+__SUPER__ 
 mobj_t *P_SpawnMobj(fixed_t x, fixed_t y, fixed_t z, mobjtype_t type) {
   mobj_t *mobj;
   state_t *st;
@@ -598,6 +599,7 @@ void P_RespawnSpecials(void) {
 // Most of the player structure stays unchanged
 //  between levels.
 //
+__SUPER__ 
 void P_SpawnPlayer(mapthing_t *mthing) {
   player_t *p;
   fixed_t x;
@@ -668,6 +670,7 @@ void P_SpawnPlayer(mapthing_t *mthing) {
 // The fields of the mapthing should
 // already be in host byte order.
 //
+__SUPER__ 
 void P_SpawnMapThing(mapthing_t *mthing) {
   int i;
   int bit;
diff --git a/src/ipu/p_setup.c b/src/ipu/p_setup.c
index daf175b..a0c1f8f 100644
--- a/src/ipu/p_setup.c
+++ b/src/ipu/p_setup.c
@@ -12,7 +12,7 @@
 #include "print.h"
 
 
-void P_SpawnMapThing(mapthing_t *mthing);
+__SUPER__ void P_SpawnMapThing(mapthing_t *mthing);
 
 //
 // MAP related Lookup tables.
@@ -80,6 +80,7 @@ mapthing_t playerstarts[MAXPLAYERS];
 //
 // P_LoadVertexes
 //
+__SUPER__ 
 void P_LoadVertexes(const unsigned char *buf) {
   byte *data;
   int i;
@@ -130,6 +131,7 @@ sector_t *GetSectorAtNullAddress(void) {
 //
 // P_LoadSegs
 //
+__SUPER__ 
 void P_LoadSegs(const unsigned char *buf) {
   byte *data;
   int i;
@@ -195,6 +197,7 @@ void P_LoadSegs(const unsigned char *buf) {
 //
 // P_LoadSectors
 //
+__SUPER__ 
 void P_LoadSectors(const unsigned char *buf) {
   byte *data;
   int i;
@@ -224,6 +227,7 @@ void P_LoadSectors(const unsigned char *buf) {
 //
 // P_LoadNodes
 //
+__SUPER__ 
 __attribute__((optnone))  // Workaround for https://graphcore.atlassian.net/browse/LLVM-330
 void P_LoadNodes(const unsigned char *buf) {
   int i;
@@ -258,6 +262,7 @@ void P_LoadNodes(const unsigned char *buf) {
 //
 // P_LoadThings
 //
+__SUPER__ 
 void P_LoadThings(const unsigned char *buf) {
   byte *data;
   int i;
@@ -308,6 +313,7 @@ void P_LoadThings(const unsigned char *buf) {
 //
 // P_LoadSideDefs
 //
+__SUPER__ 
 void P_LoadSideDefs(const unsigned char *buf) {
   byte *data;
   int i;
@@ -335,6 +341,7 @@ void P_LoadSideDefs(const unsigned char *buf) {
 // P_LoadLineDefs
 // Also counts secret lines for intermissions.
 //
+__SUPER__ 
 void P_LoadLineDefs(const unsigned char *buf) {
   byte *data;
   int i;
@@ -404,6 +411,7 @@ void P_LoadLineDefs(const unsigned char *buf) {
 //
 // P_LoadSubsectors
 //
+__SUPER__ 
 void P_LoadSubsectors(const unsigned char *buf) {
   byte *data;
   int i;
@@ -428,6 +436,7 @@ void P_LoadSubsectors(const unsigned char *buf) {
 //
 // P_SetupLevel
 //
+__SUPER__ 
 void P_SetupLevel_pt0(const unsigned char unused) {
   int i;
 
@@ -521,6 +530,7 @@ void P_SetupLevel_pt0(const unsigned char unused) {
 //
 // P_LoadBlockMap
 //
+__SUPER__ 
 void P_LoadBlockMap(const unsigned char *buf) {
   int i;
   int count;
@@ -553,6 +563,7 @@ void P_LoadBlockMap(const unsigned char *buf) {
 // Builds sector line lists and subsector sector numbers.
 // Finds block bounding boxes for sectors.
 //
+__SUPER__ 
 void P_GroupLines(const unsigned char *buf) {
   (void) buf;
 
diff --git a/src/ipu/p_setup.h b/src/ipu/p_setup.h
index a30febc..7d61ecc 100644
--- a/src/ipu/p_setup.h
+++ b/src/ipu/p_setup.h
@@ -21,9 +21,9 @@
 
 
 // NOT called by W_Ticker. Fixme.
-void P_SetupLevel();
+__SUPER__ void P_SetupLevel();
 
 // Called by startup code.
-void P_Init(void);
+__SUPER__ void P_Init(void);
 
 #endif
diff --git a/src/ipu/p_setup_codelets.cpp b/src/ipu/p_setup_codelets.cpp
index 5407d5f..82261b4 100644
--- a/src/ipu/p_setup_codelets.cpp
+++ b/src/ipu/p_setup_codelets.cpp
@@ -10,17 +10,17 @@
 
 
 extern "C" {
-    void P_SetupLevel_pt0(const unsigned char *buf);
-    void P_LoadBlockMap(const unsigned char *buf);
-    void P_LoadVertexes(const unsigned char *buf);
-    void P_LoadSectors(const unsigned char *buf);
-    void P_LoadSideDefs(const unsigned char *buf);
-    void P_LoadLineDefs(const unsigned char *buf);
-    void P_LoadSubsectors(const unsigned char *buf);
-    void P_LoadNodes(const unsigned char *buf);
-    void P_LoadSegs(const unsigned char *buf);
-    void P_LoadThings(const unsigned char *buf);
-    void P_GroupLines(const unsigned char *buf);
+    __SUPER__ void P_SetupLevel_pt0(const unsigned char *buf);
+    __SUPER__ void P_LoadBlockMap(const unsigned char *buf);
+    __SUPER__ void P_LoadVertexes(const unsigned char *buf);
+    __SUPER__ void P_LoadSectors(const unsigned char *buf);
+    __SUPER__ void P_LoadSideDefs(const unsigned char *buf);
+    __SUPER__ void P_LoadLineDefs(const unsigned char *buf);
+    __SUPER__ void P_LoadSubsectors(const unsigned char *buf);
+    __SUPER__ void P_LoadNodes(const unsigned char *buf);
+    __SUPER__ void P_LoadSegs(const unsigned char *buf);
+    __SUPER__ void P_LoadThings(const unsigned char *buf);
+    __SUPER__ void P_GroupLines(const unsigned char *buf);
 };
 
 
@@ -28,11 +28,11 @@ extern "C" {
 // --------------- P_Setup ----------------- //
 
 
-class P_SetupLevel_Vertex : public poplar::Vertex {
+class P_SetupLevel_Vertex : public poplar::SupervisorVertex {
   poplar::Input<poplar::Vector<unsigned char>> lumpBuf;
   poplar::Output<int> lumpNum;
  public:
-  bool compute() {
+  __SUPER__ bool compute() {
     static int step = 0;
     int next;
     // Switch statements are bonkers
diff --git a/src/ipu/p_tick.c b/src/ipu/p_tick.c
index 18220b3..134be09 100644
--- a/src/ipu/p_tick.c
+++ b/src/ipu/p_tick.c
@@ -42,12 +42,13 @@ thinker_t thinkercap;
 //
 // P_InitThinkers
 //
-void P_InitThinkers(void) { thinkercap.prev = thinkercap.next = &thinkercap; }
+__SUPER__ void P_InitThinkers(void) { thinkercap.prev = thinkercap.next = &thinkercap; }
 
 //
 // P_AddThinker
 // Adds a new thinker at the end of the list.
 //
+__SUPER__ 
 void P_AddThinker(thinker_t *thinker) {
   thinkercap.prev->next = thinker;
   thinker->next = &thinkercap;
@@ -60,6 +61,7 @@ void P_AddThinker(thinker_t *thinker) {
 // Deallocation is lazy -- it will not actually be freed
 // until its thinking turn comes up.
 //
+__SUPER__ 
 void P_RemoveThinker(thinker_t *thinker) {
   // FIXME: NOP.
   thinker->function.acv = (actionf_v)(-1);
@@ -69,6 +71,7 @@ void P_RemoveThinker(thinker_t *thinker) {
 // P_AllocateThinker
 // Allocates memory and adds a new thinker at the end of the list.
 //
+__SUPER__ 
 void P_AllocateThinker(thinker_t *thinker) {}
 
 //
diff --git a/src/ipu/p_tick.h b/src/ipu/p_tick.h
index 804d86e..14e0b46 100644
--- a/src/ipu/p_tick.h
+++ b/src/ipu/p_tick.h
@@ -22,6 +22,6 @@
 // Called by C_Ticker,
 // can call G_PlayerExited.
 // Carries out all thinking of monsters and players.
-void P_Ticker(void);
+__SUPER__ void P_Ticker(void);
 
 #endif
diff --git a/src/ipu/r_bsp.c b/src/ipu/r_bsp.c
index 001d6f6..67ec3db 100644
--- a/src/ipu/r_bsp.c
+++ b/src/ipu/r_bsp.c
@@ -47,11 +47,13 @@ sector_t *backsector;
 drawseg_t drawsegs[MAXDRAWSEGS];
 drawseg_t *ds_p;
 
+__SUPER__
 void R_StoreWallRange(int start, int stop);
 
 //
 // R_ClearDrawSegs
 //
+__SUPER__
 void R_ClearDrawSegs(void) { ds_p = drawsegs; }
 
 //
@@ -90,6 +92,7 @@ int bspnum;
 //  e.g. single sided LineDefs (middle texture)
 //  that entirely block the view.
 //
+__SUPER__
 void R_ClipSolidWallSegment(int first, int last) {
   cliprange_t *next;
   cliprange_t *start;
@@ -169,6 +172,7 @@ crunch:
 // Does handle windows,
 //  e.g. LineDefs with upper and lower texture.
 //
+__SUPER__
 void R_ClipPassWallSegment(int first, int last) {
   cliprange_t *start;
 
@@ -209,6 +213,7 @@ void R_ClipPassWallSegment(int first, int last) {
 //
 // R_ClearClipSegs
 //
+__SUPER__ 
 void R_ClearClipSegs(void) {
   solidsegs[0].first = -0x7fffffff;
   // solidsegs[0].last = -1;            // JOSEF
@@ -224,6 +229,7 @@ void R_ClearClipSegs(void) {
 // Clips the given segment
 // and adds any visible pieces to the line list.
 //
+__SUPER__ 
 void R_AddLine(seg_t *line) {
   int x1;
   int x2;
@@ -336,6 +342,7 @@ int checkcoord[12][4] = {{3, 0, 2, 1}, {3, 0, 2, 0}, {3, 1, 2, 0}, {0},
                          {2, 0, 2, 1}, {0, 0, 0, 0}, {3, 1, 3, 0}, {0},
                          {2, 0, 3, 1}, {2, 1, 3, 1}, {2, 1, 3, 0}};
 
+__SUPER__
 boolean R_CheckBBox(fixed_t *bspcoord) {
   int boxx;
   int boxy;
@@ -444,6 +451,7 @@ boolean R_CheckBBox(fixed_t *bspcoord) {
 // Add sprites of things in sector.
 // Draw one or more line segments.
 //
+__SUPER__
 void R_Subsector(int num) {
   int count;
   seg_t *line;
@@ -526,6 +534,7 @@ void R_RenderBSPNode(int bspnum) {
 //
 // R_RenderBSPNodeNonRecursive
 // Non-recursive version is more suitable to IPU software stack
+__SUPER__
 void R_RenderBSPNodeNonRecursive() {
   node_t *bsp;
   int side;
diff --git a/src/ipu/r_bsp.h b/src/ipu/r_bsp.h
index 735d351..6259f00 100644
--- a/src/ipu/r_bsp.h
+++ b/src/ipu/r_bsp.h
@@ -48,10 +48,10 @@ extern lighttable_t **dscalelight;
 typedef void (*drawfunc_t)(int start, int stop);
 
 // BSP?
-void R_ClearClipSegs(void);
-void R_ClearDrawSegs(void);
+__SUPER__ void R_ClearClipSegs(void);
+__SUPER__ void R_ClearDrawSegs(void);
 
-void R_RenderBSPNode(int bspnum);
-void R_RenderBSPNodeNonRecursive(void); // Nonrecursive verion for IPU
+__SUPER__ void R_RenderBSPNode(int bspnum);
+__SUPER__ void R_RenderBSPNodeNonRecursive(void); // Nonrecursive verion for IPU
 
 #endif
diff --git a/src/ipu/r_codelets.cpp b/src/ipu/r_codelets.cpp
index f964936..18c4d80 100644
--- a/src/ipu/r_codelets.cpp
+++ b/src/ipu/r_codelets.cpp
@@ -12,18 +12,19 @@
 
 
 extern "C" {
-  void R_InitTextures(int* maptex, R_Init_MiscValues_t* miscVals);
-  void R_RenderPlayerView(player_t *player);
-  void R_ExecuteSetViewSize(void);
+  __SUPER__ void R_InitTextures(int* maptex, R_Init_MiscValues_t* miscVals);
+  __SUPER__ void R_RenderPlayerView(player_t *player);
+  __SUPER__ void R_ExecuteSetViewSize(void);
 };
 
 
-struct R_Init_Vertex: public poplar::Vertex {
+struct R_Init_Vertex: public poplar::SupervisorVertex {
   
   poplar::Input<poplar::Vector<unsigned char>> miscValues;
   poplar::Input<poplar::Vector<unsigned char>> lumpBuf;
   poplar::Output<int> lumpNum;
 
+  __SUPER__ 
   void compute() {
     static int step = 0;
 
@@ -43,11 +44,11 @@ struct R_Init_Vertex: public poplar::Vertex {
 };
 
 
-struct R_RenderPlayerView_Vertex : public poplar::Vertex {
+struct R_RenderPlayerView_Vertex : public poplar::SupervisorVertex {
   poplar::Input<poplar::Vector<unsigned char>> miscValues;
   poplar::InOut<poplar::Vector<unsigned char>> frame;
 
-  void compute() {
+  __SUPER__ void compute() {
     assert(&frame[0] == I_VideoBuffer);
 
     IPU_R_RenderPlayerView_UnpackMiscValues(
@@ -60,14 +61,14 @@ struct R_RenderPlayerView_Vertex : public poplar::Vertex {
 };
 
 
-class R_ExecuteSetViewSize_Vertex : public poplar::Vertex {
+class R_ExecuteSetViewSize_Vertex : public poplar::SupervisorVertex {
  public:
   poplar::Input<poplar::Vector<unsigned char>> miscValues;
-  bool compute() {
+  __SUPER__ 
+  void compute() {
     IPU_R_ExecuteSetViewSize_UnpackMiscValues(
       (R_ExecuteSetViewSize_MiscValues_t*) &miscValues[0]
     );
     R_ExecuteSetViewSize();
-    return true;
   }
 };
\ No newline at end of file
diff --git a/src/ipu/r_data.c b/src/ipu/r_data.c
index 984ea02..0e4a33c 100644
--- a/src/ipu/r_data.c
+++ b/src/ipu/r_data.c
@@ -373,6 +373,7 @@ byte *R_GetColumn(int tex, int col) {
   return texturecomposite[tex] + ofs;
 }
 */
+__SUPER__
 static void GenerateTextureHashTable(void) {
   IPUpatchlesstexture_t **rover;
   int i;
@@ -417,6 +418,7 @@ static void GenerateTextureHashTable(void) {
 // Initializes the texture list
 //  with the textures from the world map.
 //
+__SUPER__
 void R_InitTextures(int* maptex, R_Init_MiscValues_t* miscVals) {
   maptexture_t *mtexture;
   IPUpatchlesstexture_t *texture;
@@ -694,6 +696,7 @@ int R_FlatNumForName(char *name) {
 }
 */
 
+__SUPER__
 int JOSEFstrsame(const char* restrict a, const char* restrict b) {
   for (int i = 0; i < 8; ++i) {
     char c_a = JOSEFtoupper(a[i]), c_b = JOSEFtoupper(b[i]);
@@ -708,6 +711,7 @@ int JOSEFstrsame(const char* restrict a, const char* restrict b) {
 // Check whether texture is available.
 // Filter out NoTexture indicator.
 //
+__SUPER__
 int R_CheckTextureNumForName(char *name) {
   IPUpatchlesstexture_t *texture;
   int key;
@@ -736,6 +740,7 @@ int R_CheckTextureNumForName(char *name) {
 // Calls R_CheckTextureNumForName,
 //  aborts with error message.
 //
+__SUPER__
 int R_TextureNumForName(char *name) {
   int i;
 
diff --git a/src/ipu/r_data.h b/src/ipu/r_data.h
index cb9305d..7fd5498 100644
--- a/src/ipu/r_data.h
+++ b/src/ipu/r_data.h
@@ -21,22 +21,23 @@
 #define __R_DATA__
 
 #include "doomtype.h"
+#include "ipu_utils.h"
 
 // Retrieve column data for span blitting.
-byte *R_GetColumn(int tex, int col);
+__SUPER__ byte *R_GetColumn(int tex, int col);
 
 // I/O, setting up the stuff.
-void R_InitData(void);
-void R_PrecacheLevel(void);
+__SUPER__ void R_InitData(void);
+__SUPER__ void R_PrecacheLevel(void);
 
 // Retrieval.
 // Floor/ceiling opaque texture tiles,
 // lookup by name. For animation?
-int R_FlatNumForName(char *name);
+__SUPER__ int R_FlatNumForName(char *name);
 
 // Called by P_Ticker for switches and animations,
 // returns the texture number for the texture name.
-int R_TextureNumForName(char *name);
-int R_CheckTextureNumForName(char *name);
+__SUPER__ int R_TextureNumForName(char *name);
+__SUPER__ int R_CheckTextureNumForName(char *name);
 
 #endif
diff --git a/src/ipu/r_draw.c b/src/ipu/r_draw.c
index 9cd9bd8..6405846 100644
--- a/src/ipu/r_draw.c
+++ b/src/ipu/r_draw.c
@@ -107,6 +107,7 @@ extern int lightnum;
 // Thus a special case loop for very fast rendering can
 //  be used. It has also been used with Wolfenstein 3D.
 //
+__SUPER__ 
 void R_DrawColumn(void) {
   int count;
   pixel_t *dest;
@@ -695,6 +696,7 @@ void R_DrawSpanLow(void) {
 //  for getting the framebuffer address
 //  of a pixel to draw.
 //
+__SUPER__ 
 void R_InitBuffer(int width, int height) {
   int i;
 
diff --git a/src/ipu/r_draw.h b/src/ipu/r_draw.h
index 35358e8..9dedb13 100644
--- a/src/ipu/r_draw.h
+++ b/src/ipu/r_draw.h
@@ -36,8 +36,8 @@ extern byte *dc_source;
 // The span blitting interface.
 // Hook in assembler or system specific BLT
 //  here.
-void R_DrawColumn(void);
-void R_DrawColumnLow(void);
+__SUPER__ void R_DrawColumn(void);
+__SUPER__ void R_DrawColumnLow(void);
 
 // The Spectre/Invisibility effect.
 void R_DrawFuzzColumn(void);
@@ -75,7 +75,7 @@ void R_DrawSpan(void);
 // Low resolution mode, 160x200?
 void R_DrawSpanLow(void);
 
-void R_InitBuffer(int width, int height);
+__SUPER__ void R_InitBuffer(int width, int height);
 
 // Initialize color translation tables,
 //  for player rendering etc.
diff --git a/src/ipu/r_main.c b/src/ipu/r_main.c
index 334c6ea..45d9c0a 100644
--- a/src/ipu/r_main.c
+++ b/src/ipu/r_main.c
@@ -111,6 +111,7 @@ int extralight;
 
 // void (*colfunc)(void); // JOSEF
 
+__SUPER__
 void colfunc() { // JOSEF: Don't call through pointer, explicit call
   if (!detailshift) R_DrawColumn();
   // else           R_DrawColumnLow(); // LATER
@@ -133,6 +134,7 @@ int abs(int);
 // Expand a given bbox
 // so that it encloses a given point.
 //
+__SUPER__
 void R_AddPointToBox(int x, int y, fixed_t *box) {
   if (x < box[BOXLEFT])
     box[BOXLEFT] = x;
@@ -150,6 +152,7 @@ void R_AddPointToBox(int x, int y, fixed_t *box) {
 //  check point against partition plane.
 // Returns side 0 (front) or 1 (back).
 //
+__SUPER__
 int R_PointOnSide(fixed_t x, fixed_t y, node_t *node) {
   fixed_t dx;
   fixed_t dy;
@@ -255,6 +258,7 @@ int R_PointOnSegSide(fixed_t x, fixed_t y, seg_t *line) {
 //  tangent (slope) value which is looked up in the
 //  tantoangle[] table.
 //
+__SUPER__
 angle_t R_PointToAngle(fixed_t x, fixed_t y) {
   x -= viewx;
   y -= viewy;
@@ -315,6 +319,7 @@ angle_t R_PointToAngle(fixed_t x, fixed_t y) {
   return 0;
 }
 
+__SUPER__
 angle_t R_PointToAngle2(fixed_t x1, fixed_t y1, fixed_t x2, fixed_t y2) {
   viewx = x1;
   viewy = y1;
@@ -322,6 +327,7 @@ angle_t R_PointToAngle2(fixed_t x1, fixed_t y1, fixed_t x2, fixed_t y2) {
   return R_PointToAngle(x2, y2);
 }
 
+__SUPER__
 fixed_t R_PointToDist(fixed_t x, fixed_t y) {
   int angle;
   fixed_t dx;
@@ -358,6 +364,7 @@ fixed_t R_PointToDist(fixed_t x, fixed_t y) {
 //
 // R_InitPointToAngle
 //
+__SUPER__
 void R_InitPointToAngle(void) {
 // UNUSED - now getting from tables.c
 #if 0
@@ -383,6 +390,7 @@ void R_InitPointToAngle(void) {
 //  at the given angle.
 // rw_distance must be calculated first.
 //
+__SUPER__
 fixed_t R_ScaleFromGlobalAngle(angle_t visangle) {
   fixed_t scale;
   angle_t anglea;
@@ -413,8 +421,8 @@ fixed_t R_ScaleFromGlobalAngle(angle_t visangle) {
   angleb = ANG90 + (visangle - rw_normalangle);
 
   // both sines are allways positive
-  sinea = IPU_finesine(anglea >> ANGLETOFINESHIFT);
-  sineb = IPU_finesine(angleb >> ANGLETOFINESHIFT);
+  sinea = finesine[anglea >> ANGLETOFINESHIFT];
+  sineb = finesine[angleb >> ANGLETOFINESHIFT];
 
   num = FixedMul(projection, sineb) << detailshift;
   den = FixedMul(rw_distance, sinea);
@@ -468,6 +476,7 @@ void R_InitTables(void) {
 //
 // R_InitTextureMapping
 //
+__SUPER__
 void R_InitTextureMapping(void) {
   int i;
   int x;
@@ -482,7 +491,6 @@ void R_InitTextureMapping(void) {
   //  so FIELDOFVIEW angles covers SCREENWIDTH.
   focallength =
       FixedDiv(centerxfrac, finetangent[FINEANGLES / 4 + FIELDOFVIEW / 2]);
-
   for (i = 0; i < FINEANGLES / 2; i++) {
     if (finetangent[i] > FRACUNIT * 2)
       t = -1;
@@ -570,6 +578,7 @@ boolean setsizeneeded;
 int setblocks;
 int setdetail;
 
+__SUPER__
 void R_SetViewSize(int blocks, int detail) {
   setsizeneeded = true;
   setblocks = blocks;
@@ -579,6 +588,7 @@ void R_SetViewSize(int blocks, int detail) {
 //
 // R_ExecuteSetViewSize
 //
+__SUPER__
 void R_ExecuteSetViewSize(void) {
   fixed_t cosadj;
   fixed_t dy;
@@ -728,6 +738,7 @@ subsector_t *R_PointInSubsector(fixed_t x, fixed_t y) {
 //
 // R_SetupFrame
 //
+__SUPER__ 
 void R_SetupFrame(player_t *player) {
   int i;
 
@@ -739,7 +750,7 @@ void R_SetupFrame(player_t *player) {
 
   viewz = player->viewz;
 
-  viewsin = IPU_finesine(viewangle >> ANGLETOFINESHIFT);
+  viewsin = finesine[viewangle >> ANGLETOFINESHIFT];
   viewcos = finecosine[viewangle >> ANGLETOFINESHIFT];
 
   // JOSEF TMP, some tests to about replacing sin LUTs!
@@ -773,6 +784,7 @@ void R_SetupFrame(player_t *player) {
 //
 // R_RenderView
 //
+__SUPER__ 
 void R_RenderPlayerView(player_t *player) {
   R_SetupFrame(player);
 
diff --git a/src/ipu/r_main.h b/src/ipu/r_main.h
index 53c6824..8fdaf43 100644
--- a/src/ipu/r_main.h
+++ b/src/ipu/r_main.h
@@ -83,7 +83,7 @@ extern int detailshift;
 //
 
 // extern void (*colfunc)(void); // JOSEF
-void colfunc(void); // Don't want to call via pointer on IPU
+__SUPER__ void colfunc(void); // Don't want to call via pointer on IPU
 extern void (*transcolfunc)(void);
 extern void (*basecolfunc)(void);
 extern void (*fuzzcolfunc)(void);
@@ -97,33 +97,33 @@ extern int tileRightClip;
 
 //
 // Utility functions.
-int R_PointOnSide(fixed_t x, fixed_t y, node_t *node);
+__SUPER__ int R_PointOnSide(fixed_t x, fixed_t y, node_t *node);
 
-int R_PointOnSegSide(fixed_t x, fixed_t y, seg_t *line);
+__SUPER__ int R_PointOnSegSide(fixed_t x, fixed_t y, seg_t *line);
 
-angle_t R_PointToAngle(fixed_t x, fixed_t y);
+__SUPER__ angle_t R_PointToAngle(fixed_t x, fixed_t y);
 
-angle_t R_PointToAngle2(fixed_t x1, fixed_t y1, fixed_t x2, fixed_t y2);
+__SUPER__ angle_t R_PointToAngle2(fixed_t x1, fixed_t y1, fixed_t x2, fixed_t y2);
 
-fixed_t R_PointToDist(fixed_t x, fixed_t y);
+__SUPER__ fixed_t R_PointToDist(fixed_t x, fixed_t y);
 
-fixed_t R_ScaleFromGlobalAngle(angle_t visangle);
+__SUPER__ fixed_t R_ScaleFromGlobalAngle(angle_t visangle);
 
-subsector_t *R_PointInSubsector(fixed_t x, fixed_t y);
+__SUPER__ subsector_t *R_PointInSubsector(fixed_t x, fixed_t y);
 
-void R_AddPointToBox(int x, int y, fixed_t *box);
+__SUPER__ void R_AddPointToBox(int x, int y, fixed_t *box);
 
 //
 // REFRESH - the actual rendering functions.
 //
 
 // Called by G_Drawer.
-void R_RenderPlayerView(player_t *player);
+__SUPER__ void R_RenderPlayerView(player_t *player);
 
 // Called by startup code.
-void R_Init(void);
+__SUPER__ void R_Init(void);
 
 // Called by M_Responder.
-void R_SetViewSize(int blocks, int detail);
+__SUPER__ void R_SetViewSize(int blocks, int detail);
 
 #endif
diff --git a/src/ipu/r_plane.c b/src/ipu/r_plane.c
index 330abb6..dbc657a 100644
--- a/src/ipu/r_plane.c
+++ b/src/ipu/r_plane.c
@@ -93,6 +93,7 @@ fixed_t cachedystep[SCREENHEIGHT];
 // R_InitPlanes
 // Only at game startup.
 //
+__SUPER__ 
 void R_InitPlanes(void) {
   // Doh!
 }
@@ -161,6 +162,7 @@ void R_MapPlane(int y, int x1, int x2) {
 // R_ClearPlanes
 // At begining of frame.
 //
+__SUPER__ 
 void R_ClearPlanes(void) {
   int i;
   angle_t angle;
diff --git a/src/ipu/r_plane.h b/src/ipu/r_plane.h
index 2f8ff90..d880513 100644
--- a/src/ipu/r_plane.h
+++ b/src/ipu/r_plane.h
@@ -37,17 +37,17 @@ extern short ceilingclip[SCREENWIDTH];
 extern fixed_t yslope[SCREENHEIGHT];
 extern fixed_t distscale[SCREENWIDTH];
 
-void R_InitPlanes(void);
-void R_ClearPlanes(void);
+__SUPER__ void R_InitPlanes(void);
+__SUPER__ void R_ClearPlanes(void);
 
-void R_MapPlane(int y, int x1, int x2);
+__SUPER__ void R_MapPlane(int y, int x1, int x2);
 
-void R_MakeSpans(int x, int t1, int b1, int t2, int b2);
+__SUPER__ void R_MakeSpans(int x, int t1, int b1, int t2, int b2);
 
-void R_DrawPlanes(void);
+__SUPER__ void R_DrawPlanes(void);
 
-visplane_t *R_FindPlane(fixed_t height, int picnum, int lightlevel);
+__SUPER__ visplane_t *R_FindPlane(fixed_t height, int picnum, int lightlevel);
 
-visplane_t *R_CheckPlane(visplane_t *pl, int start, int stop);
+__SUPER__ visplane_t *R_CheckPlane(visplane_t *pl, int start, int stop);
 
 #endif
diff --git a/src/ipu/r_segs.c b/src/ipu/r_segs.c
index faa5a8f..1b1d9ed 100644
--- a/src/ipu/r_segs.c
+++ b/src/ipu/r_segs.c
@@ -38,6 +38,7 @@
 #include "v_patch.h"
 
 #include <print.h>
+#include "ipu_utils.h"
 #include "ipu_texturetiles.h"
 
 
@@ -93,6 +94,7 @@ lighttable_t **walllights;
 
 short *maskedtexturecol;
 
+__SUPER__ 
 int abs(int); // JOSEF
 
 /*
@@ -190,7 +192,7 @@ void R_RenderMaskedSegRange(drawseg_t *ds, int x1, int x2) {
 //
 #define HEIGHTBITS 12
 #define HEIGHTUNIT (1 << HEIGHTBITS)
-
+__SUPER__ 
 void R_RenderSegLoop(void) {
   angle_t angle;
   unsigned index;
@@ -341,6 +343,7 @@ int lightnum; // JOSEF: Temprarily made this global for vis
 // A wall segment will be drawn
 //  between start and stop pixels (inclusive).
 //
+__SUPER__ 
 void R_StoreWallRange(int start, int stop) {
   fixed_t hyp;
   fixed_t sineval;
diff --git a/src/ipu/r_segs.h b/src/ipu/r_segs.h
index b22e539..609d9ba 100644
--- a/src/ipu/r_segs.h
+++ b/src/ipu/r_segs.h
@@ -19,8 +19,9 @@
 #ifndef __R_SEGS__
 #define __R_SEGS__
 
+#include "ipu_texturetiles.h"
 #include "r_defs.h"
 
-void R_RenderMaskedSegRange(drawseg_t *ds, int x1, int x2);
+__SUPER__ void R_RenderMaskedSegRange(drawseg_t *ds, int x1, int x2);
 
 #endif
diff --git a/src/ipu/tables.c b/src/ipu/tables.c
index 8803442..d62ff0e 100644
--- a/src/ipu/tables.c
+++ b/src/ipu/tables.c
@@ -32,13 +32,14 @@
 
 #include "tables.h"
 #include <math.h> // JOSEF
+#include "ipu_texturetiles.h"
 
 // to get a global angle from cartesian coordinates, the coordinates are
 // flipped until they are in the first octant of the coordinate system, then
 // the y (<=x) is scaled and divided by x to get a tangent (slope) value
 // which is looked up in the tantoangle[] table.  The +1 size is to handle
 // the case when x==y without additional checking.
-
+__SUPER__
 int SlopeDiv(unsigned int num, unsigned int den)
 {
     unsigned ans;
@@ -579,6 +580,7 @@ const fixed_t finetangent[4096] =
 };
 
 // 40K lookup table is a bit dear for IPU, eventually calculate live instead
+__SUPER__
 fixed_t IPU_finesine(int i) {
     return finesine[i];
 
diff --git a/src/ipu/tables.h b/src/ipu/tables.h
index 516f2b6..10fbc5b 100644
--- a/src/ipu/tables.h
+++ b/src/ipu/tables.h
@@ -47,7 +47,7 @@
 
 // Effective size is 10240.
 extern const fixed_t finesine[5*FINEANGLES/4];
-fixed_t IPU_finesine(int i); // Replace big lookup on IPU
+__SUPER__ fixed_t IPU_finesine(int i); // Replace big lookup on IPU
 
 // Re-use data, is just PI/2 pahse shift.
 extern const fixed_t *finecosine;
@@ -90,7 +90,7 @@ extern const angle_t tantoangle[SLOPERANGE+1];
 
 // Utility function,
 //  called by R_PointToAngle.
-int SlopeDiv(unsigned int num, unsigned int den);
+__SUPER__ int SlopeDiv(unsigned int num, unsigned int den);
 
 
 #endif
diff --git a/src/ipu/v_video.c b/src/ipu/v_video.c
index 8e48fa8..bb0f29a 100644
--- a/src/ipu/v_video.c
+++ b/src/ipu/v_video.c
@@ -133,7 +133,7 @@ void V_SetPatchClipCallback(vpatchclipfunc_t func)
 // V_DrawPatch
 // Masks a column based masked pic to the screen. 
 //
-
+__SUPER__
 void V_DrawPatch(int x, int y, patch_t *patch)
 { 
     int count;
@@ -597,6 +597,7 @@ void V_DrawRawScreen(byte *raw)
 //
 // V_Init
 // 
+__SUPER__
 void V_Init (void) 
 { 
     // no-op!
@@ -606,6 +607,7 @@ void V_Init (void)
 
 // Set the buffer that the code draws to.
 
+__SUPER__
 void V_UseBuffer(pixel_t *buffer)
 {
     dest_screen = buffer;
@@ -613,6 +615,7 @@ void V_UseBuffer(pixel_t *buffer)
 
 // Restore screen buffer to the i_video screen buffer.
 
+__SUPER__
 void V_RestoreBuffer(void)
 {
     dest_screen = I_VideoBuffer;
diff --git a/src/ipu/v_video.h b/src/ipu/v_video.h
index 7b98c35..624c371 100644
--- a/src/ipu/v_video.h
+++ b/src/ipu/v_video.h
@@ -27,6 +27,8 @@
 // Needed because we are refering to patches.
 #include "v_patch.h"
 
+#include "ipu_utils.h"
+
 //
 // VIDEO
 //
@@ -42,67 +44,67 @@ extern byte *tinttable;
 // haleyjd 08/28/10: Patch clipping callback, implemented to support Choco
 // Strife.
 typedef boolean (*vpatchclipfunc_t)(patch_t *, int, int);
-void V_SetPatchClipCallback(vpatchclipfunc_t func);
+__SUPER__ void V_SetPatchClipCallback(vpatchclipfunc_t func);
 
 
 // Allocates buffer screens, call before R_Init.
-void V_Init (void);
+__SUPER__ void V_Init (void);
 
 // Draw a block from the specified source screen to the screen.
 
-void V_CopyRect(int srcx, int srcy, pixel_t *source,
+__SUPER__ void V_CopyRect(int srcx, int srcy, pixel_t *source,
                 int width, int height,
                 int destx, int desty);
 
-void V_DrawPatch(int x, int y, patch_t *patch);
-void V_DrawPatchFlipped(int x, int y, patch_t *patch);
-void V_DrawTLPatch(int x, int y, patch_t *patch);
-void V_DrawAltTLPatch(int x, int y, patch_t * patch);
-void V_DrawShadowedPatch(int x, int y, patch_t *patch);
-void V_DrawXlaPatch(int x, int y, patch_t * patch);     // villsa [STRIFE]
-void V_DrawPatchDirect(int x, int y, patch_t *patch);
+__SUPER__ void V_DrawPatch(int x, int y, patch_t *patch);
+__SUPER__ void V_DrawPatchFlipped(int x, int y, patch_t *patch);
+__SUPER__ void V_DrawTLPatch(int x, int y, patch_t *patch);
+__SUPER__ void V_DrawAltTLPatch(int x, int y, patch_t * patch);
+__SUPER__ void V_DrawShadowedPatch(int x, int y, patch_t *patch);
+__SUPER__ void V_DrawXlaPatch(int x, int y, patch_t * patch);     // villsa [STRIFE]
+__SUPER__ void V_DrawPatchDirect(int x, int y, patch_t *patch);
 
 // Draw a linear block of pixels into the view buffer.
 
-void V_DrawBlock(int x, int y, int width, int height, pixel_t *src);
+__SUPER__ void V_DrawBlock(int x, int y, int width, int height, pixel_t *src);
 
-void V_MarkRect(int x, int y, int width, int height);
+__SUPER__ void V_MarkRect(int x, int y, int width, int height);
 
-void V_DrawFilledBox(int x, int y, int w, int h, int c);
-void V_DrawHorizLine(int x, int y, int w, int c);
-void V_DrawVertLine(int x, int y, int h, int c);
-void V_DrawBox(int x, int y, int w, int h, int c);
+__SUPER__ void V_DrawFilledBox(int x, int y, int w, int h, int c);
+__SUPER__ void V_DrawHorizLine(int x, int y, int w, int c);
+__SUPER__ void V_DrawVertLine(int x, int y, int h, int c);
+__SUPER__ void V_DrawBox(int x, int y, int w, int h, int c);
 
 // Draw a raw screen lump
 
-void V_DrawRawScreen(byte *raw);
+__SUPER__ void V_DrawRawScreen(byte *raw);
 
 // Temporarily switch to using a different buffer to draw graphics, etc.
 
-void V_UseBuffer(pixel_t *buffer);
+__SUPER__ void V_UseBuffer(pixel_t *buffer);
 
 // Return to using the normal screen buffer to draw graphics.
 
-void V_RestoreBuffer(void);
+__SUPER__ void V_RestoreBuffer(void);
 
 // Save a screenshot of the current screen to a file, named in the 
 // format described in the string passed to the function, eg.
 // "DOOM%02i.pcx"
 
-void V_ScreenShot(char *format);
+__SUPER__ void V_ScreenShot(char *format);
 
 // Load the lookup table for translucency calculations from the TINTTAB
 // lump.
 
-void V_LoadTintTable(void);
+__SUPER__ void V_LoadTintTable(void);
 
 // villsa [STRIFE]
 // Load the lookup table for translucency calculations from the XLATAB
 // lump.
 
-void V_LoadXlaTable(void);
+__SUPER__ void V_LoadXlaTable(void);
 
-void V_DrawMouseSpeedBox(int speed);
+__SUPER__ void V_DrawMouseSpeedBox(int speed);
 
 #endif
 
diff --git a/src/ipu/w_wad.c b/src/ipu/w_wad.c
index ed90cb6..f977996 100644
--- a/src/ipu/w_wad.c
+++ b/src/ipu/w_wad.c
@@ -68,13 +68,14 @@ static lumpinfo_t *reloadlumps = NULL;
 static char *reloadname = NULL;
 static int reloadlump = -1;
 */
-
+__SUPER__ 
 char JOSEFtoupper(const char c) {
     if (c < 97 || c > 122) return c;
     return c ^ 0x20;    
 }
 
 // Hash function used for lump names.
+__SUPER__ 
 unsigned int W_LumpNameHash(const char *s)
 {
     // This is the djb2 string hash function, modded to work on strings
diff --git a/src/ipu/w_wad.h b/src/ipu/w_wad.h
index ad860ca..b504421 100644
--- a/src/ipu/w_wad.h
+++ b/src/ipu/w_wad.h
@@ -52,25 +52,25 @@ struct lumpinfo_s
 extern lumpinfo_t **lumpinfo;
 extern unsigned int numlumps;
 
-wad_file_t *W_AddFile(char *filename);
-void W_Reload(void);
+__SUPER__ wad_file_t *W_AddFile(char *filename);
+__SUPER__ void W_Reload(void);
 
-lumpindex_t W_CheckNumForName(char *name);
-lumpindex_t W_GetNumForName(char *name);
+__SUPER__ lumpindex_t W_CheckNumForName(char *name);
+__SUPER__ lumpindex_t W_GetNumForName(char *name);
 
-int W_LumpLength(lumpindex_t lump);
-void W_ReadLump(lumpindex_t lump, void *dest);
+__SUPER__ int W_LumpLength(lumpindex_t lump);
+__SUPER__ void W_ReadLump(lumpindex_t lump, void *dest);
 
-void *W_CacheLumpNum(lumpindex_t lump, int tag);
-void *W_CacheLumpName(char *name, int tag);
+__SUPER__ void *W_CacheLumpNum(lumpindex_t lump, int tag);
+__SUPER__ void *W_CacheLumpName(char *name, int tag);
 
-void W_GenerateHashTable(void);
+__SUPER__ void W_GenerateHashTable(void);
 
-extern unsigned int W_LumpNameHash(const char *s);
+__SUPER__ extern unsigned int W_LumpNameHash(const char *s);
 
-void W_ReleaseLumpNum(lumpindex_t lump);
-void W_ReleaseLumpName(char *name);
+__SUPER__ void W_ReleaseLumpNum(lumpindex_t lump);
+__SUPER__ void W_ReleaseLumpName(char *name);
 
-char JOSEFtoupper(const char c);
+__SUPER__ char JOSEFtoupper(const char c);
 
 #endif