diff --git a/Source/Plugins/Plugin_VideoSoftware/Src/CommandProcessor.cpp b/Source/Plugins/Plugin_VideoSoftware/Src/CommandProcessor.cpp
index 76ddf9cc72..0d3f8f7874 100644
--- a/Source/Plugins/Plugin_VideoSoftware/Src/CommandProcessor.cpp
+++ b/Source/Plugins/Plugin_VideoSoftware/Src/CommandProcessor.cpp
@@ -46,6 +46,8 @@ const int maxCommandBufferWrite = commandBufferSize - commandBufferCopySize;
 u8 commandBuffer[commandBufferSize];
 u32 readPos;
 u32 writePos;
+bool interruptSet;
+int et_UpdateInterrupts;
 
 CPReg cpreg; // shared between gfx and emulator thread
 
@@ -55,20 +57,13 @@ void DoState(PointerWrap &p)
 	p.Do(cpreg);
 }
 
-// function
-void UpdateFifoRegister();
-void UpdateInterrupts();
-
 // does it matter that there is no synchronization between threads during writes?
 inline void WriteLow (u32& _reg, u16 lowbits)  {_reg = (_reg & 0xFFFF0000) | lowbits;}
 inline void WriteHigh(u32& _reg, u16 highbits) {_reg = (_reg & 0x0000FFFF) | ((u32)highbits << 16);}
-//inline void WriteLow (volatile u32& _reg, u16 lowbits)  {Common::SyncInterlockedExchange((LONG*)&_reg,(_reg & 0xFFFF0000) | lowbits);}
-//inline void WriteHigh(volatile u32& _reg, u16 highbits) {Common::SyncInterlockedExchange((LONG*)&_reg,(_reg & 0x0000FFFF) | ((u32)highbits << 16));}
 
 inline u16 ReadLow  (u32 _reg)  {return (u16)(_reg & 0xFFFF);}
 inline u16 ReadHigh (u32 _reg)  {return (u16)(_reg >> 16);}
 
-int et_UpdateInterrupts;
 
 void UpdateInterrupts_Wrapper(u64 userdata, int cyclesLate)
 {
@@ -97,25 +92,13 @@ void Init()
     readPos = 0;
     writePos = 0;
 
+    interruptSet = false;
+
     g_pVideoData = 0;
 }
 
 void Shutdown()
 {
-#ifndef _WIN32
-  //  delete fifo.sync;
-#endif
-}
-
-void Read16(u16& _rReturnValue, const u32 _Address)
-{
-	DEBUG_LOG(COMMANDPROCESSOR, "(r): 0x%08x", _Address);
-
-    u32 regAddr = (_Address & 0xFFF) >> 1;
-    if (regAddr < 0x20)
-        _rReturnValue = ((u16*)&cpreg)[regAddr];
-    else
-        _rReturnValue = 0;
 }
 
 void RunGpu()
@@ -127,12 +110,25 @@ void RunGpu()
 	    LoadDefaultSSEState();			
         
         // run the opcode decoder
+        do {
         RunBuffer();
+        } while (cpreg.ctrl.GPReadEnable && !cpreg.status.Breakpoint && cpreg.rwdistance >= commandBufferCopySize);
 
         LoadSSEState();
     }
 }
 
+void Read16(u16& _rReturnValue, const u32 _Address)
+{
+	DEBUG_LOG(COMMANDPROCESSOR, "(r): 0x%08x", _Address);
+
+    u32 regAddr = (_Address & 0xFFF) >> 1;
+    if (regAddr < 0x20)
+        _rReturnValue = ((u16*)&cpreg)[regAddr];
+    else
+        _rReturnValue = 0;
+}
+
 void Write16(const u16 _Value, const u32 _Address)
 {
 	INFO_LOG(COMMANDPROCESSOR, "(write16): 0x%04x @ 0x%08x",_Value,_Address);
@@ -149,6 +145,8 @@ void Write16(const u16 _Value, const u32 _Address)
             cpreg.status.Hex = _Value;
 
 			INFO_LOG(COMMANDPROCESSOR,"\t write to STATUS_REGISTER : %04x", _Value);
+
+            UpdateInterrupts();
 		}
 		break;
 
@@ -158,23 +156,23 @@ void Write16(const u16 _Value, const u32 _Address)
 
             // clear breakpoint if BPEnable and CPIntEnable are 0
             if (!cpreg.ctrl.BPEnable) {
-                if (!cpreg.ctrl.CPIntEnable) {
+                if (!cpreg.ctrl.BreakPointIntEnable) {
                     cpreg.status.Breakpoint = 0;
                 }
             }
 
-			UpdateInterrupts();
-
 			DEBUG_LOG(COMMANDPROCESSOR,"\t write to CTRL_REGISTER : %04x", _Value);
-			DEBUG_LOG(COMMANDPROCESSOR, "\t GPREAD %s | CPULINK %s | BP %s || CPIntEnable %s | OvF %s | UndF %s"
+			DEBUG_LOG(COMMANDPROCESSOR, "\t GPREAD %s | CPULINK %s | BP %s || BPIntEnable %s | OvF %s | UndF %s"
                 , cpreg.ctrl.GPReadEnable ?				"ON" : "OFF"
                 , cpreg.ctrl.GPLinkEnable ?				"ON" : "OFF"
                 , cpreg.ctrl.BPEnable ?					"ON" : "OFF"
-                , cpreg.ctrl.CPIntEnable ?				"ON" : "OFF"
+                , cpreg.ctrl.BreakPointIntEnable ?		"ON" : "OFF"
 				, cpreg.ctrl.FifoOverflowIntEnable ?	"ON" : "OFF"
 				, cpreg.ctrl.FifoUnderflowIntEnable ?	"ON" : "OFF"
 				);
 
+            UpdateInterrupts();
+
 		}
 		break;
 
@@ -188,6 +186,8 @@ void Write16(const u16 _Value, const u32 _Address)
                 cpreg.status.UnderflowLoWatermark = 0;
 
             INFO_LOG(COMMANDPROCESSOR,"\t write to CLEAR_REGISTER : %04x",_Value);
+
+            UpdateInterrupts();
 		}		
 		break;
 
@@ -303,19 +303,21 @@ void STACKALIGN GatherPipeBursted()
 
 void UpdateInterrupts()
 {
-    bool bpInt = cpreg.status.Breakpoint && cpreg.ctrl.BPEnable;
+    bool bpInt = cpreg.status.Breakpoint && cpreg.ctrl.BreakPointIntEnable;
     bool ovfInt = cpreg.status.OverflowHiWatermark && cpreg.ctrl.FifoOverflowIntEnable;
     bool undfInt = cpreg.status.UnderflowLoWatermark && cpreg.ctrl.FifoUnderflowIntEnable;    
 
     DEBUG_LOG(COMMANDPROCESSOR, "\tUpdate Interrupts");
-    DEBUG_LOG(COMMANDPROCESSOR, "\tCPIntEnable %s | BP %s | OvF %s | UndF %s"
-                , cpreg.ctrl.CPIntEnable ?				"ON" : "OFF"
+    DEBUG_LOG(COMMANDPROCESSOR, "\tBreakPointIntEnable %s | BP %s | OvF %s | UndF %s"
+        , cpreg.ctrl.BreakPointIntEnable ?				"ON" : "OFF"
                 , cpreg.ctrl.BPEnable ?					"ON" : "OFF"                
 				, cpreg.ctrl.FifoOverflowIntEnable ?	"ON" : "OFF"
 				, cpreg.ctrl.FifoUnderflowIntEnable ?	"ON" : "OFF"
 				);
 
-    if (cpreg.ctrl.CPIntEnable && (bpInt || ovfInt || undfInt))
+    interruptSet = bpInt || ovfInt || undfInt;
+
+    if (interruptSet)
 	{
 		DEBUG_LOG(COMMANDPROCESSOR,"Interrupt set");
         g_VideoInitialize.pSetInterrupt(INT_CAUSE_CP, true);        
@@ -334,8 +336,6 @@ void UpdateInterruptsFromVideoPlugin()
 
 void ReadFifo()
 {
-    bool updateInterrupts = false;
-
     cpreg.status.ReadIdle = 0;
 
     // update rwdistance
@@ -347,7 +347,6 @@ void ReadFifo()
 
     // overflow check
     cpreg.status.OverflowHiWatermark = cpreg.rwdistance < cpreg.hiwatermark?0:1;
-    updateInterrupts |= cpreg.ctrl.FifoOverflowIntEnable && cpreg.status.OverflowHiWatermark;
 
     // read from fifo    
     u8 *ptr = g_VideoInitialize.pGetMemoryPointer(cpreg.readptr);    
@@ -363,8 +362,6 @@ void ReadFifo()
         {
             cpreg.status.Breakpoint = 1;
             DEBUG_LOG(VIDEO,"Hit breakpoint at %x", readptr);
-            if (cpreg.ctrl.CPIntEnable)
-                updateInterrupts = true;
         }
         else
         {
@@ -389,12 +386,20 @@ void ReadFifo()
 
     // underflow check
     cpreg.status.UnderflowLoWatermark = cpreg.rwdistance > cpreg.lowatermark?0:1;
-    updateInterrupts |= cpreg.ctrl.FifoUnderflowIntEnable && cpreg.status.UnderflowLoWatermark;
 
     cpreg.status.ReadIdle = 1;
 
-    if (updateInterrupts)
+    bool bpInt = cpreg.status.Breakpoint && cpreg.ctrl.BreakPointIntEnable;
+    bool ovfInt = cpreg.status.OverflowHiWatermark && cpreg.ctrl.FifoOverflowIntEnable;
+    bool undfInt = cpreg.status.UnderflowLoWatermark && cpreg.ctrl.FifoUnderflowIntEnable;
+
+    bool interrupt = bpInt || ovfInt || undfInt;
+
+    if (interrupt != interruptSet)
+    {
+        interruptSet = interrupt;
         UpdateInterruptsFromVideoPlugin();
+    }
 }
 
 bool RunBuffer()
@@ -440,4 +445,4 @@ bool RunBuffer()
     return ranDecoder;
 }
 
-} // end of namespace CommandProcessor
\ No newline at end of file
+} // end of namespace CommandProcessor
diff --git a/Source/Plugins/Plugin_VideoSoftware/Src/CommandProcessor.h b/Source/Plugins/Plugin_VideoSoftware/Src/CommandProcessor.h
index 8a7b0ef52f..ed2fc5d1ac 100644
--- a/Source/Plugins/Plugin_VideoSoftware/Src/CommandProcessor.h
+++ b/Source/Plugins/Plugin_VideoSoftware/Src/CommandProcessor.h
@@ -76,7 +76,7 @@ namespace CommandProcessor
 	    struct
 	    {
 		    unsigned GPReadEnable			:	1;
-		    unsigned CPIntEnable			:	1;
+		    unsigned BreakPointIntEnable	:	1;
 		    unsigned FifoOverflowIntEnable	:	1;
 		    unsigned FifoUnderflowIntEnable	:	1;
 		    unsigned GPLinkEnable			:	1;
diff --git a/Source/Plugins/Plugin_VideoSoftware/Src/DebugUtil.cpp b/Source/Plugins/Plugin_VideoSoftware/Src/DebugUtil.cpp
index 80232e1b1d..c5ca3b41cd 100644
--- a/Source/Plugins/Plugin_VideoSoftware/Src/DebugUtil.cpp
+++ b/Source/Plugins/Plugin_VideoSoftware/Src/DebugUtil.cpp
@@ -32,6 +32,20 @@ namespace DebugUtil
 {
 
 u32 skipFrames = 0;
+const int NumObjectBuffers = 32;
+u8 ObjectBuffer[NumObjectBuffers][EFB_WIDTH*EFB_HEIGHT*4];
+bool DrawnToBuffer[NumObjectBuffers];
+const char* ObjectBufferName[NumObjectBuffers];
+
+void Init()
+{
+    for (int i = 0; i < NumObjectBuffers; i++)
+    {
+        memset(ObjectBuffer[i], 0, sizeof(ObjectBuffer[i]));
+        DrawnToBuffer[i] = false;
+        ObjectBufferName[i] = 0;
+    }
+}
 
 bool SaveTexture(const char* filename, u32 texmap, int width, int height)
 {
@@ -133,11 +147,24 @@ void DumpDepth(const char* filename)
     delete []data;
 }
 
+void DrawObjectBuffer(s16 x, s16 y, u8 *color, int buffer, const char *name)
+{
+    u32 offset = (x + y * EFB_WIDTH) * 4;
+    u8 *dst = &ObjectBuffer[buffer][offset];
+    *(dst++) = color[2];
+    *(dst++) = color[1];
+    *(dst++) = color[0];
+    *(dst++) = color[3];
+
+    DrawnToBuffer[buffer] = true;
+    ObjectBufferName[buffer] = name;
+}
+
 void OnObjectBegin()
 {
     if (!g_SkipFrame)
     {
-        if (g_Config.bDumpTextures)
+        if (g_Config.bDumpTextures && stats.thisFrame.numDrawnObjects >= g_Config.drawStart && stats.thisFrame.numDrawnObjects < g_Config.drawEnd)
             DumpActiveTextures();
 
         if (g_Config.bHwRasterizer)
@@ -149,12 +176,23 @@ void OnObjectEnd()
 {
     if (!g_SkipFrame)
     {
-        if (g_Config.bDumpObjects)
+        if (g_Config.bDumpObjects && stats.thisFrame.numDrawnObjects >= g_Config.drawStart && stats.thisFrame.numDrawnObjects < g_Config.drawEnd)
             DumpEfb(StringFromFormat("%s/object%i.tga", FULL_FRAMES_DIR, stats.thisFrame.numDrawnObjects).c_str());
 
         if (g_Config.bHwRasterizer)
             HwRasterizer::EndTriangles();
 
+        for (int i = 0; i < NumObjectBuffers; i++)
+        {
+            if (DrawnToBuffer[i])
+            {
+                DrawnToBuffer[i] = false;
+                SaveTGA(StringFromFormat("%s/object%i_%s(%i).tga", FULL_FRAMES_DIR,
+                    stats.thisFrame.numDrawnObjects, ObjectBufferName[i], i).c_str(), EFB_WIDTH, EFB_HEIGHT, ObjectBuffer[i]);
+                memset(ObjectBuffer[i], 0, sizeof(ObjectBuffer[i]));
+            }
+        }
+
         stats.thisFrame.numDrawnObjects++;
     }
 }
diff --git a/Source/Plugins/Plugin_VideoSoftware/Src/DebugUtil.h b/Source/Plugins/Plugin_VideoSoftware/Src/DebugUtil.h
index f26ffcd4c4..0d3ff77fe4 100644
--- a/Source/Plugins/Plugin_VideoSoftware/Src/DebugUtil.h
+++ b/Source/Plugins/Plugin_VideoSoftware/Src/DebugUtil.h
@@ -20,6 +20,8 @@
 
 namespace DebugUtil
 {
+    void Init();
+
     void GetTextureBGRA(u8 *dst, u32 texmap, int width, int height);
 
     void DumpActiveTextures();
@@ -28,6 +30,8 @@ namespace DebugUtil
     void OnObjectEnd();
 
     void OnFrameEnd();
+
+    void DrawObjectBuffer(s16 x, s16 y, u8 *color, int buffer, const char *name);
 }
 
 #endif 
\ No newline at end of file
diff --git a/Source/Plugins/Plugin_VideoSoftware/Src/PixelEngine.cpp b/Source/Plugins/Plugin_VideoSoftware/Src/PixelEngine.cpp
index debea567dd..24dd94d26d 100644
--- a/Source/Plugins/Plugin_VideoSoftware/Src/PixelEngine.cpp
+++ b/Source/Plugins/Plugin_VideoSoftware/Src/PixelEngine.cpp
@@ -136,8 +136,7 @@ void UpdateInterrupts()
 // Called only if BPMEM_PE_TOKEN_INT_ID is ack by GP
 void SetToken_OnMainThread(u64 userdata, int cyclesLate)
 {
-	g_bSignalTokenInterrupt = true;	
-	//_dbg_assert_msg_(PIXELENGINE, (CommandProcessor::fifo.PEToken == (userdata&0xFFFF)), "WTF? BPMEM_PE_TOKEN_INT_ID's token != BPMEM_PE_TOKEN_ID's token" );
+	g_bSignalTokenInterrupt = true;
     INFO_LOG(PIXELENGINE, "VIDEO Plugin raises INT_CAUSE_PE_TOKEN (btw, token: %04x)", pereg.token);
 	UpdateInterrupts();
 }
@@ -158,10 +157,6 @@ void SetToken(const u16 _token, const int _bSetTokenAcknowledge)
         g_VideoInitialize.pScheduleEvent_Threadsafe(
 			0, et_SetTokenOnMainThread, _token | (_bSetTokenAcknowledge << 16));
 	}
-	else // set token value
-	{
-		ERROR_LOG(PIXELENGINE, "VIDEO plugin should set the token directly");
-	}
 }
 
 // SetFinish
diff --git a/Source/Plugins/Plugin_VideoSoftware/Src/Tev.cpp b/Source/Plugins/Plugin_VideoSoftware/Src/Tev.cpp
index bbe82bf2fb..2d961b44b3 100644
--- a/Source/Plugins/Plugin_VideoSoftware/Src/Tev.cpp
+++ b/Source/Plugins/Plugin_VideoSoftware/Src/Tev.cpp
@@ -21,6 +21,8 @@
 #include "EfbInterface.h"
 #include "TextureSampler.h"
 #include "Statistics.h"
+#include "VideoConfig.h"
+#include "DebugUtil.h"
 
 #include <math.h>
 
@@ -120,6 +122,11 @@ inline s16 Clamp255(s16 in)
     return in>255?255:(in<0?0:in);
 }
 
+inline s16 Clamp1024(s16 in)
+{
+     return in>1023?1023:(in<-1024?-1024:in);
+}
+
 inline void Tev::SetRasColor(int colorChan, int swaptable)
 {
     switch(colorChan)
@@ -168,12 +175,7 @@ inline void Tev::SetRasColor(int colorChan, int swaptable)
 
 void Tev::DrawColorRegular(TevStageCombiner::ColorCombiner &cc)
 {
-    struct {
-        unsigned a : 8;
-        unsigned b : 8;
-        unsigned c : 8;
-        signed   d : 11;
-    } InputReg;
+    InputRegType InputReg;
 
     for (int i = 0; i < 3; i++)
     {
@@ -202,12 +204,7 @@ void Tev::DrawColorCompare(TevStageCombiner::ColorCombiner &cc)
     u32 a;
     u32 b;
 
-    struct {
-        unsigned a : 8;
-        unsigned b : 8;
-        unsigned c : 8;
-        signed   d : 11;
-    } InputReg;
+    InputRegType InputReg;
 
     switch(cmp) {
     case TEVCMP_R8_GT:        
@@ -308,13 +305,7 @@ void Tev::DrawColorCompare(TevStageCombiner::ColorCombiner &cc)
 
 void Tev::DrawAlphaRegular(TevStageCombiner::AlphaCombiner &ac)
 {
-    struct {
-        unsigned a : 8;
-        unsigned b : 8;
-        unsigned c : 8;
-        signed   d : 11;
-    } InputReg;
-
+    InputRegType InputReg;
 
     InputReg.a = *m_AlphaInputLUT[ac.a];
     InputReg.b = *m_AlphaInputLUT[ac.b];
@@ -340,12 +331,7 @@ void Tev::DrawAlphaCompare(TevStageCombiner::AlphaCombiner &ac)
     u32 a;
     u32 b;
     
-    struct {
-        unsigned a : 8;
-        unsigned b : 8;
-        unsigned c : 8;
-        signed   d : 11;
-    } InputReg;
+    InputRegType InputReg;
 
     switch(cmp) {
     case TEVCMP_R8_GT:        
@@ -641,7 +627,7 @@ void Tev::Draw()
         StageKonst[ALP_C] = *(m_KonstLUT[ka][ALP_C]);
 
         // set color        
-        SetRasColor(order.getColorChan(stageOdd), ac.rswap * 2);        
+        SetRasColor(order.getColorChan(stageOdd), ac.rswap * 2);
 
         // combine inputs
         if (cc.bias != 3)
@@ -655,6 +641,12 @@ void Tev::Draw()
             Reg[cc.dest][GRN_C] = Clamp255(Reg[cc.dest][GRN_C]);
             Reg[cc.dest][BLU_C] = Clamp255(Reg[cc.dest][BLU_C]);
         }
+        else
+        {
+            Reg[cc.dest][RED_C] = Clamp1024(Reg[cc.dest][RED_C]);
+            Reg[cc.dest][GRN_C] = Clamp1024(Reg[cc.dest][GRN_C]);
+            Reg[cc.dest][BLU_C] = Clamp1024(Reg[cc.dest][BLU_C]);
+        }
 
         if (ac.bias != 3)
             DrawAlphaRegular(ac);
@@ -663,7 +655,16 @@ void Tev::Draw()
 
         if (ac.clamp)
             Reg[ac.dest][ALP_C] = Clamp255(Reg[ac.dest][ALP_C]);
+        else
+            Reg[ac.dest][ALP_C] = Clamp1024(Reg[ac.dest][ALP_C]);
 
+#ifdef _DEBUG
+        if (g_Config.bDumpTevStages)
+        {
+            u8 stage[4] = {(u8)Reg[0][0], (u8)Reg[0][1], (u8)Reg[0][2], (u8)Reg[0][3]};
+            DebugUtil::DrawObjectBuffer(Position[0], Position[1], stage, stageNum, "Stage");
+        }
+#endif
     }
     
     // convert to 8 bits per component
@@ -719,4 +720,4 @@ void Tev::SetRegColor(int reg, int comp, bool konst, s16 color)
     {
         Reg[reg][comp] = color;
     }
-}
\ No newline at end of file
+}
diff --git a/Source/Plugins/Plugin_VideoSoftware/Src/Tev.h b/Source/Plugins/Plugin_VideoSoftware/Src/Tev.h
index 953251433b..dbaa6e5ac0 100644
--- a/Source/Plugins/Plugin_VideoSoftware/Src/Tev.h
+++ b/Source/Plugins/Plugin_VideoSoftware/Src/Tev.h
@@ -51,6 +51,13 @@ class Tev
 
     void Indirect(unsigned int stageNum, float s, float t);    
 
+    struct InputRegType {
+        unsigned a : 8;
+        unsigned b : 8;
+        unsigned c : 8;
+        signed   d : 11;
+    };
+
 public:
     s32 Position[3];
     u8 Color[2][4];
diff --git a/Source/Plugins/Plugin_VideoSoftware/Src/TransformUnit.cpp b/Source/Plugins/Plugin_VideoSoftware/Src/TransformUnit.cpp
index da174085a5..38c1cc853f 100644
--- a/Source/Plugins/Plugin_VideoSoftware/Src/TransformUnit.cpp
+++ b/Source/Plugins/Plugin_VideoSoftware/Src/TransformUnit.cpp
@@ -199,6 +199,11 @@ inline float Clamp(float val, float a, float b)
     return val<a?a:val>b?b:val;
 }
 
+inline float SafeDivide(float n, float d)
+{
+    return (d==0)?(n>0?1:0):n/d;
+}
+
 void LightColor(const float *vertexPos, const float *normal, u8 lightNum, const LitChannel &chan, Vec3 &lightCol)
 {
     // must be the size of 3 32bit floats for the light pointer to be valid
@@ -244,15 +249,16 @@ void LightColor(const float *vertexPos, const float *normal, u8 lightNum, const
 
             float cosAtt = light->cosatt.x + (light->cosatt.y * attn) + (light->cosatt.z * attn * attn);
             float distAtt = light->distatt.x + (light->distatt.y * dist) + (light->distatt.z * dist2);
-            attn = distAtt==0.0f?0.0f:(max(0.0f, cosAtt) / distAtt); 
+            attn = SafeDivide(max(0.0f, cosAtt), distAtt);
         }
         else if (chan.attnfunc == 1) { // specular
-            attn = (light->pos * (*norm0)) > 0 ? max(0.0f, (light->dir * (*norm0))) : 0;
+            // donko - what is going on here?  655.36 is a guess but seems about right.
+            attn = (light->pos * (*norm0)) > -655.36 ? max(0.0f, (light->dir * (*norm0))) : 0;
             ldir.set(1.0f, attn, attn * attn);
 
-            float cosAtt = light->cosatt * ldir;
+            float cosAtt = max(0.0f, light->cosatt * ldir);
             float distAtt = light->distatt * ldir;
-            attn = distAtt==0.0f?1.0f:(max(0.0f, cosAtt) / distAtt); 
+            attn = SafeDivide(max(0.0f, cosAtt), distAtt);
         }
 
         switch (chan.diffusefunc) {
@@ -321,15 +327,16 @@ void LightAlpha(const float *vertexPos, const float *normal, u8 lightNum, const
 
             float cosAtt = light->cosatt.x + (light->cosatt.y * attn) + (light->cosatt.z * attn * attn);
             float distAtt = light->distatt.x + (light->distatt.y * dist) + (light->distatt.z * dist2);
-            attn = distAtt==0.0f?0.0f:(max(0.0f, cosAtt) / distAtt); 
+            attn = SafeDivide(max(0.0f, cosAtt), distAtt);
         }
         else if (chan.attnfunc == 1) { // specular
-            attn = (light->pos * (*norm0)) > 0 ? max(0.0f, (light->dir * (*norm0))) : 0;
+            // donko - what is going on here?  655.36 is a guess but seems about right.
+            attn = (light->pos * (*norm0)) > -655.36 ? max(0.0f, (light->dir * (*norm0))) : 0;
             ldir.set(1.0f, attn, attn * attn);
 
             float cosAtt = light->cosatt * ldir;
             float distAtt = light->distatt * ldir;
-            attn = distAtt==0.0f?1.0f:(max(0.0f, cosAtt) / distAtt); 
+            attn = SafeDivide(max(0.0f, cosAtt), distAtt);
         }
 
         switch (chan.diffusefunc) {
diff --git a/Source/Plugins/Plugin_VideoSoftware/Src/VideoConfig.cpp b/Source/Plugins/Plugin_VideoSoftware/Src/VideoConfig.cpp
index a0b3c252d0..06ff420c56 100644
--- a/Source/Plugins/Plugin_VideoSoftware/Src/VideoConfig.cpp
+++ b/Source/Plugins/Plugin_VideoSoftware/Src/VideoConfig.cpp
@@ -32,6 +32,7 @@ Config::Config()
     bDumpTextures = false;
     bDumpObjects = false;
     bDumpFrames = false;
+    bDumpTevStages = false;
 
     bHwRasterizer = false;
 
diff --git a/Source/Plugins/Plugin_VideoSoftware/Src/VideoConfig.h b/Source/Plugins/Plugin_VideoSoftware/Src/VideoConfig.h
index 8cf0b976e0..ea6d9511d7 100644
--- a/Source/Plugins/Plugin_VideoSoftware/Src/VideoConfig.h
+++ b/Source/Plugins/Plugin_VideoSoftware/Src/VideoConfig.h
@@ -38,6 +38,7 @@ struct Config
     bool bDumpTextures;
     bool bDumpObjects;
     bool bDumpFrames;
+    bool bDumpTevStages;
 
     bool bHwRasterizer;
 
diff --git a/Source/Plugins/Plugin_VideoSoftware/Src/main.cpp b/Source/Plugins/Plugin_VideoSoftware/Src/main.cpp
index 035a39a9bb..7a7d57a064 100644
--- a/Source/Plugins/Plugin_VideoSoftware/Src/main.cpp
+++ b/Source/Plugins/Plugin_VideoSoftware/Src/main.cpp
@@ -33,6 +33,7 @@
 #include "HwRasterizer.h"
 #include "LogManager.h"
 #include "EfbInterface.h"
+#include "DebugUtil.h"
 
 
 PLUGIN_GLOBALS* globals = NULL;
@@ -88,6 +89,7 @@ void Initialize(void *init)
     Rasterizer::Init();
     HwRasterizer::Init();
     Renderer::Init(_pVideoInitialize);
+    DebugUtil::Init();
 }
 
 void DoState(unsigned char **ptr, int mode)
@@ -109,6 +111,7 @@ void Video_Prepare(void)
 // Run from the CPU thread (from VideoInterface.cpp)
 void Video_BeginField(u32 xfbAddr, FieldType field, u32 fbWidth, u32 fbHeight)
 {	
+    g_VideoInitialize.pCopiedToXFB(true);
 }
 
 // Run from the CPU thread (from VideoInterface.cpp)