optimize memory access. reduce Memory_Read_U##_type call(it uses Memory::GetPointer, which is complex), and gain speed, may be 1 or more fps faster, I've tested on zelda tww,tp, starfox assault only.

Please review and may be the code need clean.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@2144 8ced0084-cf51-0410-be5f-012b33b47a6e

5 files changed, 185 insertions, 119 deletions

diff --git a/Source/Core/VideoCommon/Src/VertexLoader_Color.cpp b/Source/Core/VideoCommon/Src/VertexLoader_Color.cpp
index e204f927db..5dbeae4750 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader_Color.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader_Color.cpp
@@ -174,9 +174,12 @@ void LOADERDECL Color_ReadIndex8_24b_6666()
 {
 	u8 Index = DataReadU8();
 	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
-	u32 val = Memory_Read_U8(iAddress+2) | 
-			  (Memory_Read_U8(iAddress+1)<<8) |
-			  (Memory_Read_U8(iAddress)<<16); 
+	u8* pData = Memory_Read_U8_Ptr(iAddress);
+
+	u32 val = *(pData+2) | ((*(pData+1))<<8) | ((*pData)<<16);
+	//u32 val = Memory_Read_U8(iAddress+2) | 
+	//		  (Memory_Read_U8(iAddress+1)<<8) |
+	//		  (Memory_Read_U8(iAddress)<<16); 
 	
 	_SetCol6666(val);
 }
@@ -219,9 +222,13 @@ void LOADERDECL Color_ReadIndex16_24b_6666()
 {
 	u16 Index = DataReadU16();
 	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
-	u32 val = Memory_Read_U8(iAddress+2) | 
-			   (Memory_Read_U8(iAddress+1)<<8) |
-			   (Memory_Read_U8(iAddress)<<16); 
+
+	u8* pData = Memory_Read_U8_Ptr(iAddress);
+
+	u32 val = *(pData+2) | ((*(pData+1))<<8) | ((*pData)<<16);
+	//u32 val = Memory_Read_U8(iAddress+2) | 
+	//		   (Memory_Read_U8(iAddress+1)<<8) |
+	//		   (Memory_Read_U8(iAddress)<<16); 
 	_SetCol6666(val);
 }
 void LOADERDECL Color_ReadIndex16_32b_8888()
diff --git a/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp b/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp
index da2f990594..df5975c31e 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp
@@ -184,9 +184,10 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Byte()
 {
     u8 Index = DataReadU8();
     u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
-    *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress);
-    *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress+1);
-    *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress+2);
+	u8* pData = Memory_Read_U8_Ptr(iAddress);
+    *VertexManager::s_pCurBufferPointer++ = *(pData);	//Memory_Read_U8(iAddress);
+    *VertexManager::s_pCurBufferPointer++ = *(pData+1);	//Memory_Read_U8(iAddress+1);
+    *VertexManager::s_pCurBufferPointer++ = *(pData+2);	//Memory_Read_U8(iAddress+2);
 	VertexManager::s_pCurBufferPointer++;
 //    ((float*)VertexManager::s_pCurBufferPointer)[0] = ((float)(signed char)Memory_Read_U8(iAddress)+0.5f) / 127.5f;
 //    ((float*)VertexManager::s_pCurBufferPointer)[1] = ((float)(signed char)Memory_Read_U8(iAddress+1)+0.5f) / 127.5f;
@@ -199,9 +200,10 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Short()
 {
     u8 Index = DataReadU8();
     u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
-    ((u16*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U16(iAddress);
-    ((u16*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U16(iAddress+2);
-    ((u16*)VertexManager::s_pCurBufferPointer)[2] = Memory_Read_U16(iAddress+4);
+	u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
+	((u16*)VertexManager::s_pCurBufferPointer)[0] = Common::swap16(*pData);		//Memory_Read_U16(iAddress);
+    ((u16*)VertexManager::s_pCurBufferPointer)[1] = Common::swap16(*(pData+1));	//Memory_Read_U16(iAddress+2);
+    ((u16*)VertexManager::s_pCurBufferPointer)[2] = Common::swap16(*(pData+2));	//Memory_Read_U16(iAddress+4);
     VertexManager::s_pCurBufferPointer += 8;
     LOG_NORM16();
 }
@@ -210,9 +212,10 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Float()
 {
     u8 Index = DataReadU8();
     u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
-    ((u32*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U32(iAddress);
-    ((u32*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U32(iAddress+4);
-    ((u32*)VertexManager::s_pCurBufferPointer)[2] = Memory_Read_U32(iAddress+8);
+	u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);
+	((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*pData);		//Memory_Read_U32(iAddress);
+    ((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(*(pData+1));	//Memory_Read_U32(iAddress+4);
+    ((u32*)VertexManager::s_pCurBufferPointer)[2] = Common::swap32(*(pData+2));	//Memory_Read_U32(iAddress+8);
     VertexManager::s_pCurBufferPointer += 12;
     LOG_NORMF();
 }
@@ -220,12 +223,15 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Float()
 void LOADERDECL VertexLoader_Normal::Normal_Index8_Byte3_Indices1()
 {
     u8 Index = DataReadU8();
+	u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
+	u8* pData = Memory_Read_U8_Ptr(iAddress);
+
     for (int i = 0; i < 3; i++)
 	{
-        u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 1*3*i;
-        *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress);
-        *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress+1);
-        *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress+2);
+        //u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 1*3*i;
+        *VertexManager::s_pCurBufferPointer++ = *(pData + 3*i);		//Memory_Read_U8(iAddress);
+        *VertexManager::s_pCurBufferPointer++ = *(pData + 3*i + 1);	//Memory_Read_U8(iAddress+1);
+        *VertexManager::s_pCurBufferPointer++ = *(pData + 3*i + 2);	//Memory_Read_U8(iAddress+2);
 		VertexManager::s_pCurBufferPointer++;
         LOG_NORM8();
     }
@@ -234,12 +240,15 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Byte3_Indices1()
 void LOADERDECL VertexLoader_Normal::Normal_Index8_Short3_Indices1()
 {
     u8 Index = DataReadU8();
+	u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
+	u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
+
     for (int i = 0; i < 3; i++)
 	{
-        u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 2*3*i;
-        ((u16*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U16(iAddress);
-        ((u16*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U16(iAddress+2);
-        ((u16*)VertexManager::s_pCurBufferPointer)[2] = Memory_Read_U16(iAddress+4);
+        //u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 2*3*i;
+        ((u16*)VertexManager::s_pCurBufferPointer)[0] = Common::swap16(*(pData + 3*i));		//Memory_Read_U16(iAddress);
+        ((u16*)VertexManager::s_pCurBufferPointer)[1] = Common::swap16(*(pData + 3*i + 1));	//Memory_Read_U16(iAddress+2);
+        ((u16*)VertexManager::s_pCurBufferPointer)[2] = Common::swap16(*(pData + 3*i + 2));	//Memory_Read_U16(iAddress+4);
         VertexManager::s_pCurBufferPointer += 8;
         LOG_NORM16();
     }    
@@ -248,12 +257,14 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Short3_Indices1()
 void LOADERDECL VertexLoader_Normal::Normal_Index8_Float3_Indices1()
 {
     u8 Index = DataReadU8();
+	u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
+	u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);
     for (int i = 0; i < 3; i++)
 	{
-        u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 4*3*i;
-        ((u32*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U32(iAddress);
-        ((u32*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U32(iAddress+4);
-        ((u32*)VertexManager::s_pCurBufferPointer)[2] = Memory_Read_U32(iAddress+8);
+        //u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 4*3*i;
+        ((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*(pData + 3*i));		//Memory_Read_U32(iAddress);
+        ((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(*(pData + 3*i + 1));	//Memory_Read_U32(iAddress+4);
+        ((u32*)VertexManager::s_pCurBufferPointer)[2] = Common::swap32(*(pData + 3*i + 2));	//Memory_Read_U32(iAddress+8);
         VertexManager::s_pCurBufferPointer += 12;
         LOG_NORMF();
     }    
@@ -265,9 +276,10 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Byte3_Indices3()
 	{
         u8 Index = DataReadU8();
         u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 1*3*i;
-        *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress);
-        *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress+1);
-        *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress+2);
+		u8* pData = Memory_Read_U8_Ptr(iAddress);
+        *VertexManager::s_pCurBufferPointer++ = *(pData);	//Memory_Read_U8(iAddress);
+        *VertexManager::s_pCurBufferPointer++ = *(pData+1);	//Memory_Read_U8(iAddress+1);
+        *VertexManager::s_pCurBufferPointer++ = *(pData+2);	//Memory_Read_U8(iAddress+2);
         *VertexManager::s_pCurBufferPointer++;
         LOG_NORM8();
     }
@@ -279,9 +291,10 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Short3_Indices3()
 	{
         u8 Index = DataReadU8();
         u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 2*3*i;
-        ((u16*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U16(iAddress);
-        ((u16*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U16(iAddress+2);
-        ((u16*)VertexManager::s_pCurBufferPointer)[2] = Memory_Read_U16(iAddress+4);
+		u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
+        ((u16*)VertexManager::s_pCurBufferPointer)[0] = Common::swap16(*(pData));	//Memory_Read_U16(iAddress);
+        ((u16*)VertexManager::s_pCurBufferPointer)[1] = Common::swap16(*(pData+1));	//Memory_Read_U16(iAddress+2);
+        ((u16*)VertexManager::s_pCurBufferPointer)[2] = Common::swap16(*(pData+2));	//Memory_Read_U16(iAddress+4);
         VertexManager::s_pCurBufferPointer += 8;
         LOG_NORM16();
     }
@@ -293,9 +306,10 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Float3_Indices3()
 	{
         u8 Index = DataReadU8();
         u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 4*3*i;
-        ((u32*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U32(iAddress);
-        ((u32*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U32(iAddress+4);
-        ((u32*)VertexManager::s_pCurBufferPointer)[2] = Memory_Read_U32(iAddress+8);
+		u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);
+        ((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*(pData));	//Memory_Read_U32(iAddress);
+        ((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(*(pData+1));	//Memory_Read_U32(iAddress+4);
+        ((u32*)VertexManager::s_pCurBufferPointer)[2] = Common::swap32(*(pData+2));	//Memory_Read_U32(iAddress+8);
         VertexManager::s_pCurBufferPointer += 12;
         LOG_NORMF();
     }    
@@ -309,9 +323,10 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Byte()
 {
     u16 Index = DataReadU16();
     u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
-    *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress);
-    *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress+1);
-    *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress+2);
+	u8* pData = Memory_Read_U8_Ptr(iAddress);
+    *VertexManager::s_pCurBufferPointer++ = *(pData);	//Memory_Read_U8(iAddress);
+    *VertexManager::s_pCurBufferPointer++ = *(pData+1);	//Memory_Read_U8(iAddress+1);
+    *VertexManager::s_pCurBufferPointer++ = *(pData+2);	//Memory_Read_U8(iAddress+2);
 	VertexManager::s_pCurBufferPointer++;
     LOG_NORM8();
 }
@@ -320,9 +335,11 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Short()
 {
     u16 Index = DataReadU16();
     u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
-    ((u16*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U16(iAddress);
-    ((u16*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U16(iAddress+2);
-    ((u16*)VertexManager::s_pCurBufferPointer)[2] = Memory_Read_U16(iAddress+4);
+	u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
+
+    ((u16*)VertexManager::s_pCurBufferPointer)[0] = Common::swap16(*(pData));//Memory_Read_U16(iAddress);
+    ((u16*)VertexManager::s_pCurBufferPointer)[1] = Common::swap16(*(pData+1));//Memory_Read_U16(iAddress+2);
+    ((u16*)VertexManager::s_pCurBufferPointer)[2] = Common::swap16(*(pData+2));//Memory_Read_U16(iAddress+4);
     VertexManager::s_pCurBufferPointer += 8;
     LOG_NORM16();
 }
@@ -331,9 +348,10 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Float()
 {
     u16 Index = DataReadU16();
     u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
-    ((u32*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U32(iAddress);
-    ((u32*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U32(iAddress+4);
-    ((u32*)VertexManager::s_pCurBufferPointer)[2] = Memory_Read_U32(iAddress+8);
+	u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);
+    ((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*(pData));//Memory_Read_U32(iAddress);
+    ((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(*(pData+1));//Memory_Read_U32(iAddress+4);
+    ((u32*)VertexManager::s_pCurBufferPointer)[2] = Common::swap32(*(pData+2));//Memory_Read_U32(iAddress+8);
     VertexManager::s_pCurBufferPointer += 12;
     LOG_NORMF();
 }
@@ -341,12 +359,14 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Float()
 void LOADERDECL VertexLoader_Normal::Normal_Index16_Byte3_Indices1()
 {
     u16 Index = DataReadU16();
+	u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
+	u8* pData = Memory_Read_U8_Ptr(iAddress);
     for (int i = 0; i < 3; i++)
 	{
-        u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 1*3*i;
-        *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress);
-        *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress+1);
-        *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress+2);
+        //u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 1*3*i;
+        *VertexManager::s_pCurBufferPointer++ =  *(pData + 3*i);		//Memory_Read_U8(iAddress);
+        *VertexManager::s_pCurBufferPointer++ =  *(pData + 3*i + 1);	//Memory_Read_U8(iAddress+1);
+        *VertexManager::s_pCurBufferPointer++ =  *(pData + 3*i + 2);	//Memory_Read_U8(iAddress+2);
 		VertexManager::s_pCurBufferPointer++;
         LOG_NORM8();
     }
@@ -355,12 +375,15 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Byte3_Indices1()
 void LOADERDECL VertexLoader_Normal::Normal_Index16_Short3_Indices1()
 {
     u16 Index = DataReadU16();
+	u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
+	u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
+
     for (int i = 0; i < 3; i++)
     {
-        u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 2*3*i;
-        ((u16*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U16(iAddress);
-        ((u16*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U16(iAddress+2);
-        ((u16*)VertexManager::s_pCurBufferPointer)[2] = Memory_Read_U16(iAddress+4);
+        //u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 2*3*i;
+        ((u16*)VertexManager::s_pCurBufferPointer)[0] = Common::swap16(*(pData + 3*i));		//Memory_Read_U16(iAddress);
+        ((u16*)VertexManager::s_pCurBufferPointer)[1] = Common::swap16(*(pData + 3*i + 1));	//Memory_Read_U16(iAddress+2);
+        ((u16*)VertexManager::s_pCurBufferPointer)[2] = Common::swap16(*(pData + 3*i + 2));	//Memory_Read_U16(iAddress+4);
         VertexManager::s_pCurBufferPointer += 8;
         LOG_NORM16();
     }
@@ -369,12 +392,15 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Short3_Indices1()
 void LOADERDECL VertexLoader_Normal::Normal_Index16_Float3_Indices1()
 {
     u16 Index = DataReadU16();
+	u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
+	u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);
+
     for (int i = 0; i < 3; i++)
     {
-        u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 4*3*i;
-        ((u32*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U32(iAddress);
-        ((u32*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U32(iAddress+4);
-		((u32*)VertexManager::s_pCurBufferPointer)[2] = Memory_Read_U32(iAddress+8);
+        //u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 4*3*i;
+        ((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*(pData + 3*i));		//Memory_Read_U32(iAddress);
+        ((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(*(pData + 3*i + 1));	//Memory_Read_U32(iAddress+4);
+		((u32*)VertexManager::s_pCurBufferPointer)[2] = Common::swap32(*(pData + 3*i + 2));	//Memory_Read_U32(iAddress+8);
         VertexManager::s_pCurBufferPointer += 12;
         LOG_NORMF();
     }
@@ -386,9 +412,10 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Byte3_Indices3()
 	{
         u16 Index = DataReadU16();
         u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 1*3*i;
-        *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress);
-        *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress+1);
-        *VertexManager::s_pCurBufferPointer++ = Memory_Read_U8(iAddress+2);
+		u8* pData = Memory_Read_U8_Ptr(iAddress);
+        *VertexManager::s_pCurBufferPointer++ = *(pData);	//Memory_Read_U8(iAddress);
+        *VertexManager::s_pCurBufferPointer++ = *(pData+1);	//Memory_Read_U8(iAddress+1);
+        *VertexManager::s_pCurBufferPointer++ = *(pData+2);	//Memory_Read_U8(iAddress+2);
 		VertexManager::s_pCurBufferPointer++;
         LOG_NORM8();
     }    
@@ -400,9 +427,10 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Short3_Indices3()
     {
         u16 Index = DataReadU16();
         u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 2*3*i;
-        ((u16*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U16(iAddress);
-        ((u16*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U16(iAddress+2);
-        ((u16*)VertexManager::s_pCurBufferPointer)[2] = Memory_Read_U16(iAddress+4);
+		u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
+        ((u16*)VertexManager::s_pCurBufferPointer)[0] = Common::swap16(*(pData));	//Memory_Read_U16(iAddress);
+        ((u16*)VertexManager::s_pCurBufferPointer)[1] = Common::swap16(*(pData+1));	//Memory_Read_U16(iAddress+2);
+        ((u16*)VertexManager::s_pCurBufferPointer)[2] = Common::swap16(*(pData+2));	//Memory_Read_U16(iAddress+4);
         VertexManager::s_pCurBufferPointer += 8;
         LOG_NORM16();
     }
@@ -415,9 +443,10 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Float3_Indices3()
     {
         u16 Index = DataReadU16();
         u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 4*3*i;
-        ((u32*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U32(iAddress);
-        ((u32*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U32(iAddress+4);
-        ((u32*)VertexManager::s_pCurBufferPointer)[2] = Memory_Read_U32(iAddress+8);
+		u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);
+        ((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*(pData));	//Memory_Read_U32(iAddress);
+        ((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(*(pData+1));	//Memory_Read_U32(iAddress+4);
+        ((u32*)VertexManager::s_pCurBufferPointer)[2] = Common::swap32(*(pData+2));	//Memory_Read_U32(iAddress+8);
         VertexManager::s_pCurBufferPointer += 12;
         LOG_NORMF();
     }    
diff --git a/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp b/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp
index 2a9d4bcb09..922d1d4db5 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp
@@ -137,40 +137,43 @@ void LOADERDECL Pos_ReadDirect_Float()
 	VertexManager::s_pCurBufferPointer += 12;
 }
 
-#define Pos_ReadIndex_Byte(T) { \
-	u32 iAddress = arraybases[ARRAY_POSITION] + ((u32)Index * arraystrides[ARRAY_POSITION]); \
-	((float*)VertexManager::s_pCurBufferPointer)[0] = ((float)(T)Memory_Read_U8(iAddress)) * posScale; \
-	((float*)VertexManager::s_pCurBufferPointer)[1] = ((float)(T)Memory_Read_U8(iAddress+1)) * posScale; \
-	if (pVtxAttr->PosElements) \
-		((float*)VertexManager::s_pCurBufferPointer)[2] = ((float)(T)Memory_Read_U8(iAddress+2)) * posScale; \
-	else \
-		((float*)VertexManager::s_pCurBufferPointer)[2] = 1.0f; \
-	LOG_VTX(); \
-	VertexManager::s_pCurBufferPointer += 12; \
-}
-
-#define Pos_ReadIndex_Short(T) { \
-	u32 iAddress = arraybases[ARRAY_POSITION] + ((u32)Index * arraystrides[ARRAY_POSITION]); \
-	((float*)VertexManager::s_pCurBufferPointer)[0] = ((float)(T)Memory_Read_U16(iAddress)) * posScale; \
-	((float*)VertexManager::s_pCurBufferPointer)[1] = ((float)(T)Memory_Read_U16(iAddress+2)) * posScale; \
-	if (pVtxAttr->PosElements) \
-		((float*)VertexManager::s_pCurBufferPointer)[2] = ((float)(T)Memory_Read_U16(iAddress+4)) * posScale; \
-	else \
-		((float*)VertexManager::s_pCurBufferPointer)[2] = 1.0f; \
-	LOG_VTX(); \
-	VertexManager::s_pCurBufferPointer += 12; \
-}
-
-#define Pos_ReadIndex_Float() { \
+#define Pos_ReadIndex_Byte(T) {																	\
+	u32 iAddress = arraybases[ARRAY_POSITION] + ((u32)Index * arraystrides[ARRAY_POSITION]);	\
+	u8* pData = Memory_Read_U8_Ptr(iAddress);													\
+	((float*)VertexManager::s_pCurBufferPointer)[0] = ((float)(T)(*(pData))) * posScale;		\
+	((float*)VertexManager::s_pCurBufferPointer)[1] = ((float)(T)(*(pData+1))) * posScale;		\
+	if (pVtxAttr->PosElements)																	\
+		((float*)VertexManager::s_pCurBufferPointer)[2] = ((float)(T)(*(pData+2))) * posScale;	\
+	else																						\
+		((float*)VertexManager::s_pCurBufferPointer)[2] = 1.0f;									\
+	LOG_VTX();																					\
+	VertexManager::s_pCurBufferPointer += 12;													\
+}
+
+#define Pos_ReadIndex_Short(T) {																		\
+	u32 iAddress = arraybases[ARRAY_POSITION] + ((u32)Index * arraystrides[ARRAY_POSITION]);			\
+	u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);												\
+	((float*)VertexManager::s_pCurBufferPointer)[0] = ((float)(T)Common::swap16(*(pData))) * posScale;	\
+	((float*)VertexManager::s_pCurBufferPointer)[1] = ((float)(T)Common::swap16(*(pData+1))) * posScale;\
+	if (pVtxAttr->PosElements)																			\
+	((float*)VertexManager::s_pCurBufferPointer)[2] = ((float)(T)Common::swap16(*(pData+2))) * posScale;\
+	else																								\
+		((float*)VertexManager::s_pCurBufferPointer)[2] = 1.0f;											\
+	LOG_VTX();																							\
+	VertexManager::s_pCurBufferPointer += 12;															\
+}
+
+#define Pos_ReadIndex_Float() {															\
 	u32 iAddress = arraybases[ARRAY_POSITION] + (Index * arraystrides[ARRAY_POSITION]); \
-	((u32*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U32(iAddress); \
-	((u32*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U32(iAddress+4); \
-	if (pVtxAttr->PosElements) \
-		((u32*)VertexManager::s_pCurBufferPointer)[2] = Memory_Read_U32(iAddress+8); \
-	else \
-		((float*)VertexManager::s_pCurBufferPointer)[2] = 1.0f; \
-	LOG_VTX(); \
-	VertexManager::s_pCurBufferPointer += 12; \
+	u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);								\
+	((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*(pData));			\
+	((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(*(pData+1));			\
+	if (pVtxAttr->PosElements)															\
+		((u32*)VertexManager::s_pCurBufferPointer)[2] = Common::swap32(*(pData+2));		\
+	else																				\
+		((float*)VertexManager::s_pCurBufferPointer)[2] = 1.0f;							\
+	LOG_VTX();																			\
+	VertexManager::s_pCurBufferPointer += 12;											\
 }
 
 // ==============================================================================
diff --git a/Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp b/Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp
index bd6cb43f99..1015049097 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp
@@ -130,9 +130,10 @@ void LOADERDECL TexCoord_ReadIndex8_UByte2()
 {
 	u8 Index = DataReadU8();
 	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
+	u8* pData = Memory_Read_U8_Ptr(iAddress);
 
-	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(u8)Memory_Read_U8(iAddress) * tcScale[tcIndex];
-	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(u8)Memory_Read_U8(iAddress+1) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(u8)(*(pData)) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(u8)(*(pData+1)) * tcScale[tcIndex];
 	LOG_TEX2();
 	VertexManager::s_pCurBufferPointer += 8;
 	tcIndex++;
@@ -152,9 +153,10 @@ void LOADERDECL TexCoord_ReadIndex8_Byte2()
 {
 	u8 Index = DataReadU8();
 	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
+	u8* pData = Memory_Read_U8_Ptr(iAddress);
 
-	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s8)Memory_Read_U8(iAddress) * tcScale[tcIndex];
-	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(s8)Memory_Read_U8(iAddress+1) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s8)(*(pData)) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(s8)(*(pData+1)) * tcScale[tcIndex];
 	LOG_TEX2();
 	VertexManager::s_pCurBufferPointer += 8;
 	tcIndex++;
@@ -174,9 +176,10 @@ void LOADERDECL TexCoord_ReadIndex8_UShort2()
 {
 	u8 Index = DataReadU8();
 	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
+	u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
 
-	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(u16)Memory_Read_U16(iAddress) * tcScale[tcIndex];
-	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(u16)Memory_Read_U16(iAddress+2) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(u16)Common::swap16(*(pData)) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(u16)Common::swap16(*(pData+1)) * tcScale[tcIndex];
 	LOG_TEX2();
 	VertexManager::s_pCurBufferPointer += 8;
 	tcIndex++;
@@ -196,9 +199,10 @@ void LOADERDECL TexCoord_ReadIndex8_Short2()
 {
 	u8 Index = DataReadU8();
 	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
+	u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
 
-	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s16)Memory_Read_U16(iAddress) * tcScale[tcIndex];
-	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(s16)Memory_Read_U16(iAddress+2) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s16)Common::swap16(*(pData)) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(s16)Common::swap16(*(pData+1)) * tcScale[tcIndex];
 	LOG_TEX2();
 	VertexManager::s_pCurBufferPointer += 8;
 	tcIndex++;
@@ -217,8 +221,10 @@ void LOADERDECL TexCoord_ReadIndex8_Float2()
 {
 	u16 Index = DataReadU8(); 
 	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
-	((u32*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U32(iAddress);
-	((u32*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U32(iAddress+4);
+	u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);
+
+	((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*(pData));
+	((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(*(pData+1));
 	LOG_TEX2();
 	VertexManager::s_pCurBufferPointer += 8;
 	tcIndex++;
@@ -239,9 +245,10 @@ void LOADERDECL TexCoord_ReadIndex16_UByte2()
 {
 	u16 Index = DataReadU16(); 
 	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
+	u8* pData = Memory_Read_U8_Ptr(iAddress);
 
-	((float*)VertexManager::s_pCurBufferPointer)[0] =  (float)(u8)Memory_Read_U8(iAddress) * tcScale[tcIndex];
-	((float*)VertexManager::s_pCurBufferPointer)[1] =  (float)(u8)Memory_Read_U8(iAddress+1) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[0] =  (float)(u8)(*(pData)) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[1] =  (float)(u8)(*(pData+1)) * tcScale[tcIndex];
 	LOG_TEX2();
 	VertexManager::s_pCurBufferPointer += 8;
 	tcIndex++;
@@ -261,9 +268,10 @@ void LOADERDECL TexCoord_ReadIndex16_Byte2()
 {
 	u16 Index = DataReadU16(); 
 	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
+	u8* pData = Memory_Read_U8_Ptr(iAddress);
 
-	((float*)VertexManager::s_pCurBufferPointer)[0] =  (float)(s8)Memory_Read_U8(iAddress) * tcScale[tcIndex];
-	((float*)VertexManager::s_pCurBufferPointer)[1] =  (float)(s8)Memory_Read_U8(iAddress+1) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[0] =  (float)(s8)(*(pData)) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[1] =  (float)(s8)(*(pData+1)) * tcScale[tcIndex];
 	LOG_TEX2();
 	VertexManager::s_pCurBufferPointer += 8;
 	tcIndex++;
@@ -283,9 +291,10 @@ void LOADERDECL TexCoord_ReadIndex16_UShort2()
 {
 	u16 Index = DataReadU16(); 
 	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
+	u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
 
-	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(u16)Memory_Read_U16(iAddress) * tcScale[tcIndex];
-	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(u16)Memory_Read_U16(iAddress+2) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(u16)Common::swap16(*(pData)) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(u16)Common::swap16(*(pData+1)) * tcScale[tcIndex];
 	LOG_TEX2();
 	VertexManager::s_pCurBufferPointer += 8;
 	tcIndex++;
@@ -305,9 +314,10 @@ void LOADERDECL TexCoord_ReadIndex16_Short2()
 {
 	u16 Index = DataReadU16(); 
 	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
+	u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
 
-	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s16)Memory_Read_U16(iAddress) * tcScale[tcIndex];
-	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(s16)Memory_Read_U16(iAddress+2) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s16)Common::swap16(*(pData)) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(s16)Common::swap16(*(pData+1)) * tcScale[tcIndex];
 	LOG_TEX2();
 	VertexManager::s_pCurBufferPointer += 8;
 	tcIndex++;
@@ -327,9 +337,10 @@ void LOADERDECL TexCoord_ReadIndex16_Float2()
 {
 	u16 Index = DataReadU16(); 
 	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
+	u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);
 
-	((u32*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U32(iAddress);
-	((u32*)VertexManager::s_pCurBufferPointer)[1] = Memory_Read_U32(iAddress + 4);
+	((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*(pData));
+	((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(*(pData+1));
 	LOG_TEX2();
 	VertexManager::s_pCurBufferPointer += 8;
 	tcIndex++;
diff --git a/Source/Core/VideoCommon/Src/VideoCommon.h b/Source/Core/VideoCommon/Src/VideoCommon.h
index f4c99b845a..0a99340c02 100644
--- a/Source/Core/VideoCommon/Src/VideoCommon.h
+++ b/Source/Core/VideoCommon/Src/VideoCommon.h
@@ -48,6 +48,7 @@ extern volatile u32 g_XFBUpdateRequested;
 

 void DebugLog(const char* _fmt, ...);

 

+//////////////////////////////////////////////////////////////////////////

 inline u8 *Memory_GetPtr(u32 _uAddress)

 {

 	return g_VideoInitialize.pGetMemoryPointer(_uAddress);//&g_pMemory[_uAddress & RAM_MASK];

@@ -67,7 +68,22 @@ inline u32 Memory_Read_U32(u32 _uAddress)
 {

 	return Common::swap32(*(u32*)g_VideoInitialize.pGetMemoryPointer(_uAddress));

 }

+//////////////////////////////////////////////////////////////////////////

+inline u8* Memory_Read_U8_Ptr(u32 _uAddress)

+{

+	return (u8*)g_VideoInitialize.pGetMemoryPointer(_uAddress);//g_pMemory[_uAddress & RAM_MASK];

+}

 

+inline u16* Memory_Read_U16_Unswapped_Ptr(u32 _uAddress)

+{

+	return (u16*)g_VideoInitialize.pGetMemoryPointer(_uAddress);

+}

+

+inline u32* Memory_Read_U32_Unswapped_Ptr(u32 _uAddress)

+{

+	return (u32*)g_VideoInitialize.pGetMemoryPointer(_uAddress);

+}

+//////////////////////////////////////////////////////////////////////////

 inline u32 Memory_Read_U32_Unswapped(u32 _uAddress)

 {

 	return *(u32*)g_VideoInitialize.pGetMemoryPointer(_uAddress);