Cache vertex array pointers, a la magumagu's idea. Misc other stuff. Should give very minor speed boost.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@2185 8ced0084-cf51-0410-be5f-012b33b47a6e

12 files changed, 132 insertions, 176 deletions

diff --git a/Source/Core/Core/Src/HW/MemmapFunctions.cpp b/Source/Core/Core/Src/HW/MemmapFunctions.cpp
index 9e49c98d20..50db4dc741 100644
--- a/Source/Core/Core/Src/HW/MemmapFunctions.cpp
+++ b/Source/Core/Core/Src/HW/MemmapFunctions.cpp
@@ -265,7 +265,7 @@ u32 Read_U32(const u32 _Address)
 	if (mc)

 	{

 		mc->numHits++;

-		mc->Action(_var, _Address,false,4,PC);

+		mc->Action(_var, _Address, false, 4, PC);

 	}

 #endif

 	return _var;

@@ -281,7 +281,7 @@ u64 Read_U64(const u32 _Address)
 	if (mc)

 	{

 		mc->numHits++;

-		mc->Action((u32)_var, _Address,false,8,PC);

+		mc->Action((u32)_var, _Address, false, 8, PC);

 	}

 #endif

 	return _var;

diff --git a/Source/Core/DebuggerWX/Src/JitWindow.cpp b/Source/Core/DebuggerWX/Src/JitWindow.cpp
index 77c6d26770..63afb0c228 100644
--- a/Source/Core/DebuggerWX/Src/JitWindow.cpp
+++ b/Source/Core/DebuggerWX/Src/JitWindow.cpp
@@ -214,13 +214,13 @@ void CJitWindow::Compare(u32 em_address)
 
 		sptr += sprintf(sptr, "%i estimated cycles\n", st.numCycles);
 
-		sptr += sprintf(sptr, "Num instr: PPC: %i  x86: %i  (blowup: %i%%)\n", size, num_x86_instructions, 100 * num_x86_instructions / size);
-		sptr += sprintf(sptr, "Num bytes: PPC: %i  x86: %i  (blowup: %i%%)\n", size * 4, block->codeSize, 100 * block->codeSize / (4 * size));
+		sptr += sprintf(sptr, "Num instr: PPC: %i  x86: %i  (blowup: %i%%)\n", size, num_x86_instructions, 100 * (num_x86_instructions / size - 1));
+		sptr += sprintf(sptr, "Num bytes: PPC: %i  x86: %i  (blowup: %i%%)\n", size * 4, block->codeSize, 100 * (block->codeSize / (4 * size) - 1));
 
 		ppc_box->SetValue(wxString::FromAscii((char*)xDis));
 	} else {
 		ppc_box->SetValue(wxString::FromAscii(StringFromFormat("(non-code address: %08x)", em_address).c_str()));
-		x86_box->SetValue(wxString::FromAscii("---"));
+		x86_box->SetValue(wxString::FromAscii("---"));	
 	}
 	
 
diff --git a/Source/Core/VideoCommon/Src/CPMemory.cpp b/Source/Core/VideoCommon/Src/CPMemory.cpp
index 32cd4f0b32..a51c2d1454 100644
--- a/Source/Core/VideoCommon/Src/CPMemory.cpp
+++ b/Source/Core/VideoCommon/Src/CPMemory.cpp
@@ -21,6 +21,7 @@
 // CP state
 // STATE_TO_SAVE
 u32 arraybases[16];
+u8 *cached_arraybases[16];
 u32 arraystrides[16];
 TMatrixIndexA MatrixIndexA;
 TMatrixIndexB MatrixIndexB;
diff --git a/Source/Core/VideoCommon/Src/CPMemory.h b/Source/Core/VideoCommon/Src/CPMemory.h
index 62a5fa118d..af5be4952f 100644
--- a/Source/Core/VideoCommon/Src/CPMemory.h
+++ b/Source/Core/VideoCommon/Src/CPMemory.h
@@ -244,6 +244,7 @@ union TMatrixIndexB
 #pragma pack()
 
 extern u32 arraybases[16];
+extern u8 *cached_arraybases[16];
 extern u32 arraystrides[16];
 extern TMatrixIndexA MatrixIndexA;
 extern TMatrixIndexB MatrixIndexB;
diff --git a/Source/Core/VideoCommon/Src/VertexLoader.cpp b/Source/Core/VideoCommon/Src/VertexLoader.cpp
index e3abb2d3ef..65be39a539 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader.cpp
@@ -702,22 +702,42 @@ void VertexLoader::SetVAT(u32 _group0, u32 _group1, u32 _group2)
 
 void VertexLoader::AppendToString(std::string *dest)
 {
-	dest->reserve(150);
+	dest->reserve(250);
 	static const char *posMode[4] = {
 		"Inv",
 		"Dir",
-		"I8 ",
+		"I8",
 		"I16",
 	};
 	static const char *posFormats[5] = {
-		"u8 ", "s8 ", "u16", "s16", "flt",
+		"u8", "s8", "u16", "s16", "flt",
 	};
-	dest->append(StringFromFormat("%ib skin: %i P: %i %s %s ",
+	static const char *colorFormat[8] = {
+		"565",
+		"888",
+		"888x",
+		"4444",
+		"6666",
+		"8888",
+		"Inv",
+		"Inv",
+	};
+
+	dest->append(StringFromFormat("%ib skin: %i P: %i %s-%s ",
 		m_VertexSize, m_VtxDesc.PosMatIdx,
 		m_VtxAttr.PosElements ? 3 : 2, posMode[m_VtxDesc.Position], posFormats[m_VtxAttr.PosFormat]));
-	dest->append(StringFromFormat("Nrm: %i %s %s ",
-		m_VtxAttr.NormalElements, posMode[m_VtxDesc.Normal], posFormats[m_VtxAttr.NormalFormat]));
-
+	if (m_VtxDesc.Normal) {
+		dest->append(StringFromFormat("Nrm: %i %s-%s ",
+			m_VtxAttr.NormalElements, posMode[m_VtxDesc.Normal], posFormats[m_VtxAttr.NormalFormat]));
+	}
+	int color_mode[2] = {m_VtxDesc.Color0, m_VtxDesc.Color1};
+	for (int i = 0; i < 2; i++)
+	{
+		if (color_mode[i])
+		{
+			dest->append(StringFromFormat("C%i: %i %s-%s ", i, m_VtxAttr.color[i].Elements, posMode[color_mode[i]], colorFormat[m_VtxAttr.color[i].Comp]));
+		}
+	}
 	int tex_mode[8] = {
 		m_VtxDesc.Tex0Coord, m_VtxDesc.Tex1Coord, m_VtxDesc.Tex2Coord, m_VtxDesc.Tex3Coord, 
 		m_VtxDesc.Tex4Coord, m_VtxDesc.Tex5Coord, m_VtxDesc.Tex6Coord, m_VtxDesc.Tex7Coord
@@ -726,7 +746,7 @@ void VertexLoader::AppendToString(std::string *dest)
 	{
 		if (tex_mode[i])
 		{
-			dest->append(StringFromFormat("T%i: %i %s %s ",
+			dest->append(StringFromFormat("T%i: %i %s-%s ",
 				i, m_VtxAttr.texCoord[i].Elements, posMode[tex_mode[i]], posFormats[m_VtxAttr.texCoord[i].Format]));
 		}
 	}
diff --git a/Source/Core/VideoCommon/Src/VertexLoaderManager.h b/Source/Core/VideoCommon/Src/VertexLoaderManager.h
index 1061d85480..e6313f1f80 100644
--- a/Source/Core/VideoCommon/Src/VertexLoaderManager.h
+++ b/Source/Core/VideoCommon/Src/VertexLoaderManager.h
@@ -34,4 +34,6 @@ namespace VertexLoaderManager
 	void AppendListToString(std::string *dest);
 };
 
+void RecomputeCachedArraybases();
+
 #endif  // _VERTEXLOADERMANAGER_H
diff --git a/Source/Core/VideoCommon/Src/VertexLoader_Color.cpp b/Source/Core/VideoCommon/Src/VertexLoader_Color.cpp
index 5dbeae4750..16b8bb2a3b 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader_Color.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader_Color.cpp
@@ -67,23 +67,14 @@ void _SetCol565(u16 val)
 }
 //////////////////////////////////////////////////////////////////////////
 //////////////////////////////////////////////////////////////////////////
-inline u32 _Read24(u32 iAddress)
+inline u32 _Read24(const u8 *iAddress)
 {
-//	u32 col = Memory_Read_U8(iAddress)   << RSHIFT;      //should just get a pointer to main memory instead of going thru slow memhandler
-//	col     |= Memory_Read_U8(iAddress+1) << GSHIFT;    //we can guarantee that it is reading from main memory
-//	col     |= Memory_Read_U8(iAddress+2) << BSHIFT;
-//	return col | (0xFF<<ASHIFT);
-	return Memory_Read_U32_Unswapped(iAddress)|0xFF000000;
+	return *(const u32 *)iAddress | 0xFF000000;
 }
 
-inline u32 _Read32(u32 iAddress)
+inline u32 _Read32(const u8 *iAddress)
 {
-//	u32 col = Memory_Read_U8(iAddress)   << RSHIFT;      //should just get a pointer to main memory instead of going thru slow memhandler
-//	col     |= Memory_Read_U8(iAddress+1) << GSHIFT;    //we can guarantee that it is reading from main memory
-//	col     |= Memory_Read_U8(iAddress+2) << BSHIFT;
-//	col     |= Memory_Read_U8(iAddress+3) << ASHIFT;
-//	return col;
-	return Memory_Read_U32_Unswapped(iAddress);
+	return *(const u32 *)iAddress;
 }
 
 //////////////////////////////////////////////////////////////////////////
@@ -91,17 +82,17 @@ inline u32 _Read32(u32 iAddress)
 
 void LOADERDECL Color_ReadDirect_24b_888()
 {
-	u32 col = DataReadU8()<<RSHIFT;
-	col     |= DataReadU8()<<GSHIFT;
-	col     |= DataReadU8()<<BSHIFT;
-	_SetCol(col | (0xFF<<ASHIFT));
+	u32 col = DataReadU8() << RSHIFT;
+	col    |= DataReadU8() << GSHIFT;
+	col    |= DataReadU8() << BSHIFT;
+	_SetCol(col | (0xFF << ASHIFT));
 }
 
 void LOADERDECL Color_ReadDirect_32b_888x(){
-	u32 col = DataReadU8()<<RSHIFT;
-	col     |= DataReadU8()<<GSHIFT;
-	col     |= DataReadU8()<<BSHIFT;
-	_SetCol(col | (0xFF<<ASHIFT));
+	u32 col = DataReadU8() << RSHIFT;
+	col    |= DataReadU8() << GSHIFT;
+	col    |= DataReadU8() << BSHIFT;
+	_SetCol(col | (0xFF << ASHIFT));
 	DataReadU8();
 }
 void LOADERDECL Color_ReadDirect_16b_565()
@@ -131,13 +122,13 @@ void LOADERDECL Color_ReadDirect_32b_8888()
 {
 	// TODO (mb2): check this
 	u32 col = DataReadU8()<<RSHIFT;
-	col     |= DataReadU8()<<GSHIFT;
-	col     |= DataReadU8()<<BSHIFT;
-	col		|= DataReadU8()<<ASHIFT;
+	col    |= DataReadU8()<<GSHIFT;
+	col    |= DataReadU8()<<BSHIFT;
+	col    |= DataReadU8()<<ASHIFT;
 
 	// "kill" the alpha
 	if (!colElements[colIndex])	
-		col |= 0xFF<<ASHIFT;
+		col |= 0xFF << ASHIFT;
 
 	_SetCol(col);
 }
@@ -147,46 +138,38 @@ void LOADERDECL Color_ReadDirect_32b_8888()
 void LOADERDECL Color_ReadIndex8_16b_565()
 {
 	u8 Index = DataReadU8();
-	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
-	u16 val = Memory_Read_U16(iAddress);
+	u16 val = Common::swap16(*(const u16 *)(cached_arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex])));
 	_SetCol565(val);
 }
 void LOADERDECL Color_ReadIndex8_24b_888()
 {
 	u8 Index = DataReadU8();
-	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
+	const u8 *iAddress = cached_arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
 	_SetCol(_Read24(iAddress));
 }
 void LOADERDECL Color_ReadIndex8_32b_888x()
 {
 	u8 Index = DataReadU8();
-	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR]+colIndex);
+	const u8 *iAddress = cached_arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR]+colIndex);
 	_SetCol(_Read24(iAddress));
 }
 void LOADERDECL Color_ReadIndex8_16b_4444()
 {
 	u8 Index = DataReadU8();
-	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
-	u16 val = Memory_Read_U16(iAddress);
+	u16 val = Common::swap16(*(const u16 *)(cached_arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex])));
 	_SetCol4444(val);
 }
 void LOADERDECL Color_ReadIndex8_24b_6666()
 {
 	u8 Index = DataReadU8();
-	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
-	u8* pData = Memory_Read_U8_Ptr(iAddress);
-
+	const u8* pData = cached_arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
 	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_ReadIndex8_32b_8888()
 {
 	u8 Index = DataReadU8();
-	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
+	const u8 *iAddress = cached_arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
 	_SetCol(_Read32(iAddress));
 }
 //////////////////////////////////////////////////////////////////////////
@@ -195,46 +178,39 @@ void LOADERDECL Color_ReadIndex8_32b_8888()
 void LOADERDECL Color_ReadIndex16_16b_565()
 {
 	u16 Index = DataReadU16(); 
-	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
-	u16 val = Memory_Read_U16(iAddress);
+	u16 val = Common::swap16(*(const u16 *)(cached_arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex])));
 	_SetCol565(val);
 }
 void LOADERDECL Color_ReadIndex16_24b_888()
 {
 	u16 Index = DataReadU16(); 
-	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
+	const u8 *iAddress = cached_arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
 	_SetCol(_Read24(iAddress));
 }
 void LOADERDECL Color_ReadIndex16_32b_888x()
 {
 	u16 Index = DataReadU16(); 
-	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
+	const u8 *iAddress = cached_arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
 	_SetCol(_Read24(iAddress));
 }
 void LOADERDECL Color_ReadIndex16_16b_4444()
 {
 	u16 Index = DataReadU16();
-	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
-	u16 val = Memory_Read_U16(iAddress);
+	u16 val = Common::swap16(*(const u16 *)(cached_arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex])));
 	_SetCol4444(val);
 }
 void LOADERDECL Color_ReadIndex16_24b_6666()
 {
 	u16 Index = DataReadU16();
-	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
-
-	u8* pData = Memory_Read_U8_Ptr(iAddress);
-
+	const u8 *iAddress = cached_arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
+	const u8* pData = 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()
 {
 	u16 Index = DataReadU16(); 
-	u32 iAddress = arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
+	const u8 *iAddress = cached_arraybases[ARRAY_COLOR+colIndex] + (Index * arraystrides[ARRAY_COLOR+colIndex]);
 	_SetCol(_Read32(iAddress));
 }
 
diff --git a/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp b/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp
index df5975c31e..75fb9aef74 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp
@@ -183,8 +183,7 @@ void LOADERDECL VertexLoader_Normal::Normal_DirectFloat3()
 void LOADERDECL VertexLoader_Normal::Normal_Index8_Byte()
 {
     u8 Index = DataReadU8();
-    u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
-	u8* pData = Memory_Read_U8_Ptr(iAddress);
+	const u8* pData = cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
     *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);
@@ -199,8 +198,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Byte()
 void LOADERDECL VertexLoader_Normal::Normal_Index8_Short()
 {
     u8 Index = DataReadU8();
-    u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
-	u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
+	const u16* pData = (const u16 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
 	((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);
@@ -211,8 +209,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Short()
 void LOADERDECL VertexLoader_Normal::Normal_Index8_Float()
 {
     u8 Index = DataReadU8();
-    u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
-	u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);
+	const u32* pData = (const u32 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
 	((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);
@@ -223,12 +220,9 @@ 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);
-
+	const u8* pData = cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
     for (int i = 0; i < 3; i++)
 	{
-        //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);
@@ -240,12 +234,9 @@ 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);
-
+	const u16* pData = (const u16 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
     for (int i = 0; i < 3; i++)
 	{
-        //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);
@@ -257,11 +248,9 @@ 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);
+	const u32* pData = (const u32 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
     for (int i = 0; i < 3; i++)
 	{
-        //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);
@@ -275,8 +264,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Byte3_Indices3()
     for (int i = 0; i < 3; i++)
 	{
         u8 Index = DataReadU8();
-        u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 1*3*i;
-		u8* pData = Memory_Read_U8_Ptr(iAddress);
+		const u8* pData = cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 1*3*i;
         *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);
@@ -290,8 +278,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Short3_Indices3()
     for (int i = 0; i < 3; i++)
 	{
         u8 Index = DataReadU8();
-        u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 2*3*i;
-		u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
+		const u16* pData = (const u16 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 2*3*i);
         ((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);
@@ -305,8 +292,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Float3_Indices3()
     for (int i = 0; i < 3; i++)
 	{
         u8 Index = DataReadU8();
-        u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 4*3*i;
-		u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);
+		const u32* pData = (const u32 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 4*3*i);
         ((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);
@@ -322,8 +308,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Float3_Indices3()
 void LOADERDECL VertexLoader_Normal::Normal_Index16_Byte()
 {
     u16 Index = DataReadU16();
-    u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
-	u8* pData = Memory_Read_U8_Ptr(iAddress);
+	const u8* pData = cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
     *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);
@@ -334,8 +319,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Byte()
 void LOADERDECL VertexLoader_Normal::Normal_Index16_Short()
 {
     u16 Index = DataReadU16();
-    u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
-	u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
+	const u16* pData = (const u16 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
 
     ((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);
@@ -347,8 +331,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Short()
 void LOADERDECL VertexLoader_Normal::Normal_Index16_Float()
 {
     u16 Index = DataReadU16();
-    u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
-	u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);
+	const u32* pData = (const u32 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
     ((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);
@@ -359,11 +342,9 @@ 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);
+	const u8* pData = cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]);
     for (int i = 0; i < 3; i++)
 	{
-        //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);
@@ -375,12 +356,10 @@ 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);
+	const u16* pData = (const u16 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
 
     for (int i = 0; i < 3; i++)
     {
-        //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);
@@ -392,12 +371,10 @@ 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);
+	const u32* pData = (const u32 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
 
     for (int i = 0; i < 3; i++)
     {
-        //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);
@@ -411,8 +388,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Byte3_Indices3()
     for (int i = 0; i < 3; i++)
 	{
         u16 Index = DataReadU16();
-        u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 1*3*i;
-		u8* pData = Memory_Read_U8_Ptr(iAddress);
+		const u8* pData = cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 1*3*i;
         *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);
@@ -426,8 +402,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Short3_Indices3()
     for (int i = 0; i < 3; i++)
     {
         u16 Index = DataReadU16();
-        u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 2*3*i;
-		u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);
+		const u16* pData = (const u16 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 2*3*i);
         ((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);
@@ -442,8 +417,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Float3_Indices3()
     for (int i = 0; i < 3; i++)
     {
         u16 Index = DataReadU16();
-        u32 iAddress = arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 4*3*i;
-		u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);
+		const u32* pData = (const u32 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 4*3*i);
         ((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);
diff --git a/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp b/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp
index 922d1d4db5..44d15ccc23 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader_Position.cpp
@@ -138,8 +138,7 @@ void LOADERDECL Pos_ReadDirect_Float()
 }
 
 #define Pos_ReadIndex_Byte(T) {																	\
-	u32 iAddress = arraybases[ARRAY_POSITION] + ((u32)Index * arraystrides[ARRAY_POSITION]);	\
-	u8* pData = Memory_Read_U8_Ptr(iAddress);													\
+	const u8* pData = cached_arraybases[ARRAY_POSITION] + ((u32)Index * arraystrides[ARRAY_POSITION]);\
 	((float*)VertexManager::s_pCurBufferPointer)[0] = ((float)(T)(*(pData))) * posScale;		\
 	((float*)VertexManager::s_pCurBufferPointer)[1] = ((float)(T)(*(pData+1))) * posScale;		\
 	if (pVtxAttr->PosElements)																	\
@@ -151,8 +150,7 @@ void LOADERDECL Pos_ReadDirect_Float()
 }
 
 #define Pos_ReadIndex_Short(T) {																		\
-	u32 iAddress = arraybases[ARRAY_POSITION] + ((u32)Index * arraystrides[ARRAY_POSITION]);			\
-	u16* pData = Memory_Read_U16_Unswapped_Ptr(iAddress);												\
+	const u16* pData = (const u16 *)(cached_arraybases[ARRAY_POSITION] + ((u32)Index * arraystrides[ARRAY_POSITION])); \
 	((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)																			\
@@ -164,8 +162,7 @@ void LOADERDECL Pos_ReadDirect_Float()
 }
 
 #define Pos_ReadIndex_Float() {															\
-	u32 iAddress = arraybases[ARRAY_POSITION] + (Index * arraystrides[ARRAY_POSITION]); \
-	u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);								\
+	const u32* pData = (const u32 *)(cached_arraybases[ARRAY_POSITION] + (Index * arraystrides[ARRAY_POSITION])); \
 	((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*(pData));			\
 	((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(*(pData+1));			\
 	if (pVtxAttr->PosElements)															\
diff --git a/Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp b/Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp
index 1015049097..e277ca17ca 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader_TextCoord.cpp
@@ -119,9 +119,8 @@ void LOADERDECL TexCoord_ReadDirect_Float2()
 void LOADERDECL TexCoord_ReadIndex8_UByte1()	
 {
 	u8 Index = DataReadU8();
-	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
-
-	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(u8)Memory_Read_U8(iAddress) * tcScale[tcIndex];
+	const u8 *pData = cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
+	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(*pData) * tcScale[tcIndex];
 	LOG_TEX1();
 	VertexManager::s_pCurBufferPointer += 4;
 	tcIndex++;
@@ -129,9 +128,7 @@ void LOADERDECL TexCoord_ReadIndex8_UByte1()
 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);
-
+	const u8 *pData = cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
 	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(u8)(*(pData)) * tcScale[tcIndex];
 	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(u8)(*(pData+1)) * tcScale[tcIndex];
 	LOG_TEX2();
@@ -142,9 +139,9 @@ void LOADERDECL TexCoord_ReadIndex8_UByte2()
 void LOADERDECL TexCoord_ReadIndex8_Byte1()		
 {
 	u8 Index = DataReadU8();
-	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
+	const u8 *pData = cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
 
-	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s8)Memory_Read_U8(iAddress) * tcScale[tcIndex];
+	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s8)(*pData) * tcScale[tcIndex];
 	LOG_TEX1();
 	VertexManager::s_pCurBufferPointer += 4;
 	tcIndex++;
@@ -152,9 +149,7 @@ void LOADERDECL TexCoord_ReadIndex8_Byte1()
 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);
-
+	const u8 *pData = cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
 	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s8)(*(pData)) * tcScale[tcIndex];
 	((float*)VertexManager::s_pCurBufferPointer)[1] = (float)(s8)(*(pData+1)) * tcScale[tcIndex];
 	LOG_TEX2();
@@ -165,9 +160,8 @@ void LOADERDECL TexCoord_ReadIndex8_Byte2()
 void LOADERDECL TexCoord_ReadIndex8_UShort1()	
 {
 	u8 Index = DataReadU8();
-	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
-
-	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(u16)Memory_Read_U16(iAddress) * tcScale[tcIndex];
+	const u16 *pData = (const u16 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
+	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(u16)Common::swap16(*pData) * tcScale[tcIndex];
 	LOG_TEX1();
 	VertexManager::s_pCurBufferPointer += 4;
 	tcIndex++;
@@ -175,9 +169,7 @@ void LOADERDECL TexCoord_ReadIndex8_UShort1()
 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);
-
+	const u16 *pData = (const u16 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+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();
@@ -188,9 +180,8 @@ void LOADERDECL TexCoord_ReadIndex8_UShort2()
 void LOADERDECL TexCoord_ReadIndex8_Short1()	
 {
 	u8 Index = DataReadU8();
-	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
-
-	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s16)Memory_Read_U16(iAddress) * tcScale[tcIndex];
+	const u16 *pData = (const u16 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
+	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s16)Common::swap16(*(pData)) * tcScale[tcIndex];
 	LOG_TEX1();
 	VertexManager::s_pCurBufferPointer += 4;
 	tcIndex++;
@@ -198,9 +189,7 @@ void LOADERDECL TexCoord_ReadIndex8_Short1()
 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);
-
+	const u16 *pData = (const u16 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+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();
@@ -211,8 +200,8 @@ void LOADERDECL TexCoord_ReadIndex8_Short2()
 void LOADERDECL TexCoord_ReadIndex8_Float1()	
 {
 	u16 Index = DataReadU8(); 
-	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
-	((u32*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U32(iAddress);
+	const u32 *pData = (const u32 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
+	((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*(pData));
 	LOG_TEX1();
 	VertexManager::s_pCurBufferPointer += 4;
 	tcIndex++;
@@ -220,9 +209,7 @@ void LOADERDECL TexCoord_ReadIndex8_Float1()
 void LOADERDECL TexCoord_ReadIndex8_Float2()	
 {
 	u16 Index = DataReadU8(); 
-	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
-	u32* pData = Memory_Read_U32_Unswapped_Ptr(iAddress);
-
+	const u32 *pData = (const u32 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
 	((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*(pData));
 	((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(*(pData+1));
 	LOG_TEX2();
@@ -234,9 +221,8 @@ void LOADERDECL TexCoord_ReadIndex8_Float2()
 void LOADERDECL TexCoord_ReadIndex16_UByte1()	
 {
 	u16 Index = DataReadU16(); 
-	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
-
-	((float*)VertexManager::s_pCurBufferPointer)[0] =  (float)(u8)Memory_Read_U8(iAddress) * tcScale[tcIndex];
+	const u8 *pData = cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
+	((float*)VertexManager::s_pCurBufferPointer)[0] =  (float)(u8)(*(pData)) * tcScale[tcIndex];
 	LOG_TEX1();
 	VertexManager::s_pCurBufferPointer += 4;
 	tcIndex++;
@@ -244,9 +230,7 @@ void LOADERDECL TexCoord_ReadIndex16_UByte1()
 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);
-
+	const u8 *pData = cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
 	((float*)VertexManager::s_pCurBufferPointer)[0] =  (float)(u8)(*(pData)) * tcScale[tcIndex];
 	((float*)VertexManager::s_pCurBufferPointer)[1] =  (float)(u8)(*(pData+1)) * tcScale[tcIndex];
 	LOG_TEX2();
@@ -257,9 +241,8 @@ void LOADERDECL TexCoord_ReadIndex16_UByte2()
 void LOADERDECL TexCoord_ReadIndex16_Byte1()	
 {
 	u16 Index = DataReadU16(); 
-	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
-
-	((float*)VertexManager::s_pCurBufferPointer)[0] =  (float)(s8)Memory_Read_U8(iAddress) * tcScale[tcIndex];
+	const u8 *pData = cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
+	((float*)VertexManager::s_pCurBufferPointer)[0] =  (float)(s8)(*(pData)) * tcScale[tcIndex];
 	LOG_TEX1();
 	VertexManager::s_pCurBufferPointer += 4;
 	tcIndex++;
@@ -267,9 +250,7 @@ void LOADERDECL TexCoord_ReadIndex16_Byte1()
 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);
-
+	const u8 *pData = cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
 	((float*)VertexManager::s_pCurBufferPointer)[0] =  (float)(s8)(*(pData)) * tcScale[tcIndex];
 	((float*)VertexManager::s_pCurBufferPointer)[1] =  (float)(s8)(*(pData+1)) * tcScale[tcIndex];
 	LOG_TEX2();
@@ -280,9 +261,8 @@ void LOADERDECL TexCoord_ReadIndex16_Byte2()
 void LOADERDECL TexCoord_ReadIndex16_UShort1()	
 {
 	u16 Index = DataReadU16(); 
-	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
-
-	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(u16)Memory_Read_U16(iAddress) * tcScale[tcIndex];
+	const u16* pData = (const u16 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
+	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(u16)Common::swap16(*(pData)) * tcScale[tcIndex];
 	LOG_TEX1();
 	VertexManager::s_pCurBufferPointer += 4;
 	tcIndex++;
@@ -290,9 +270,7 @@ void LOADERDECL TexCoord_ReadIndex16_UShort1()
 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);
-
+	const u16* pData = (const u16 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+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();
@@ -303,9 +281,8 @@ void LOADERDECL TexCoord_ReadIndex16_UShort2()
 void LOADERDECL TexCoord_ReadIndex16_Short1()	
 {
 	u16 Index = DataReadU16(); 
-	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
-
-	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s16)Memory_Read_U16(iAddress) * tcScale[tcIndex];
+	const u16 *pData = (const u16 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
+	((float*)VertexManager::s_pCurBufferPointer)[0] = (float)(s16)Common::swap16(*pData) * tcScale[tcIndex];
 	LOG_TEX1();
 	VertexManager::s_pCurBufferPointer += 4;
 	tcIndex++;
@@ -313,9 +290,7 @@ void LOADERDECL TexCoord_ReadIndex16_Short1()
 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);
-
+	const u16 *pData = (const u16 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+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();
@@ -326,9 +301,8 @@ void LOADERDECL TexCoord_ReadIndex16_Short2()
 void LOADERDECL TexCoord_ReadIndex16_Float1()	
 {
 	u16 Index = DataReadU16(); 
-	u32 iAddress = arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]);
-
-	((u32*)VertexManager::s_pCurBufferPointer)[0] = Memory_Read_U32(iAddress);
+	const u32 *pData = (const u32 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
+	((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*(pData));
 	LOG_TEX1();
 	VertexManager::s_pCurBufferPointer += 4;
 	tcIndex++;
@@ -336,9 +310,7 @@ void LOADERDECL TexCoord_ReadIndex16_Float1()
 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);
-
+	const u32 *pData = (const u32 *)(cached_arraybases[ARRAY_TEXCOORD0+tcIndex] + (Index * arraystrides[ARRAY_TEXCOORD0+tcIndex]));
 	((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(*(pData));
 	((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(*(pData+1));
 	LOG_TEX2();
diff --git a/Source/Plugins/Plugin_VideoOGL/Src/VertexLoaderManager.cpp b/Source/Plugins/Plugin_VideoOGL/Src/VertexLoaderManager.cpp
index 897e387207..88b1ec18b7 100644
--- a/Source/Plugins/Plugin_VideoOGL/Src/VertexLoaderManager.cpp
+++ b/Source/Plugins/Plugin_VideoOGL/Src/VertexLoaderManager.cpp
@@ -17,6 +17,7 @@
 
 #include <map>
 
+#include "VideoCommon.h"
 #include "Statistics.h"
 
 #include "VertexShaderManager.h"
@@ -145,7 +146,8 @@ void LoadCPReg(u32 sub_cmd, u32 value)
 
 	// Pointers to vertex arrays in GC RAM
 	case 0xA0:
-		arraybases[sub_cmd & 0xF] = value & 0xFFFFFFFF;  // huh, why the mask?
+		arraybases[sub_cmd & 0xF] = value;
+		cached_arraybases[sub_cmd & 0xF] = Memory_GetPtr(value);
 		break;
 
 	case 0xB0:
@@ -153,3 +155,11 @@ void LoadCPReg(u32 sub_cmd, u32 value)
 		break;
 	}
 }
+
+void RecomputeCachedArraybases()
+{
+	for (int i = 0; i < 16; i++)
+	{
+		cached_arraybases[i] = Memory_GetPtr(arraybases[i]);
+	}
+}
diff --git a/Source/Plugins/Plugin_VideoOGL/Src/main.cpp b/Source/Plugins/Plugin_VideoOGL/Src/main.cpp
index 119589ceff..4a24ad52fe 100644
--- a/Source/Plugins/Plugin_VideoOGL/Src/main.cpp
+++ b/Source/Plugins/Plugin_VideoOGL/Src/main.cpp
@@ -253,8 +253,11 @@ void DoState(unsigned char **ptr, int mode) {
     VideoCommon_DoState(p);
     
     // Refresh state.
-    if (mode == PointerWrap::MODE_READ)
+	if (mode == PointerWrap::MODE_READ)
+	{
         BPReload();
+		RecomputeCachedArraybases();
+	}
 }
 
 // This is called after Video_Initialize() from the Core