2 fixes one for dx9 and one for software plugin.

some little optimization to normal loading.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6898 8ced0084-cf51-0410-be5f-012b33b47a6e

1 files changed, 128 insertions, 88 deletions

diff --git a/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp b/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp
index dcfaf21779..c49a23031b 100644
--- a/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp
+++ b/Source/Core/VideoCommon/Src/VertexLoader_Normal.cpp
@@ -21,14 +21,16 @@
 #include "VertexLoader_Normal.h"
 #include "VertexManagerBase.h"
 #include "CPUDetect.h"
+#include <cmath>
 
-#if _M_SSE >= 0x301 && !(defined __GNUC__ && !defined __SSSE3__)
+#if _M_SSE >= 0x401
+#include <smmintrin.h>
+#include <emmintrin.h>
+#elif _M_SSE >= 0x301 && !(defined __GNUC__ && !defined __SSSE3__)
 #include <tmmintrin.h>
 #endif
 
-#define LOG_NORM8()  // PRIM_LOG("norm: %f %f %f, ", ((float*)VertexManager::s_pCurBufferPointer)[-3], ((float*)VertexManager::s_pCurBufferPointer)[-2], ((float*)VertexManager::s_pCurBufferPointer)[-1]);
-#define LOG_NORM16() // PRIM_LOG("norm: %f %f %f, ", ((float*)VertexManager::s_pCurBufferPointer)[-3], ((float*)VertexManager::s_pCurBufferPointer)[-2], ((float*)VertexManager::s_pCurBufferPointer)[-1]);
-#define LOG_NORMF()  // PRIM_LOG("norm: %f %f %f, ", ((float*)VertexManager::s_pCurBufferPointer)[-3], ((float*)VertexManager::s_pCurBufferPointer)[-2], ((float*)VertexManager::s_pCurBufferPointer)[-1]);
+#define LOG_NORM()  // PRIM_LOG("norm: %f %f %f, ", ((float*)VertexManager::s_pCurBufferPointer)[-3], ((float*)VertexManager::s_pCurBufferPointer)[-2], ((float*)VertexManager::s_pCurBufferPointer)[-1]);
 
 VertexLoader_Normal::Set VertexLoader_Normal::m_Table[NUM_NRM_TYPE][NUM_NRM_INDICES][NUM_NRM_ELEMENTS][NUM_NRM_FORMAT];
 
@@ -118,98 +120,156 @@ TPipelineFunction VertexLoader_Normal::GetFunction(unsigned int _type, unsigned
 #define S16FRAC 0.00006103515625f; // 1.0f / (1U << 14)
 // --- Direct ---
 
-
-inline void ReadDirectS8()
+inline void ReadIndirectS8x3(const s8* pData)
 {
-	((float*)VertexManager::s_pCurBufferPointer)[0] = DataReadS8() * S8FRAC;
-    ((float*)VertexManager::s_pCurBufferPointer)[1] = DataReadS8() * S8FRAC;
-    ((float*)VertexManager::s_pCurBufferPointer)[2] = DataReadS8() * S8FRAC;
+	((float*)VertexManager::s_pCurBufferPointer)[0] = pData[0] * S8FRAC;
+    ((float*)VertexManager::s_pCurBufferPointer)[1] = pData[1] * S8FRAC;
+    ((float*)VertexManager::s_pCurBufferPointer)[2] = pData[2] * S8FRAC;
 	VertexManager::s_pCurBufferPointer += 12;
-	LOG_NORM8();
-}
-
-inline void ReadDirectS16()
-{
-	((float*)VertexManager::s_pCurBufferPointer)[0] = ((s16)DataReadU16()) * S16FRAC;
-    ((float*)VertexManager::s_pCurBufferPointer)[1] = ((s16)DataReadU16()) * S16FRAC;
-    ((float*)VertexManager::s_pCurBufferPointer)[2] = ((s16)DataReadU16()) * S16FRAC;
-    VertexManager::s_pCurBufferPointer += 12;
-    LOG_NORM16()
+	LOG_NORM();
 }
 
-inline void ReadDirectFloat()
-{
-	((u32*)VertexManager::s_pCurBufferPointer)[0] = DataReadU32();
-    ((u32*)VertexManager::s_pCurBufferPointer)[1] = DataReadU32();
-    ((u32*)VertexManager::s_pCurBufferPointer)[2] = DataReadU32();
-    VertexManager::s_pCurBufferPointer += 12;
-    LOG_NORMF()
-}
-
-inline void ReadIndirectS8(const s8* pData)
+inline void ReadIndirectS8x9(const s8* pData)
 {
 	((float*)VertexManager::s_pCurBufferPointer)[0] = pData[0] * S8FRAC;
     ((float*)VertexManager::s_pCurBufferPointer)[1] = pData[1] * S8FRAC;
     ((float*)VertexManager::s_pCurBufferPointer)[2] = pData[2] * S8FRAC;
-	VertexManager::s_pCurBufferPointer += 12;
-	LOG_NORM8();
+	LOG_NORM();
+	((float*)VertexManager::s_pCurBufferPointer)[3] = pData[3] * S8FRAC;
+    ((float*)VertexManager::s_pCurBufferPointer)[4] = pData[4] * S8FRAC;
+    ((float*)VertexManager::s_pCurBufferPointer)[5] = pData[5] * S8FRAC;
+	LOG_NORM();
+	((float*)VertexManager::s_pCurBufferPointer)[6] = pData[6] * S8FRAC;
+    ((float*)VertexManager::s_pCurBufferPointer)[7] = pData[7] * S8FRAC;
+    ((float*)VertexManager::s_pCurBufferPointer)[8] = pData[8] * S8FRAC;
+	LOG_NORM();
+	VertexManager::s_pCurBufferPointer += 36;	
 }
 
-inline void ReadIndirectS16(const u16* pData)
+inline void ReadIndirectS16x3(const u16* pData)
 {
 	((float*)VertexManager::s_pCurBufferPointer)[0] = ((s16)Common::swap16(pData[0])) * S16FRAC;
     ((float*)VertexManager::s_pCurBufferPointer)[1] = ((s16)Common::swap16(pData[1])) * S16FRAC;
     ((float*)VertexManager::s_pCurBufferPointer)[2] = ((s16)Common::swap16(pData[2])) * S16FRAC;
 	VertexManager::s_pCurBufferPointer += 12;
-    LOG_NORM16()
+    LOG_NORM()
+}
+
+inline void ReadIndirectS16x9(const u16* pData)
+{
+	((float*)VertexManager::s_pCurBufferPointer)[0] = ((s16)Common::swap16(pData[0])) * S16FRAC;
+    ((float*)VertexManager::s_pCurBufferPointer)[1] = ((s16)Common::swap16(pData[1])) * S16FRAC;
+    ((float*)VertexManager::s_pCurBufferPointer)[2] = ((s16)Common::swap16(pData[2])) * S16FRAC;
+	LOG_NORM()
+	((float*)VertexManager::s_pCurBufferPointer)[3] = ((s16)Common::swap16(pData[3])) * S16FRAC;
+    ((float*)VertexManager::s_pCurBufferPointer)[4] = ((s16)Common::swap16(pData[4])) * S16FRAC;
+    ((float*)VertexManager::s_pCurBufferPointer)[5] = ((s16)Common::swap16(pData[5])) * S16FRAC;
+	LOG_NORM()
+	((float*)VertexManager::s_pCurBufferPointer)[6] = ((s16)Common::swap16(pData[6])) * S16FRAC;
+    ((float*)VertexManager::s_pCurBufferPointer)[7] = ((s16)Common::swap16(pData[7])) * S16FRAC;
+    ((float*)VertexManager::s_pCurBufferPointer)[8] = ((s16)Common::swap16(pData[8])) * S16FRAC;
+	LOG_NORM()
+	VertexManager::s_pCurBufferPointer += 36;
 }
 
-inline void ReadIndirectFloat(const u32* pData)
+inline void ReadIndirectFloatx3(const u32* pData)
 {
 	((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(pData[0]);	
     ((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(pData[1]);
     ((u32*)VertexManager::s_pCurBufferPointer)[2] = Common::swap32(pData[2]);
     VertexManager::s_pCurBufferPointer += 12;
-    LOG_NORMF();
+    LOG_NORM();
+}
+
+inline void ReadIndirectFloatx9(const u32* pData)
+{
+	((u32*)VertexManager::s_pCurBufferPointer)[0] = Common::swap32(pData[0]);	
+    ((u32*)VertexManager::s_pCurBufferPointer)[1] = Common::swap32(pData[1]);
+    ((u32*)VertexManager::s_pCurBufferPointer)[2] = Common::swap32(pData[2]);
+	LOG_NORM();
+	((u32*)VertexManager::s_pCurBufferPointer)[3] = Common::swap32(pData[3]);	
+    ((u32*)VertexManager::s_pCurBufferPointer)[4] = Common::swap32(pData[4]);
+    ((u32*)VertexManager::s_pCurBufferPointer)[5] = Common::swap32(pData[5]);
+	LOG_NORM();
+	((u32*)VertexManager::s_pCurBufferPointer)[6] = Common::swap32(pData[6]);	
+    ((u32*)VertexManager::s_pCurBufferPointer)[7] = Common::swap32(pData[7]);
+    ((u32*)VertexManager::s_pCurBufferPointer)[8] = Common::swap32(pData[8]);
+	LOG_NORM();
+    VertexManager::s_pCurBufferPointer += 36;    
+}
+
+inline void ReadDirectS8x3()
+{
+	const s8* Source = (const s8*)DataGetPosition();
+	ReadIndirectS8x3(Source);
+	DataSkip(3);
+}
+
+inline void ReadDirectS8x9()
+{
+	const s8* Source = (const s8*)DataGetPosition();
+	ReadIndirectS8x9(Source);
+	DataSkip(9);
+}
+
+inline void ReadDirectS16x3()
+{
+	const u16* Source = (const u16*)DataGetPosition();
+	ReadIndirectS16x3(Source);
+	DataSkip(6);
+}
+
+inline void ReadDirectS16x9()
+{
+	const u16* Source = (const u16*)DataGetPosition();
+	ReadIndirectS16x9(Source);
+	DataSkip(18);
 }
 
+inline void ReadDirectFloatx3()
+{
+	const u32* Source = (const u32*)DataGetPosition();
+	ReadIndirectFloatx3(Source);
+	DataSkip(12);	
+}
+
+inline void ReadDirectFloatx9()
+{
+	const u32* Source = (const u32*)DataGetPosition();
+	ReadIndirectFloatx9(Source);
+	DataSkip(36);	
+}
+
+
+
 void LOADERDECL VertexLoader_Normal::Normal_DirectByte()
 {
-    ReadDirectS8();
+    ReadDirectS8x3();
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_DirectShort()
 {
-	ReadDirectS16();
+	ReadDirectS16x3();
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_DirectFloat()
 {
-    ReadDirectFloat();
+    ReadDirectFloatx3();
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_DirectByte3()
 {
-    for (int i = 0; i < 3; i++)
-    {
-        ReadDirectS8();
-    }
+    ReadDirectS8x9();    
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_DirectShort3()
 {
-    for (int i = 0; i < 3; i++)
-    {
-        ReadDirectS16();
-    }
+    ReadDirectS16x9();    
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_DirectFloat3()
 {
-    for (int i = 0; i < 3; i++)
-    {
-        ReadDirectFloat();
-    }
+    ReadDirectFloatx9();    
 }
 
 
@@ -219,51 +279,42 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Byte()
 {
     u8 Index = DataReadU8();
 	const s8* pData = (const s8 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
-	ReadIndirectS8(pData);
+	ReadIndirectS8x3(pData);
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_Index8_Short()
 {
     u8 Index = DataReadU8();
 	const u16* pData = (const u16 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
-	ReadIndirectS16(pData);
+	ReadIndirectS16x3(pData);
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_Index8_Float()
 {
     u8 Index = DataReadU8();
 	const u32* pData = (const u32 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
-	ReadIndirectFloat(pData);	
+	ReadIndirectFloatx3(pData);	
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_Index8_Byte3_Indices1()
 {
     u8 Index = DataReadU8();
 	const s8* pData = (const s8 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
-    for (int i = 0; i < 3; i++)
-	{
-		ReadIndirectS8((const s8*)(&pData[3 * i]));
-    }
+    ReadIndirectS8x9(pData);
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_Index8_Short3_Indices1()
 {
     u8 Index = DataReadU8();
 	const u16* pData = (const u16 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
-    for (int i = 0; i < 3; i++)
-	{
-		ReadIndirectS16((const u16*)(&pData[3 * i]));
-    }    
+    ReadIndirectS16x9(pData);        
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_Index8_Float3_Indices1()
 {
     u8 Index = DataReadU8();	
 	const u32* pData = (const u32 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
-    for (int i = 0; i < 3; i++)
-	{
-		ReadIndirectFloat((const u32*)(&pData[3 * i]));
-    }    
+    ReadIndirectFloatx9(pData);        
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_Index8_Byte3_Indices3()
@@ -272,7 +323,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Byte3_Indices3()
 	{
         u8 Index = DataReadU8();
 		const s8* pData = (const s8*)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 1*3*i);
-        ReadIndirectS8(pData);
+        ReadIndirectS8x3(pData);
     }
 }
 
@@ -283,7 +334,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Short3_Indices3()
 	{
         u8 Index = DataReadU8();
 		const u16* pData = (const u16 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 2*3*i);
-        ReadIndirectS16(pData);
+        ReadIndirectS16x3(pData);
     }
 }
 
@@ -293,7 +344,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index8_Float3_Indices3()
 	{
         u8 Index = DataReadU8();
 		const u32* pData = (const u32 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 4*3*i);
-        ReadIndirectFloat(pData);
+        ReadIndirectFloatx3(pData);
     }    
 }
 
@@ -305,53 +356,42 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Byte()
 {
     u16 Index = DataReadU16();
 	const s8* pData = (const s8 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
-	ReadIndirectS8(pData);
+	ReadIndirectS8x3(pData);
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_Index16_Short()
 {
     u16 Index = DataReadU16();
 	const u16* pData = (const u16 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
-	ReadIndirectS16(pData);
+	ReadIndirectS16x3(pData);
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_Index16_Float()
 {
     u16 Index = DataReadU16();
 	const u32* pData = (const u32 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
-    ReadIndirectFloat(pData);
+    ReadIndirectFloatx3(pData);
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_Index16_Byte3_Indices1()
 {
     u16 Index = DataReadU16();
 	const s8* pData = (const s8 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
-    for (int i = 0; i < 3; i++)
-	{
-		ReadIndirectS8((const s8 *)(&pData[3 * i]));        
-    }
+    ReadIndirectS8x9(pData);            
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_Index16_Short3_Indices1()
 {
     u16 Index = DataReadU16();
 	const u16* pData = (const u16 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
-
-    for (int i = 0; i < 3; i++)
-    {
-        ReadIndirectS16((const u16 *)(&pData[3 * i]));
-    }
+	ReadIndirectS16x9(pData);
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_Index16_Float3_Indices1()
 {
     u16 Index = DataReadU16();
 	const u32* pData = (const u32 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]));
-
-    for (int i = 0; i < 3; i++)
-    {
-        ReadIndirectFloat((const u32 *)(&pData[3 * i]));
-    }
+	ReadIndirectFloatx9(pData);
 }
 
 void LOADERDECL VertexLoader_Normal::Normal_Index16_Byte3_Indices3()
@@ -360,7 +400,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Byte3_Indices3()
 	{
         u16 Index = DataReadU16();
 		const s8* pData = (const s8*)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 1*3*i);
-        ReadIndirectS8(pData);
+        ReadIndirectS8x3(pData);
     }    
 }
 
@@ -370,7 +410,7 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Short3_Indices3()
     {
         u16 Index = DataReadU16();
 		const u16* pData = (const u16 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 2*3*i);
-        ReadIndirectS16(pData);
+        ReadIndirectS16x3(pData);
     }
 }
 
@@ -380,6 +420,6 @@ void LOADERDECL VertexLoader_Normal::Normal_Index16_Float3_Indices3()
     {
         u16 Index = DataReadU16();
 		const u32* pData = (const u32 *)(cached_arraybases[ARRAY_NORMAL] + (Index * arraystrides[ARRAY_NORMAL]) + 4*3*i);
-        ReadIndirectFloat(pData);
+        ReadIndirectFloatx3(pData);
     }    
 }