D3D: minor vs constant-setting optimization, remove a stupid memcpy that doesn't do anything, don't see much benefit though :/ At least the PIX logs will be cleaner.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@4301 8ced0084-cf51-0410-be5f-012b33b47a6e

9 files changed, 81 insertions, 30 deletions

diff --git a/Source/Core/VideoCommon/Src/VertexShaderManager.cpp b/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
index c90e8531f9..3a5936e3fe 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
+++ b/Source/Core/VideoCommon/Src/VertexShaderManager.cpp
@@ -100,9 +100,8 @@ void VertexShaderManager::SetConstants()
         int startn = nNormalMatricesChanged[0] / 3;
         int endn = (nNormalMatricesChanged[1] + 2) / 3;
         const float *pnstart = (const float*)&xfmem[XFMEM_NORMALMATRICES+3*startn];
-        for (int i = startn; i < endn; ++i, pnstart += 3)
-            SetVSConstant4fv(C_NORMALMATRICES + i, pnstart);  // looks like we're reading one too much..
-        nNormalMatricesChanged[0] = nNormalMatricesChanged[1] = -1;
+        SetMultiVSConstant3fv(C_NORMALMATRICES + startn, endn - startn, pnstart);
+		nNormalMatricesChanged[0] = nNormalMatricesChanged[1] = -1;
     }
 
     if (nPostTransformMatricesChanged[0] >= 0)
@@ -165,9 +164,7 @@ void VertexShaderManager::SetConstants()
         const float *norm = (const float *)xfmem + XFMEM_NORMALMATRICES + 3 * (MatrixIndexA.PosNormalMtxIdx & 31);
 
         SetMultiVSConstant4fv(C_POSNORMALMATRIX, 3, pos);
-        SetVSConstant4fv(C_POSNORMALMATRIX+3, norm);
-        SetVSConstant4fv(C_POSNORMALMATRIX+4, norm + 3);
-        SetVSConstant4fv(C_POSNORMALMATRIX+5, norm + 6);
+		SetMultiVSConstant3fv(C_POSNORMALMATRIX + 3, 3, norm);
 	}
 
     if (bTexMatricesChanged[0])
diff --git a/Source/Core/VideoCommon/Src/VertexShaderManager.h b/Source/Core/VideoCommon/Src/VertexShaderManager.h
index 619c3af5de..be7a506716 100644
--- a/Source/Core/VideoCommon/Src/VertexShaderManager.h
+++ b/Source/Core/VideoCommon/Src/VertexShaderManager.h
@@ -46,6 +46,7 @@ public:
 
 void SetVSConstant4f(int const_number, float f1, float f2, float f3, float f4);
 void SetVSConstant4fv(int const_number, const float *f);
+void SetMultiVSConstant3fv(int const_number, int count, const float *f);
 void SetMultiVSConstant4fv(int const_number, int count, const float *f);
 
 #endif // _VERTEXSHADERMANAGER_H
diff --git a/Source/Plugins/Plugin_VideoDX9/Src/D3DBase.cpp b/Source/Plugins/Plugin_VideoDX9/Src/D3DBase.cpp
index 896eebb3a2..b37a4e4c29 100644
--- a/Source/Plugins/Plugin_VideoDX9/Src/D3DBase.cpp
+++ b/Source/Plugins/Plugin_VideoDX9/Src/D3DBase.cpp
@@ -58,6 +58,7 @@ static DWORD m_RenderStates[MaxRenderStates];
 static DWORD m_TextureStageStates[MaxTextureStages][MaxTextureTypes];
 static DWORD m_SamplerStates[MaxSamplerSize][MaxSamplerTypes];
 LPDIRECT3DBASETEXTURE9 m_Textures[16];
+LPDIRECT3DVERTEXDECLARATION9 m_VtxDecl;
 
 void Enumerate();
 
@@ -256,25 +257,18 @@ HRESULT Create(int adapter, HWND wnd, bool _fullscreen, int _resolution, int aa_
 		adapter, 
 		D3DDEVTYPE_HAL, 
 		wnd,
-		D3DCREATE_HARDWARE_VERTEXPROCESSING,
+		D3DCREATE_HARDWARE_VERTEXPROCESSING, // | D3DCREATE_PUREDEVICE,  doesn't seem to make a difference
 		&d3dpp, &dev)))
 	{
-		MessageBox(wnd,
-			_T("Sorry, but it looks like your 3D accelerator is too old,\n")
-			_T("or doesn't support features that Dolphin requires.\n")
-			_T("Falling back to software vertex processing.\n"),
-			_T("Dolphin Direct3D plugin"), MB_OK | MB_ICONERROR);
 		if (FAILED(D3D->CreateDevice( 
 			adapter, 
 			D3DDEVTYPE_HAL, 
 			wnd,
-			D3DCREATE_SOFTWARE_VERTEXPROCESSING | D3DCREATE_MULTITHREADED,
-			// |D3DCREATE_MULTITHREADED /* | D3DCREATE_PUREDEVICE*/,
-			//D3DCREATE_SOFTWARE_VERTEXPROCESSING ,
+			D3DCREATE_SOFTWARE_VERTEXPROCESSING,
 			&d3dpp, &dev)))
 		{
 			MessageBox(wnd,
-				_T("Software VP failed too. Upgrade your graphics card."),
+				_T("Failed to initialize Direct3D."),
 				_T("Dolphin Direct3D plugin"), MB_OK | MB_ICONERROR);
 			return E_FAIL;
 		}
@@ -442,6 +436,7 @@ void ApplyCachedState()
 	// so no stale state is around.
 	memset(m_Textures, 0, sizeof(m_Textures));
 	memset(m_TextureStageStates, 0xFF, sizeof(m_TextureStageStates));
+	m_VtxDecl = NULL;
 }
 
 void SetTexture(DWORD Stage, LPDIRECT3DBASETEXTURE9 pTexture)
@@ -481,4 +476,17 @@ void SetSamplerState(DWORD Sampler, D3DSAMPLERSTATETYPE Type, DWORD Value)
 	}
 }
 
+void SetVertexDeclaration(LPDIRECT3DVERTEXDECLARATION9 decl)
+{
+	if (!decl) {
+		m_VtxDecl = NULL;
+		return;
+	}
+	if (decl != m_VtxDecl)
+	{
+		D3D::dev->SetVertexDeclaration(decl);
+		m_VtxDecl = decl;
+	}
+}
+
 }  // namespace
diff --git a/Source/Plugins/Plugin_VideoDX9/Src/D3DBase.h b/Source/Plugins/Plugin_VideoDX9/Src/D3DBase.h
index 6fcbe675bf..557eef1d9e 100644
--- a/Source/Plugins/Plugin_VideoDX9/Src/D3DBase.h
+++ b/Source/Plugins/Plugin_VideoDX9/Src/D3DBase.h
@@ -73,6 +73,7 @@ void SetTexture(DWORD Stage, IDirect3DBaseTexture9 *pTexture);
 void SetRenderState(D3DRENDERSTATETYPE State, DWORD Value);
 void SetTextureStageState(DWORD Stage, D3DTEXTURESTAGESTATETYPE Type, DWORD Value);
 void SetSamplerState(DWORD Sampler, D3DSAMPLERSTATETYPE Type, DWORD Value);
+void SetVertexDeclaration(LPDIRECT3DVERTEXDECLARATION9 decl);
 void ApplyCachedState();
 
 // Utility functions for vendor specific hacks. So far, just the one.
diff --git a/Source/Plugins/Plugin_VideoDX9/Src/D3DUtil.cpp b/Source/Plugins/Plugin_VideoDX9/Src/D3DUtil.cpp
index 600c470fc9..a1b30a3dc2 100644
--- a/Source/Plugins/Plugin_VideoDX9/Src/D3DUtil.cpp
+++ b/Source/Plugins/Plugin_VideoDX9/Src/D3DUtil.cpp
@@ -227,6 +227,7 @@ void CD3DFont::SetRenderStates()
 	dev->SetPixelShader(0);
 	dev->SetVertexShader(0);
 	
+	D3D::SetVertexDeclaration(NULL);  // throw away cached vtx decl
 	dev->SetFVF(D3DFVF_FONT2DVERTEX);
 
 	for (int i = 0; i < 6; i++)
@@ -366,6 +367,7 @@ void quad2d(float x1, float y1, float x2, float y2, u32 color, float u1, float v
 	dev->SetVertexShader(0);
 	dev->SetVertexDeclaration(0);
 	
+	D3D::SetVertexDeclaration(NULL);  // throw away cached vtx decl
 	dev->SetFVF(D3DFVF_XYZRHW | D3DFVF_DIFFUSE | D3DFVF_TEX1);
 	dev->DrawPrimitiveUP(D3DPT_TRIANGLEFAN, 2, coords, sizeof(Q2DVertex));
 	
diff --git a/Source/Plugins/Plugin_VideoDX9/Src/NativeVertexFormat.cpp b/Source/Plugins/Plugin_VideoDX9/Src/NativeVertexFormat.cpp
index f4891bdcd5..1717861db2 100644
--- a/Source/Plugins/Plugin_VideoDX9/Src/NativeVertexFormat.cpp
+++ b/Source/Plugins/Plugin_VideoDX9/Src/NativeVertexFormat.cpp
@@ -162,7 +162,7 @@ void D3DVertexFormat::Initialize(const PortableVertexDeclaration &_vtx_decl)
 void D3DVertexFormat::SetupVertexPointers() const
 {
 	if (d3d_decl)
-		D3D::dev->SetVertexDeclaration(d3d_decl);
+		D3D::SetVertexDeclaration(d3d_decl);
 	else
 		ERROR_LOG(VIDEO, "invalid d3d decl");
 }
\ No newline at end of file
diff --git a/Source/Plugins/Plugin_VideoDX9/Src/VertexManager.cpp b/Source/Plugins/Plugin_VideoDX9/Src/VertexManager.cpp
index e89cfc56aa..516bcb0be5 100644
--- a/Source/Plugins/Plugin_VideoDX9/Src/VertexManager.cpp
+++ b/Source/Plugins/Plugin_VideoDX9/Src/VertexManager.cpp
@@ -99,6 +99,7 @@ void Shutdown()
 
 void CreateDeviceObjects()
 {
+	
 }
 
 void DestroyDeviceObjects()
@@ -140,12 +141,8 @@ void AddVertices(int _primitive, int _numVertices)
 
 	if (collection != type)
 	{
-		//We are NOT collecting the right type.
+		// We are NOT collecting the right type.
 		Flush();
-
-		// Copy the extra verts that we lost.
-		memcpy(s_pCurBufferPointer, fakeVBuffer, _numVertices * g_nativeVertexFmt->GetVertexStride());
-
 		collection = type;
 		u16 *ptr = 0;
 		if (type != C_POINTS)
diff --git a/Source/Plugins/Plugin_VideoDX9/Src/VertexShaderCache.cpp b/Source/Plugins/Plugin_VideoDX9/Src/VertexShaderCache.cpp
index 5beace957b..5409c217b3 100644
--- a/Source/Plugins/Plugin_VideoDX9/Src/VertexShaderCache.cpp
+++ b/Source/Plugins/Plugin_VideoDX9/Src/VertexShaderCache.cpp
@@ -43,11 +43,11 @@ void SetVSConstant4f(int const_number, float f1, float f2, float f3, float f4)
 		lastVSconstants[const_number][3] != f4)
 	{
 		const float f[4] = {f1, f2, f3, f4};
-		D3D::dev->SetVertexShaderConstantF(const_number, f, 1);
 		lastVSconstants[const_number][0] = f1;
 		lastVSconstants[const_number][1] = f2;
 		lastVSconstants[const_number][2] = f3;
 		lastVSconstants[const_number][3] = f4;
+		D3D::dev->SetVertexShaderConstantF(const_number, lastVSconstants[const_number], 1);
 	}
 }
 
@@ -58,11 +58,37 @@ void SetVSConstant4fv(int const_number, const float *f)
 		lastVSconstants[const_number][2] != f[2] ||
 		lastVSconstants[const_number][3] != f[3])
 	{
-		D3D::dev->SetVertexShaderConstantF(const_number, f, 1);
 		lastVSconstants[const_number][0] = f[0];
 		lastVSconstants[const_number][1] = f[1];
 		lastVSconstants[const_number][2] = f[2];
 		lastVSconstants[const_number][3] = f[3];
+		D3D::dev->SetVertexShaderConstantF(const_number, lastVSconstants[const_number], 1);
+	}
+}
+
+void SetMultiVSConstant3fv(int const_number, int count, const float *f)
+{
+	bool change = false;
+	for (int i = 0; i < count; i++)
+	{
+		if (lastVSconstants[const_number + i][0] != f[0 + i*3] || 
+			lastVSconstants[const_number + i][1] != f[1 + i*3] ||
+			lastVSconstants[const_number + i][2] != f[2 + i*3])
+		{
+			change = true;
+			break;
+		}
+	}
+	if (change)
+	{
+		for (int i = 0; i < count; i++)
+		{
+			lastVSconstants[const_number + i][0] = f[0 + i*3];
+			lastVSconstants[const_number + i][1] = f[1 + i*3];
+			lastVSconstants[const_number + i][2] = f[2 + i*3];
+			lastVSconstants[const_number + i][3] = 0.0f;
+		}
+		D3D::dev->SetVertexShaderConstantF(const_number, lastVSconstants[const_number], count);
 	}
 }
 
@@ -82,13 +108,14 @@ void SetMultiVSConstant4fv(int const_number, int count, const float *f)
 	}
 	if (change)
 	{
-		D3D::dev->SetVertexShaderConstantF(const_number, f, count);
-		for (int i = 0; i < count; i++) {
+		for (int i = 0; i < count; i++)
+		{
 			lastVSconstants[const_number + i][0] = f[0 + i*4];
 			lastVSconstants[const_number + i][1] = f[1 + i*4];
 			lastVSconstants[const_number + i][2] = f[2 + i*4];
 			lastVSconstants[const_number + i][3] = f[3 + i*4];
 		}
+		D3D::dev->SetVertexShaderConstantF(const_number, lastVSconstants[const_number], count);
 	}
 }
 
diff --git a/Source/Plugins/Plugin_VideoOGL/Src/VertexShaderCache.cpp b/Source/Plugins/Plugin_VideoOGL/Src/VertexShaderCache.cpp
index defa2ca440..2d6cb30ddf 100644
--- a/Source/Plugins/Plugin_VideoOGL/Src/VertexShaderCache.cpp
+++ b/Source/Plugins/Plugin_VideoOGL/Src/VertexShaderCache.cpp
@@ -50,12 +50,11 @@ void SetVSConstant4f(int const_number, float f1, float f2, float f3, float f4)
 		lastVSconstants[const_number][2] != f3 ||
 		lastVSconstants[const_number][3] != f4)
 	{
-		const float f[4] = {f1, f2, f3, f4};
-		glProgramEnvParameter4fARB(GL_VERTEX_PROGRAM_ARB, const_number, f1, f2, f3, f4);
 		lastVSconstants[const_number][0] = f1;
 		lastVSconstants[const_number][1] = f2;
 		lastVSconstants[const_number][2] = f3;
 		lastVSconstants[const_number][3] = f4;
+		glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, const_number, lastVSconstants[const_number]);
 	}
 }
 
@@ -66,11 +65,11 @@ void SetVSConstant4fv(int const_number, const float *f)
 		lastVSconstants[const_number][2] != f[2] ||
 		lastVSconstants[const_number][3] != f[3])
 	{
-		glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, const_number, f);
 		lastVSconstants[const_number][0] = f[0];
 		lastVSconstants[const_number][1] = f[1];
 		lastVSconstants[const_number][2] = f[2];
 		lastVSconstants[const_number][3] = f[3];
+		glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, const_number, lastVSconstants[const_number]);
 	}
 }
 
@@ -83,15 +82,34 @@ void SetMultiVSConstant4fv(int const_number, int count, const float *f)
 			lastVSconstants[const_number + i][2] != f[2 + i*4] ||
 			lastVSconstants[const_number + i][3] != f[3 + i*4])
 		{
-			glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, const_number + i, f + i * 4);
 			lastVSconstants[const_number + i][0] = f[0 + i*4];
 			lastVSconstants[const_number + i][1] = f[1 + i*4];
 			lastVSconstants[const_number + i][2] = f[2 + i*4];
 			lastVSconstants[const_number + i][3] = f[3 + i*4];
+			glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, const_number + i, lastVSconstants[const_number + i]);
 		}
 	}
 }
 
+void SetMultiVSConstant3fv(int const_number, int count, const float *f)
+{
+	for (int i = 0; i < count; i++)
+	{
+		if (lastVSconstants[const_number + i][0] != f[0 + i*3] || 
+			lastVSconstants[const_number + i][1] != f[1 + i*3] ||
+			lastVSconstants[const_number + i][2] != f[2 + i*3] ||
+			lastVSconstants[const_number + i][3] != 0.0f)
+		{
+			lastVSconstants[const_number + i][0] = f[0 + i*3];
+			lastVSconstants[const_number + i][1] = f[1 + i*3];
+			lastVSconstants[const_number + i][2] = f[2 + i*3];
+			lastVSconstants[const_number + i][3] = 0.0f;
+			glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, const_number + i, lastVSconstants[const_number + i]);
+		}
+	}
+}
+
+
 void VertexShaderCache::Init()
 {
 	for( int i=0;i<(C_FOGPARAMS+8)*4;i++)