revert the mess i made in the d3d plugin. Add way to use HLSL directly without going through Cg. Not that the results look much better... i do see shadows of spinning cars in Burnout 2 after clicking around in the dark. Then it crashes.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@2504 8ced0084-cf51-0410-be5f-012b33b47a6e

7 files changed, 53 insertions, 106 deletions

diff --git a/Source/Core/VideoCommon/Src/PixelShaderGen.cpp b/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
index c13d580106..287e5a6300 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
+++ b/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
@@ -133,7 +133,7 @@ void GetPixelShaderId(PIXELSHADERUID &uid, u32 s_texturemask, u32 zbufrender, u3
 static void WriteStage(char *&p, int n, u32 texture_mask);
 static void WrapNonPow2Tex(char* &p, const char* var, int texmap, u32 texture_mask);
 static void WriteAlphaCompare(char *&p, int num, int comp);
-static bool WriteAlphaTest(char *&p);
+static bool WriteAlphaTest(char *&p, bool HLSL);
 static void WriteFog(char *&p, bool bOutputZ);
 
 const float epsilon8bit = 1.0f / 255.0f;
@@ -369,7 +369,7 @@ static void BuildSwapModeTable()
     }
 }
 
-const char *GeneratePixelShader(u32 texture_mask, bool has_zbuffer_target, bool bRenderZToCol0)
+const char *GeneratePixelShader(u32 texture_mask, bool has_zbuffer_target, bool bRenderZToCol0, bool HLSL)
 {
 	text[sizeof(text) - 1] = 0x7C;  // canary
     DVSTARTPROFILE();
@@ -516,7 +516,7 @@ const char *GeneratePixelShader(u32 texture_mask, bool has_zbuffer_target, bool
 
     //if (bpmem.genMode.numindstages ) WRITE(p, "prev.rg = indtex0.xy;\nprev.b = 0;\n");
 
-    if (!WriteAlphaTest(p)) {
+    if (!WriteAlphaTest(p, HLSL)) {
         // alpha test will always fail, so restart the shader and just make it an empty function
         p = pmainstart;
 		WRITE(p, "discard;\n");
@@ -535,7 +535,9 @@ const char *GeneratePixelShader(u32 texture_mask, bool has_zbuffer_target, bool
 			   */
             WriteFog(p, bOutputZ);
             WRITE(p, "  ocol0 = prev;\n");
-        }
+		} else {
+			WRITE(p, "  ocol0 = prev;\n");
+		}
     }
 
     if (bRenderZ) {
@@ -889,7 +891,7 @@ static void WriteAlphaCompare(char *&p, int num, int comp)
     }
 }
 
-static bool WriteAlphaTest(char *&p)
+static bool WriteAlphaTest(char *&p, bool HLSL)
 {
     u32 op = bpmem.alphaFunc.logic;
     u32 comp[2] = {bpmem.alphaFunc.comp0,bpmem.alphaFunc.comp1};
@@ -928,7 +930,11 @@ static bool WriteAlphaTest(char *&p)
     }
 
 	// Seems we need discard for Cg and clip for d3d. sigh.
-    WRITE(p, "discard( ");
+	if (HLSL) 
+		WRITE(p, "clip( ");
+	else {
+		WRITE(p, "discard( ");
+	}
     WriteAlphaCompare(p, 0, bpmem.alphaFunc.comp0);
     
     // negated because testing the inverse condition
diff --git a/Source/Core/VideoCommon/Src/PixelShaderGen.h b/Source/Core/VideoCommon/Src/PixelShaderGen.h
index d69fd805c6..d6228f352b 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderGen.h
+++ b/Source/Core/VideoCommon/Src/PixelShaderGen.h
@@ -91,7 +91,7 @@ public:
 	}
 };
 
-const char *GeneratePixelShader(u32 texture_mask, bool has_zbuffer_target, bool bRenderZToCol0);
+const char *GeneratePixelShader(u32 texture_mask, bool has_zbuffer_target, bool bRenderZToCol0, bool HLSL = false);
 void GetPixelShaderId(PIXELSHADERUID &, u32 s_texturemask, u32 zbufrender, u32 zBufRenderToCol0);
 
 #endif
diff --git a/Source/Plugins/Plugin_VideoDX9/Src/D3DShader.cpp b/Source/Plugins/Plugin_VideoDX9/Src/D3DShader.cpp
index f1463e420a..7d4ae6f7fb 100644
--- a/Source/Plugins/Plugin_VideoDX9/Src/D3DShader.cpp
+++ b/Source/Plugins/Plugin_VideoDX9/Src/D3DShader.cpp
@@ -24,18 +24,18 @@
 namespace D3D
 {
 
-LPDIRECT3DVERTEXSHADER9 CompileVertexShader(const char *code, int len)
+LPDIRECT3DVERTEXSHADER9 CompileVertexShader(const char *code, int len, bool assembly)
 {
 	//try to compile
 	LPD3DXBUFFER shaderBuffer = 0;
 	LPD3DXBUFFER errorBuffer = 0;
 	LPDIRECT3DVERTEXSHADER9 vShader = 0;
- 	HRESULT hr = D3DXAssembleShader(code, len, 0, 0, 0, &shaderBuffer, &errorBuffer);
-	/*if (FAILED(hr))
-	{
-		//let's try 2.0
+ 	HRESULT hr;
+	
+	if (assembly)
+		hr = D3DXAssembleShader(code, len, 0, 0, 0, &shaderBuffer, &errorBuffer);
+	else
 		hr = D3DXCompileShader(code, len, 0, 0, "main", "vs_2_0", 0, &shaderBuffer, &errorBuffer, 0);
-	}*/
 
 	if (FAILED(hr))
 	{
@@ -67,14 +67,19 @@ LPDIRECT3DVERTEXSHADER9 CompileVertexShader(const char *code, int len)
 	return vShader;
 }
 
-LPDIRECT3DPIXELSHADER9 CompilePixelShader(const char *code, int len)
+LPDIRECT3DPIXELSHADER9 CompilePixelShader(const char *code, int len, bool assembly)
 {
 	LPD3DXBUFFER shaderBuffer = 0;
 	LPD3DXBUFFER errorBuffer = 0;
 	LPDIRECT3DPIXELSHADER9 pShader = 0;
 	static char *versions[6] = {"ERROR", "ps_1_1", "ps_1_4", "ps_2_0", "ps_3_0", "ps_4_0"};
 
-	HRESULT hr = D3DXAssembleShader(code, len, 0, 0, 0, &shaderBuffer, &errorBuffer);
+	HRESULT hr;
+	if (assembly)
+		hr = D3DXAssembleShader(code, len, 0, 0, 0, &shaderBuffer, &errorBuffer);
+	else
+		hr = D3DXCompileShader(code, len, 0, 0, "main", "ps_2_0", // Pixel Shader 2.0 is enough for all we do
+							   0, &shaderBuffer, &errorBuffer, 0);
 
 	if (FAILED(hr))
 	{
diff --git a/Source/Plugins/Plugin_VideoDX9/Src/D3DShader.h b/Source/Plugins/Plugin_VideoDX9/Src/D3DShader.h
index 0f7de68e16..7e2847e5e9 100644
--- a/Source/Plugins/Plugin_VideoDX9/Src/D3DShader.h
+++ b/Source/Plugins/Plugin_VideoDX9/Src/D3DShader.h
@@ -21,6 +21,6 @@
 
 namespace D3D
 {
-	LPDIRECT3DVERTEXSHADER9 CompileVertexShader(const char *code, int len);
-	LPDIRECT3DPIXELSHADER9 CompilePixelShader(const char *code, int len);
+	LPDIRECT3DVERTEXSHADER9 CompileVertexShader(const char *code, int len, bool assembly);
+	LPDIRECT3DPIXELSHADER9 CompilePixelShader(const char *code, int len, bool assembly);
 }
\ No newline at end of file
diff --git a/Source/Plugins/Plugin_VideoDX9/Src/PixelShaderCache.cpp b/Source/Plugins/Plugin_VideoDX9/Src/PixelShaderCache.cpp
index b4f0543bc5..c38ac2aab4 100644
--- a/Source/Plugins/Plugin_VideoDX9/Src/PixelShaderCache.cpp
+++ b/Source/Plugins/Plugin_VideoDX9/Src/PixelShaderCache.cpp
@@ -83,8 +83,9 @@ void PixelShaderCache::SetShader()
 		return;
 	}
 
-	const char *code = GeneratePixelShader(PixelShaderManager::GetTextureMask(), false, false);
-	LPDIRECT3DPIXELSHADER9 shader = CompileCgShader(code);
+	bool HLSL = false;
+	const char *code = GeneratePixelShader(PixelShaderManager::GetTextureMask(), false, false, HLSL);
+	LPDIRECT3DPIXELSHADER9 shader = HLSL ? D3D::CompilePixelShader(code, strlen(code), false) : CompileCgShader(code);
 	if (shader)
 	{
 		//Make an entry in the table
@@ -104,10 +105,7 @@ void PixelShaderCache::SetShader()
 
 LPDIRECT3DPIXELSHADER9 PixelShaderCache::CompileCgShader(const char *pstrprogram) 
 {
-	//char stropt[64];
-	//sprintf(stropt, "MaxLocalParams=256,MaxInstructions=%d", s_nMaxVertexInstructions);
 	const char *opts[] = {"-profileopts", "MaxLocalParams=256", "-O2", "-q", NULL};
-
 	//const char **opts = cgD3D9GetOptimalOptions(g_cgvProf);
 	CGprogram tempprog = cgCreateProgram(g_cgcontext, CG_SOURCE, pstrprogram, g_cgfProf, "main", opts);
 	if (!cgIsProgram(tempprog) || cgGetError() != CG_NO_ERROR) {
@@ -116,36 +114,11 @@ LPDIRECT3DPIXELSHADER9 PixelShaderCache::CompileCgShader(const char *pstrprogram
 		return false;
 	}
 
-	// This looks evil - we modify the program through the const char * we got from cgGetProgramString!
-	// It SHOULD not have any nasty side effects though - but you never know...
 	char *pcompiledprog = (char*)cgGetProgramString(tempprog, CG_COMPILED_PROGRAM);
-	LPD3DXBUFFER shader_binary;
-	LPD3DXBUFFER error_msg;
-
-	// Step one - Assemble into binary code. This binary code could be cached.
-	if (FAILED(D3DXAssembleShader(pcompiledprog, (UINT)strlen(pcompiledprog), NULL, NULL, 0, &shader_binary, &error_msg)))
-		PanicAlert("Asm fail");
-	// Destroy Cg program as early as possible - we want as little as possible to do with Cg due to
-	// our rather extreme performance requirements.
+
+	LPDIRECT3DPIXELSHADER9 pixel_shader = D3D::CompilePixelShader(pcompiledprog, (int)strlen(pcompiledprog), true);
 	cgDestroyProgram(tempprog);
 	tempprog = NULL;
-
-	// Create pixel shader from the binary code.
-	LPDIRECT3DPIXELSHADER9 pixel_shader = NULL;
-	if (SUCCEEDED(D3D::dev->CreatePixelShader((const DWORD *)shader_binary->GetBufferPointer(), &pixel_shader))) {
-		// PanicAlert("Success Pixel!");
-	} else {
-		if (error_msg) {
-			PanicAlert("failure pixel %s", error_msg->GetBufferPointer());
-			MessageBox(0, pcompiledprog, 0, 0);
-		}
-		else
-			PanicAlert("failure pixel with no error message.");
-	}
-	if (shader_binary)
-		shader_binary->Release();
-	if (error_msg)
-		error_msg->Release();
 	return pixel_shader;
 }
 
diff --git a/Source/Plugins/Plugin_VideoDX9/Src/Render.cpp b/Source/Plugins/Plugin_VideoDX9/Src/Render.cpp
index f9340c5e39..38fb20fbcf 100644
--- a/Source/Plugins/Plugin_VideoDX9/Src/Render.cpp
+++ b/Source/Plugins/Plugin_VideoDX9/Src/Render.cpp
@@ -85,10 +85,6 @@ void Renderer::Init(SVideoInitialize &_VideoInitialize)
     EmuWindow::SetSize(g_Res[g_Config.iWindowedRes][0], g_Res[g_Config.iWindowedRes][1]);
 
     D3D::Create(g_Config.iAdapter, EmuWindow::GetWnd(), g_Config.bFullscreen, g_Config.iFSResolution, g_Config.iMultisampleMode);
-
-	D3DVIEWPORT9 vp;
-	D3D::dev->GetViewport(&vp);
-
 	g_cgcontext = cgCreateContext();
 
 	cgGetError();
@@ -97,31 +93,23 @@ void Renderer::Init(SVideoInitialize &_VideoInitialize)
 	g_cgvProf = cgD3D9GetLatestVertexProfile();
 	g_cgfProf = cgD3D9GetLatestPixelProfile();
 
+	float width =  (float)D3D::GetDisplayWidth();
+	float height = (float)D3D::GetDisplayHeight();
+
 	m_x = 0;
 	m_y = 0;
-	m_width  = (float)vp.Width;
-	m_height = (float)vp.Height;
-	xScale = 640.0f / (float)vp.Width;
-	yScale = 480.0f / (float)vp.Height;
+	m_width  = width;
+	m_height = height;
+	xScale = width / (float)EFB_WIDTH;
+	yScale = height / (float)EFB_HEIGHT;
 
+	// We're not using much fixed function. Let's just set the matrices to identity.
 	D3DXMATRIX mtx;
 	D3DXMatrixIdentity(&mtx);
 	D3D::dev->SetTransform(D3DTS_VIEW, &mtx);
 	D3D::dev->SetTransform(D3DTS_WORLD, &mtx);
-	float width =  (float)D3D::GetDisplayWidth();
-	float height = (float)D3D::GetDisplayHeight();
 
-	xScale = width/640.0f;
-	yScale = height/480.0f;
-/*
-	RECT rc =
-	{
-		(LONG)(m_x*xScale),
-		(LONG)(m_y*yScale),
-		(LONG)(m_width*xScale),
-		(LONG)(m_height*yScale)
-	};
-*/
+
 	D3D::font.Init();
 	Initialize();
 }
@@ -145,6 +133,7 @@ void Renderer::Initialize()
 		m_Textures.push_back(NULL);
 	for (int i = 0; i < 8; i++)
 		D3D::dev->SetSamplerState(i, D3DSAMP_MAXANISOTROPY, 16);
+
 	Postprocess::Initialize();
 	Postprocess::BeginFrame();
 	D3D::BeginFrame(true, 0);
@@ -158,11 +147,12 @@ void Renderer::AddMessage(const std::string &message, u32 ms)
 
 void Renderer::ProcessMessages()
 {
-	if (s_listMsgs.size() > 0) {
+	if (s_listMsgs.size() > 0)
+	{
         int left = 25, top = 15;
 		std::list<Message>::iterator it = s_listMsgs.begin();
         
-        while(it != s_listMsgs.end()) 
+        while (it != s_listMsgs.end()) 
 		{
 			int time_left = (int)(it->dwTimeStamp - timeGetTime());
 			int alpha = 255;
@@ -193,7 +183,7 @@ void Renderer::RenderText(const std::string &text, int left, int top, u32 color)
 
 void dumpMatrix(D3DXMATRIX &mtx)
 {
-	for (int y=0; y<4; y++)
+	for (int y = 0; y < 4; y++)
 	{
 		char temp[256];
 		sprintf(temp,"%4.4f %4.4f %4.4f %4.4f",mtx.m[y][0],mtx.m[y][1],mtx.m[y][2],mtx.m[y][3]);
@@ -213,8 +203,6 @@ void Renderer::SwapBuffers()
 		int height = rcWindow.bottom - rcWindow.top;
 
 		::MoveWindow(EmuWindow::GetWnd(), 0, 0, width, height, FALSE);
-	//	nBackbufferWidth = width;
-	//	nBackbufferHeight = height;
 	}
 
 	//Finish up the current frame, print some stats
@@ -290,6 +278,7 @@ void Renderer::SwapBuffers()
 	rc.right  = (LONG)m_width;
 	rc.bottom = (LONG)m_height;
 	D3D::dev->SetScissorRect(&rc);
+	D3D::dev->SetRenderState(D3DRS_SCISSORTESTENABLE, false);
 
 	D3D::dev->Clear(0, 0, D3DCLEAR_TARGET, 0x101010, 0, 0);
 	u32 clearColor = (bpmem.clearcolorAR << 16) | bpmem.clearcolorGB;
@@ -347,6 +336,7 @@ void Renderer::SetViewport(float* _Viewport)
 
 void Renderer::SetScissorRect()
 {
+	/*
 	RECT rc = {0,0,0,0}; // FIXX
 	rc.left   = (int)(rc.left   * xScale);
 	rc.top    = (int)(rc.top    * yScale);
@@ -355,7 +345,7 @@ void Renderer::SetScissorRect()
 	if (rc.right >= rc.left && rc.bottom >= rc.top)
 		D3D::dev->SetScissorRect(&rc);
 	else
-		g_VideoInitialize.pLog("SCISSOR ERROR", FALSE);
+		g_VideoInitialize.pLog("SCISSOR ERROR", FALSE);*/
 }
 
 /*
@@ -493,10 +483,6 @@ void UpdateViewport()
 	int yoffs = 0;
 	int wid, hei, actualWid, actualHei;
 
-	int winw = 640;
-	int winh = 480;
-	float ratio = (float)winw / (float)winh / fourThree;
-	
 	vp.MinZ = (xfregs.rawViewport[5] - xfregs.rawViewport[2])/16777215.0f;
 	vp.MaxZ = xfregs.rawViewport[5]/16777215.0f;
 	
diff --git a/Source/Plugins/Plugin_VideoDX9/Src/VertexShaderCache.cpp b/Source/Plugins/Plugin_VideoDX9/Src/VertexShaderCache.cpp
index f013c850d5..e9aa4aff73 100644
--- a/Source/Plugins/Plugin_VideoDX9/Src/VertexShaderCache.cpp
+++ b/Source/Plugins/Plugin_VideoDX9/Src/VertexShaderCache.cpp
@@ -85,8 +85,9 @@ void VertexShaderCache::SetShader(u32 components)
 		return;
 	}
 
+	bool HLSL = false;
 	const char *code = GenerateVertexShader(components, false);
-	LPDIRECT3DVERTEXSHADER9 shader = CompileCgShader(code);
+	LPDIRECT3DVERTEXSHADER9 shader = HLSL ? D3D::CompileVertexShader(code, strlen(code), false) : CompileCgShader(code);
 	if (shader)
 	{
 		// Make an entry in the table
@@ -121,33 +122,9 @@ LPDIRECT3DVERTEXSHADER9 VertexShaderCache::CompileCgShader(const char *pstrprogr
 	}
 	const char *pcompiledprog = cgGetProgramString(tempprog, CG_COMPILED_PROGRAM);
 
-	LPD3DXBUFFER shader_binary;
-	LPD3DXBUFFER error_msg;
-
-	// Step one - Assemble into binary code. This binary code could be cached.
-	if (FAILED(D3DXAssembleShader(pcompiledprog, (UINT)strlen(pcompiledprog), NULL, NULL, 0, &shader_binary, &error_msg)))
-		PanicAlert("Asm fail");
-	// Destroy Cg program as early as possible - we want as little as possible to do with Cg due to
-	// our rather extreme performance requirements.
+	LPDIRECT3DVERTEXSHADER9 vertex_shader = D3D::CompileVertexShader(pcompiledprog, (int)strlen(pcompiledprog), true);
 	cgDestroyProgram(tempprog);
 	tempprog = NULL;
-
-	// Create vertex shader from the binary code.
-	LPDIRECT3DVERTEXSHADER9 vertex_shader = NULL;
-	if (SUCCEEDED(D3D::dev->CreateVertexShader((const DWORD *)shader_binary->GetBufferPointer(), &vertex_shader))) {
-		// PanicAlert("Successvertex!");
-	} else {
-		if (error_msg) {
-			PanicAlert("failure vertex %s", error_msg->GetBufferPointer());
-			MessageBox(0, pcompiledprog, 0, 0);
-		}
-		else
-			PanicAlert("failure vertex with no error message.");
-	}
-	if (shader_binary)
-		shader_binary->Release();
-	if (error_msg)
-		error_msg->Release();
 	return vertex_shader;
 }