DX11 plugin: Do destination-alpha in a single pass. No more drawing the same geometry twice! SMG is faster now because it uses destination-alpha extensively. This uses dual-source color blending, a DirectX 10-level feature. The equivalent OpenGL function comes from the GL_ARB_blend_func_extended extension, which was made core in OpenGL 3.3.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6294 8ced0084-cf51-0410-be5f-012b33b47a6e

7 files changed, 79 insertions, 62 deletions

diff --git a/Source/Core/Common/Src/LinearDiskCache.cpp b/Source/Core/Common/Src/LinearDiskCache.cpp
index 37dfb8a69a..25665257ab 100644
--- a/Source/Core/Common/Src/LinearDiskCache.cpp
+++ b/Source/Core/Common/Src/LinearDiskCache.cpp
@@ -22,7 +22,7 @@ static const char ID[4] = {'D', 'C', 'A', 'C'};
 // Update this to the current SVN revision every time you change shader generation code.
 // We don't automatically get this from SVN_REV because that would mean regenerating the
 // shader cache for every revision, graphics-related or not, which is simply annoying.
-const int version = 6267;
+const int version = 6294;
 
 LinearDiskCache::LinearDiskCache() 
 	: file_(NULL), num_entries_(0) {
diff --git a/Source/Core/VideoCommon/Src/PixelShaderGen.cpp b/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
index e528c6a1b3..7b638a1cbd 100644
--- a/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
+++ b/Source/Core/VideoCommon/Src/PixelShaderGen.cpp
@@ -521,7 +521,7 @@ const char *GeneratePixelShaderCode(bool dstAlphaEnable, API_TYPE ApiType,u32 co
 	if(ApiType != API_D3D11)
 		WRITE(p, "  out float4 ocol0 : COLOR0,%s\n  in float4 rawpos : %s,\n",DepthTextureEnable ? "\n  out float depth : DEPTH," : "", ApiType == API_OPENGL ? "WPOS" : "POSITION");
 	else
-		WRITE(p, "  out float4 ocol0 : SV_Target,%s\n  in float4 rawpos : SV_Position,\n",DepthTextureEnable ? "\n  out float depth : SV_Depth," : "");
+		WRITE(p, "  out float4 ocol0 : SV_Target0,\n  out float4 ocol1 : SV_Target1,%s\n  in float4 rawpos : SV_Position,\n",DepthTextureEnable ? "\n  out float depth : SV_Depth," : "");
 	
 	WRITE(p, "  in float4 colors_0 : COLOR0,\n");
 	WRITE(p, "  in float4 colors_1 : COLOR1");
@@ -544,7 +544,8 @@ const char *GeneratePixelShaderCode(bool dstAlphaEnable, API_TYPE ApiType,u32 co
 
 	char* pmainstart = p;
 	int Pretest = AlphaPreTest();
-	if (dstAlphaEnable && !DepthTextureEnable && Pretest >= 0)
+	// TODO: Re-enable the early discard on D3D11
+	if (dstAlphaEnable && !DepthTextureEnable && Pretest >= 0 && ApiType != API_D3D11)
 	{
 		if (!Pretest)
 		{
@@ -806,13 +807,24 @@ const char *GeneratePixelShaderCode(bool dstAlphaEnable, API_TYPE ApiType,u32 co
 			WRITE(p, "depth = zCoord;\n");
 		}
 
-		if (dstAlphaEnable)
+		if (dstAlphaEnable && ApiType != API_D3D11)
 			WRITE(p, "  ocol0 = float4(prev.rgb, "I_ALPHA"[0].a);\n");
 		else
 		{
 			WriteFog(p);
 			WRITE(p, "  ocol0 = prev;\n");
-		}		
+		}
+
+		// On D3D11, use dual-source color blending to perform dst alpha in a
+		// single pass
+		if (ApiType == API_D3D11)
+		{
+			// Colors will be blended against the alpha from ocol1...
+			WRITE(p, "  ocol1 = ocol0;\n");
+			// ...and the alpha from ocol0 will be written to the framebuffer.
+			if (dstAlphaEnable)
+				WRITE(p, "  ocol0.a = "I_ALPHA"[0].a;\n");
+		}
 	}
 	WRITE(p, "}\n");
 	if (text[sizeof(text) - 1] != 0x7C)
@@ -1256,7 +1268,7 @@ static bool WriteAlphaTest(char *&p, API_TYPE ApiType)
 	int Pretest = AlphaPreTest();
 	if(Pretest >= 0)
 	{
-		return Pretest;
+		return Pretest != 0;
 	}
 
 	// using discard then return works the same in cg and dx9 but not in dx11
diff --git a/Source/Plugins/Plugin_VideoDX11/Src/GfxState.cpp b/Source/Plugins/Plugin_VideoDX11/Src/GfxState.cpp
index 10d2dd2d3a..b9a3ca4b65 100644
--- a/Source/Plugins/Plugin_VideoDX11/Src/GfxState.cpp
+++ b/Source/Plugins/Plugin_VideoDX11/Src/GfxState.cpp
@@ -34,6 +34,8 @@ EmuGfxState::EmuGfxState() : vertexshader(NULL), vsbytecode(NULL), pixelshader(N
 		if(g_ActiveConfig.iMaxAnisotropy > 1) samplerdesc[k].Filter = D3D11_FILTER_ANISOTROPIC;
 	}
 
+	m_useDstAlpha = false;
+
 	memset(&blenddesc, 0, sizeof(blenddesc));
 	blenddesc.AlphaToCoverageEnable = FALSE;
 	blenddesc.IndependentBlendEnable = FALSE;
@@ -70,7 +72,7 @@ EmuGfxState::EmuGfxState() : vertexshader(NULL), vsbytecode(NULL), pixelshader(N
 EmuGfxState::~EmuGfxState()
 {
 	for (unsigned int k = 0;k < 8;k++)
-		SAFE_RELEASE(shader_resources[k])
+		SAFE_RELEASE(shader_resources[k]);
 
 	SAFE_RELEASE(vsbytecode);
 	SAFE_RELEASE(psbytecode);
@@ -224,27 +226,6 @@ void EmuGfxState::ApplyState()
 	apply_called = true;
 }
 
-void EmuGfxState::AlphaPass()
-{
-	if (!apply_called) ERROR_LOG(VIDEO, "EmuGfxState::AlphaPass called without having called ApplyState before!")
-	else stateman->PopBlendState();
-
-	// pixel shader for alpha pass is different, so update it
-	context->PSSetShader(pixelshader, NULL, 0);
-
-	ID3D11BlendState* blstate;
-	D3D11_BLEND_DESC desc = blenddesc;
-	desc.RenderTarget[0].RenderTargetWriteMask = D3D11_COLOR_WRITE_ENABLE_ALPHA;
-	desc.RenderTarget[0].BlendEnable = FALSE;
-	HRESULT hr = device->CreateBlendState(&desc, &blstate);
-	if (FAILED(hr)) PanicAlert("Failed to create blend state at %s %d\n", __FILE__, __LINE__);
-	SetDebugObjectName((ID3D11DeviceChild*)blstate, "a blend state of EmuGfxState (created during alpha pass)");
-	stateman->PushBlendState(blstate);
-	blstate->Release();
-
-	stateman->Apply();
-}
-
 void EmuGfxState::Reset()
 {
 	for (unsigned int k = 0;k < 8;k++)
@@ -274,28 +255,57 @@ void EmuGfxState::SetSrcBlend(D3D11_BLEND val)
 {
 	// TODO: Check whether e.g. the dest color check is needed here
 	blenddesc.RenderTarget[0].SrcBlend = val;
-	if (val == D3D11_BLEND_SRC_COLOR) blenddesc.RenderTarget[0].SrcBlendAlpha = D3D11_BLEND_SRC_ALPHA;
-	else if (val == D3D11_BLEND_INV_SRC_COLOR) blenddesc.RenderTarget[0].SrcBlendAlpha = D3D11_BLEND_INV_SRC_ALPHA;
-	else if (val == D3D11_BLEND_DEST_COLOR) blenddesc.RenderTarget[0].SrcBlendAlpha = D3D11_BLEND_DEST_ALPHA;
-	else if (val == D3D11_BLEND_INV_DEST_COLOR) blenddesc.RenderTarget[0].SrcBlendAlpha = D3D11_BLEND_INV_DEST_ALPHA;
-	else blenddesc.RenderTarget[0].SrcBlendAlpha = val;
+	if (m_useDstAlpha)
+	{
+		blenddesc.RenderTarget[0].SrcBlendAlpha = D3D11_BLEND_ONE;
+	}
+	else
+	{
+		if (val == D3D11_BLEND_SRC_COLOR) blenddesc.RenderTarget[0].SrcBlendAlpha = D3D11_BLEND_SRC_ALPHA;
+		else if (val == D3D11_BLEND_INV_SRC_COLOR) blenddesc.RenderTarget[0].SrcBlendAlpha = D3D11_BLEND_INV_SRC_ALPHA;
+		else if (val == D3D11_BLEND_DEST_COLOR) blenddesc.RenderTarget[0].SrcBlendAlpha = D3D11_BLEND_DEST_ALPHA;
+		else if (val == D3D11_BLEND_INV_DEST_COLOR) blenddesc.RenderTarget[0].SrcBlendAlpha = D3D11_BLEND_INV_DEST_ALPHA;
+		else blenddesc.RenderTarget[0].SrcBlendAlpha = val;
+	}
 }
 
 void EmuGfxState::SetDestBlend(D3D11_BLEND val)
 {
 	// TODO: Check whether e.g. the source color check is needed here
 	blenddesc.RenderTarget[0].DestBlend = val;
-	if (val == D3D11_BLEND_SRC_COLOR) blenddesc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_SRC_ALPHA;
-	else if (val == D3D11_BLEND_INV_SRC_COLOR) blenddesc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_INV_SRC_ALPHA;
-	else if (val == D3D11_BLEND_DEST_COLOR) blenddesc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_DEST_ALPHA;
-	else if (val == D3D11_BLEND_INV_DEST_COLOR) blenddesc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_INV_DEST_ALPHA;
-	else blenddesc.RenderTarget[0].DestBlendAlpha = val;
+	if (m_useDstAlpha)
+	{
+		blenddesc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_ZERO;
+	}
+	else
+	{
+		if (val == D3D11_BLEND_SRC_COLOR) blenddesc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_SRC_ALPHA;
+		else if (val == D3D11_BLEND_INV_SRC_COLOR) blenddesc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_INV_SRC_ALPHA;
+		else if (val == D3D11_BLEND_DEST_COLOR) blenddesc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_DEST_ALPHA;
+		else if (val == D3D11_BLEND_INV_DEST_COLOR) blenddesc.RenderTarget[0].DestBlendAlpha = D3D11_BLEND_INV_DEST_ALPHA;
+		else blenddesc.RenderTarget[0].DestBlendAlpha = val;
+	}
 }
 
 void EmuGfxState::SetBlendOp(D3D11_BLEND_OP val)
 {
 	blenddesc.RenderTarget[0].BlendOp = val;
-	blenddesc.RenderTarget[0].BlendOpAlpha = val;
+	if (m_useDstAlpha)
+	{
+		blenddesc.RenderTarget[0].BlendOpAlpha = D3D11_BLEND_OP_ADD;
+	}
+	else
+	{
+		blenddesc.RenderTarget[0].BlendOpAlpha = val;
+	}
+}
+
+void EmuGfxState::SetDstAlpha(bool enable)
+{
+	m_useDstAlpha = enable;
+	SetSrcBlend(blenddesc.RenderTarget[0].SrcBlend);
+	SetDestBlend(blenddesc.RenderTarget[0].DestBlend);
+	SetBlendOp(blenddesc.RenderTarget[0].BlendOp);
 }
 
 void EmuGfxState::SetSamplerFilter(DWORD stage, D3D11_FILTER filter)
diff --git a/Source/Plugins/Plugin_VideoDX11/Src/GfxState.h b/Source/Plugins/Plugin_VideoDX11/Src/GfxState.h
index 507eb76c34..90151a7905 100644
--- a/Source/Plugins/Plugin_VideoDX11/Src/GfxState.h
+++ b/Source/Plugins/Plugin_VideoDX11/Src/GfxState.h
@@ -39,7 +39,6 @@ public:
 	void SetShaderResource(unsigned int stage, ID3D11ShaderResourceView* srv);
 
 	void ApplyState();            // apply current state
-	void AlphaPass();             // only modify the current state to enable the alpha pass
 	void Reset();
 
 	// blend state
@@ -49,6 +48,8 @@ public:
 	void SetDestBlend(D3D11_BLEND val);
 	void SetBlendOp(D3D11_BLEND_OP val);
 
+	void SetDstAlpha(bool enable);
+
 	// sampler states
 	void SetSamplerFilter(DWORD stage, D3D11_FILTER filter);
 
@@ -80,6 +81,8 @@ private:
 	ID3D11ShaderResourceView* shader_resources[8];
 	D3D11_BLEND_DESC blenddesc;
 
+	bool m_useDstAlpha;
+
 	bool apply_called;
 };
 
diff --git a/Source/Plugins/Plugin_VideoDX11/Src/Render.cpp b/Source/Plugins/Plugin_VideoDX11/Src/Render.cpp
index c06ff9c9de..5ab0290b82 100644
--- a/Source/Plugins/Plugin_VideoDX11/Src/Render.cpp
+++ b/Source/Plugins/Plugin_VideoDX11/Src/Render.cpp
@@ -90,8 +90,8 @@ static const D3D11_BLEND d3dSrcFactors[8] =
 	D3D11_BLEND_ONE,
 	D3D11_BLEND_DEST_COLOR,
 	D3D11_BLEND_INV_DEST_COLOR,
-	D3D11_BLEND_SRC_ALPHA,
-	D3D11_BLEND_INV_SRC_ALPHA, 
+	D3D11_BLEND_SRC1_ALPHA,
+	D3D11_BLEND_INV_SRC1_ALPHA, // Use dual-source color blending for dst alpha
 	D3D11_BLEND_DEST_ALPHA,
 	D3D11_BLEND_INV_DEST_ALPHA
 };
@@ -102,8 +102,8 @@ static const D3D11_BLEND d3dDestFactors[8] =
 	D3D11_BLEND_ONE,
 	D3D11_BLEND_SRC_COLOR,
 	D3D11_BLEND_INV_SRC_COLOR,
-	D3D11_BLEND_SRC_ALPHA,
-	D3D11_BLEND_INV_SRC_ALPHA,
+	D3D11_BLEND_SRC1_ALPHA,
+	D3D11_BLEND_INV_SRC1_ALPHA, // Use dual-source color blending for dst alpha
 	D3D11_BLEND_DEST_ALPHA,
 	D3D11_BLEND_INV_DEST_ALPHA
 };
diff --git a/Source/Plugins/Plugin_VideoDX11/Src/VertexManager.cpp b/Source/Plugins/Plugin_VideoDX11/Src/VertexManager.cpp
index ff4571c054..6f9d90feb8 100644
--- a/Source/Plugins/Plugin_VideoDX11/Src/VertexManager.cpp
+++ b/Source/Plugins/Plugin_VideoDX11/Src/VertexManager.cpp
@@ -144,12 +144,9 @@ void VertexManager::LoadBuffers()
 	m_indexBufferCursor += iCount;
 }
 
-void VertexManager::Draw(UINT stride, bool alphapass)
+void VertexManager::Draw(UINT stride)
 {
-	if (!alphapass)
-		D3D::gfxstate->ApplyState();
-	else
-		D3D::gfxstate->AlphaPass();
+	D3D::gfxstate->ApplyState();
 
 	D3D::context->IASetVertexBuffers(0, 1, &m_vertexBuffer, &stride, &m_vertexDrawOffset);
 	D3D::context->IASetIndexBuffer(m_indexBuffer, DXGI_FORMAT_R16_UINT, 0);
@@ -221,7 +218,12 @@ void VertexManager::vFlush()
 	VertexShaderManager::SetConstants();
 	PixelShaderManager::SetConstants();
 
-	if (!PixelShaderCache::SetShader(false,g_nativeVertexFmt->m_components))
+	bool useDstAlpha = bpmem.dstalpha.enable && bpmem.blendmode.alphaupdate &&
+		bpmem.zcontrol.pixel_format == PIXELFMT_RGBA6_Z24;
+
+	D3D::gfxstate->SetDstAlpha(useDstAlpha);
+
+	if (!PixelShaderCache::SetShader(useDstAlpha,g_nativeVertexFmt->m_components))
 		goto shader_fail;
 	if (!VertexShaderCache::SetShader(g_nativeVertexFmt->m_components))
 		goto shader_fail;
@@ -230,18 +232,8 @@ void VertexManager::vFlush()
 	g_nativeVertexFmt->SetupVertexPointers();
 
 	LoadBuffers();
+	Draw(stride);
 
-	Draw(stride, false);
-
-	if (bpmem.dstalpha.enable && bpmem.blendmode.alphaupdate)
-	{
-		DWORD write = 0;
-		if (!PixelShaderCache::SetShader(true,g_nativeVertexFmt->m_components))
-			goto shader_fail;
-
-		// update alpha only
-		Draw(stride, true);
-	}
 	D3D::gfxstate->Reset();
 
 shader_fail:
diff --git a/Source/Plugins/Plugin_VideoDX11/Src/VertexManager.h b/Source/Plugins/Plugin_VideoDX11/Src/VertexManager.h
index b6e3ebd545..8efc08d373 100644
--- a/Source/Plugins/Plugin_VideoDX11/Src/VertexManager.h
+++ b/Source/Plugins/Plugin_VideoDX11/Src/VertexManager.h
@@ -37,7 +37,7 @@ private:
 	void CreateDeviceObjects();
 	void DestroyDeviceObjects();
 	void LoadBuffers();
-	void Draw(UINT stride, bool alphapass);
+	void Draw(UINT stride);
 	// temp
 	void vFlush();