VideoCommon: Add a separate constants buffer for the geometry shader.

Source/Core/VideoBackends/D3D/GeometryShaderCache.cpp

@@ -17,6 +17,8 @@
 
 #include "VideoCommon/Debugger.h"
 #include "VideoCommon/GeometryShaderGen.h"
+#include "VideoCommon/GeometryShaderManager.h"
+#include "VideoCommon/Statistics.h"
 #include "VideoCommon/VideoConfig.h"
 
 namespace DX11
@@ -37,6 +39,22 @@ ID3D11GeometryShader* GeometryShaderCache::GetCopyGeometryShader() { return Copy
 
 ID3D11Buffer* gscbuf = nullptr;
 
+ID3D11Buffer* &GeometryShaderCache::GetConstantBuffer()
+{
+	// TODO: divide the global variables of the generated shaders into about 5 constant buffers to speed this up
+	if (GeometryShaderManager::dirty)
+	{
+		D3D11_MAPPED_SUBRESOURCE map;
+		D3D::context->Map(gscbuf, 0, D3D11_MAP_WRITE_DISCARD, 0, &map);
+		memcpy(map.pData, &GeometryShaderManager::constants, sizeof(GeometryShaderConstants));
+		D3D::context->Unmap(gscbuf, 0);
+		GeometryShaderManager::dirty = false;
+
+		ADDSTAT(stats.thisFrame.bytesUniformStreamed, sizeof(GeometryShaderConstants));
+	}
+	return gscbuf;
+}
+
 // this class will load the precompiled shaders into our cache
 class GeometryShaderCacheInserter : public LinearDiskCacheReader<GeometryShaderUid, u8>
 {
@@ -113,6 +131,12 @@ const char copy_shader_code[] = {
 
 void GeometryShaderCache::Init()
 {
+	unsigned int gbsize = ((sizeof(GeometryShaderConstants))&(~0xf)) + 0x10; // must be a multiple of 16
+	D3D11_BUFFER_DESC gbdesc = CD3D11_BUFFER_DESC(gbsize, D3D11_BIND_CONSTANT_BUFFER, D3D11_USAGE_DYNAMIC, D3D11_CPU_ACCESS_WRITE);
+	HRESULT hr = D3D::device->CreateBuffer(&gbdesc, nullptr, &gscbuf);
+	CHECK(hr == S_OK, "Create geometry shader constant buffer (size=%u)", gbsize);
+	D3D::SetDebugObjectName((ID3D11DeviceChild*)gscbuf, "geometry shader constant buffer used to emulate the GX pipeline");
+
 	// used when drawing clear quads
 	ClearGeometryShader = D3D::CompileAndCreateGeometryShader(clear_shader_code);
 	CHECK(ClearGeometryShader != nullptr, "Create clear geometry shader");
@@ -152,6 +176,8 @@ void GeometryShaderCache::Clear()
 
 void GeometryShaderCache::Shutdown()
 {
+	SAFE_RELEASE(gscbuf);
+
 	SAFE_RELEASE(ClearGeometryShader);
 	SAFE_RELEASE(CopyGeometryShader);
 

Source/Core/VideoBackends/D3D/GeometryShaderCache.h

@@ -25,6 +25,7 @@ public:
 	static ID3D11GeometryShader* GeometryShaderCache::GetCopyGeometryShader();
 
 	static ID3D11GeometryShader* GetActiveShader() { return last_entry->shader; }
+	static ID3D11Buffer* &GetConstantBuffer();
 
 private:
 	struct GSCacheEntry

Source/Core/VideoBackends/D3D/VertexManager.cpp

@@ -140,7 +140,7 @@ void VertexManager::Draw(u32 stride)
 	if (current_primitive_type == PRIMITIVE_TRIANGLES)
 	{
 		D3D::stateman->SetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY_TRIANGLESTRIP);
-		D3D::stateman->SetGeometryConstants(VertexShaderCache::GetConstantBuffer());
+		D3D::stateman->SetGeometryConstants(GeometryShaderCache::GetConstantBuffer());
 		D3D::stateman->SetGeometryShader(g_ActiveConfig.iStereoMode > 0 ? GeometryShaderCache::GetActiveShader() : nullptr);
 
 		D3D::stateman->Apply();

Source/Core/VideoBackends/D3D/main.cpp

@@ -28,6 +28,7 @@
 #include "VideoCommon/BPStructs.h"
 #include "VideoCommon/CommandProcessor.h"
 #include "VideoCommon/Fifo.h"
+#include "VideoCommon/GeometryShaderManager.h"
 #include "VideoCommon/IndexGenerator.h"
 #include "VideoCommon/OnScreenDisplay.h"
 #include "VideoCommon/OpcodeDecoding.h"
@@ -187,6 +188,7 @@ void VideoBackend::Video_Prepare()
 	OpcodeDecoder_Init();
 	VertexShaderManager::Init();
 	PixelShaderManager::Init();
+	GeometryShaderManager::Init();
 	CommandProcessor::Init();
 	PixelEngine::Init();
 	BBox::Init();
@@ -205,6 +207,7 @@ void VideoBackend::Shutdown()
 		// VideoCommon
 		Fifo_Shutdown();
 		CommandProcessor::Shutdown();
+		GeometryShaderManager::Shutdown();
 		PixelShaderManager::Shutdown();
 		VertexShaderManager::Shutdown();
 		OpcodeDecoder_Shutdown();

Source/Core/VideoBackends/OGL/ProgramShaderCache.cpp

@@ -13,6 +13,7 @@
 
 #include "VideoCommon/Debugger.h"
 #include "VideoCommon/DriverDetails.h"
+#include "VideoCommon/GeometryShaderManager.h"
 #include "VideoCommon/ImageWrite.h"
 #include "VideoCommon/PixelShaderManager.h"
 #include "VideoCommon/Statistics.h"
@@ -71,11 +72,14 @@ void SHADER::SetProgramVariables()
 
 		GLint PSBlock_id = glGetUniformBlockIndex(glprogid, "PSBlock");
 		GLint VSBlock_id = glGetUniformBlockIndex(glprogid, "VSBlock");
+		GLint GSBlock_id = glGetUniformBlockIndex(glprogid, "GSBlock");
 
 		if (PSBlock_id != -1)
 			glUniformBlockBinding(glprogid, PSBlock_id, 1);
 		if (VSBlock_id != -1)
 			glUniformBlockBinding(glprogid, VSBlock_id, 2);
+		if (GSBlock_id != -1)
+			glUniformBlockBinding(glprogid, GSBlock_id, 3);
 
 		// Bind Texture Sampler
 		for (int a = 0; a <= 9; ++a)
@@ -133,7 +137,7 @@ void SHADER::Bind()
 
 void ProgramShaderCache::UploadConstants()
 {
-	if (PixelShaderManager::dirty || VertexShaderManager::dirty)
+	if (PixelShaderManager::dirty || VertexShaderManager::dirty || GeometryShaderManager::dirty)
 	{
 		auto buffer = s_buffer->Map(s_ubo_buffer_size, s_ubo_align);
 
@@ -143,14 +147,20 @@ void ProgramShaderCache::UploadConstants()
 		memcpy(buffer.first + ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align),
 			&VertexShaderManager::constants, sizeof(VertexShaderConstants));
 
+		memcpy(buffer.first + ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align) + ROUND_UP(sizeof(VertexShaderConstants), s_ubo_align),
+			&GeometryShaderManager::constants, sizeof(GeometryShaderConstants));
+
 		s_buffer->Unmap(s_ubo_buffer_size);
 		glBindBufferRange(GL_UNIFORM_BUFFER, 1, s_buffer->m_buffer, buffer.second,
 					sizeof(PixelShaderConstants));
 		glBindBufferRange(GL_UNIFORM_BUFFER, 2, s_buffer->m_buffer, buffer.second + ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align),
 					sizeof(VertexShaderConstants));
+		glBindBufferRange(GL_UNIFORM_BUFFER, 3, s_buffer->m_buffer, buffer.second + ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align) + ROUND_UP(sizeof(VertexShaderConstants), s_ubo_align),
+					sizeof(GeometryShaderConstants));
 
 		PixelShaderManager::dirty = false;
 		VertexShaderManager::dirty = false;
+		GeometryShaderManager::dirty = false;
 
 		ADDSTAT(stats.thisFrame.bytesUniformStreamed, s_ubo_buffer_size);
 	}
@@ -419,7 +429,7 @@ void ProgramShaderCache::Init()
 	// then the UBO will fail.
 	glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, &s_ubo_align);
 
-	s_ubo_buffer_size = ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align) + ROUND_UP(sizeof(VertexShaderConstants), s_ubo_align);
+	s_ubo_buffer_size = ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align) + ROUND_UP(sizeof(VertexShaderConstants), s_ubo_align) + ROUND_UP(sizeof(GeometryShaderConstants), s_ubo_align);
 
 	// We multiply by *4*4 because we need to get down to basic machine units.
 	// So multiply by four to get how many floats we have from vec4s

Source/Core/VideoBackends/OGL/main.cpp

@@ -65,6 +65,7 @@ Make AA apply instantly during gameplay if possible
 #include "VideoCommon/BPStructs.h"
 #include "VideoCommon/CommandProcessor.h"
 #include "VideoCommon/Fifo.h"
+#include "VideoCommon/GeometryShaderManager.h"
 #include "VideoCommon/ImageWrite.h"
 #include "VideoCommon/IndexGenerator.h"
 #include "VideoCommon/LookUpTables.h"
@@ -202,6 +203,7 @@ void VideoBackend::Video_Prepare()
 	IndexGenerator::Init();
 	VertexShaderManager::Init();
 	PixelShaderManager::Init();
+	GeometryShaderManager::Init();
 	ProgramShaderCache::Init();
 	g_texture_cache = new TextureCache();
 	g_sampler_cache = new SamplerCache();
@@ -243,6 +245,7 @@ void VideoBackend::Video_Cleanup()
 		ProgramShaderCache::Shutdown();
 		VertexShaderManager::Shutdown();
 		PixelShaderManager::Shutdown();
+		GeometryShaderManager::Shutdown();
 		delete g_perf_query;
 		g_perf_query = nullptr;
 		delete g_vertex_manager;

Source/Core/VideoCommon/BPStructs.cpp

@@ -14,6 +14,7 @@
 #include "VideoCommon/BPFunctions.h"
 #include "VideoCommon/BPStructs.h"
 #include "VideoCommon/Fifo.h"
+#include "VideoCommon/GeometryShaderManager.h"
 #include "VideoCommon/PerfQueryBase.h"
 #include "VideoCommon/PixelEngine.h"
 #include "VideoCommon/PixelShaderManager.h"
@@ -125,6 +126,7 @@ static void BPWritten(const BPCmd& bp)
 	case BPMEM_SCISSOROFFSET: // Scissor Offset
 		SetScissor();
 		VertexShaderManager::SetViewportChanged();
+		GeometryShaderManager::SetViewportChanged();
 		return;
 	case BPMEM_LINEPTWIDTH: // Line Width
 		SetLineWidth();

Source/Core/VideoCommon/CMakeLists.txt

@@ -10,6 +10,7 @@ set(SRCS	BoundingBox.cpp
 			FPSCounter.cpp
 			FramebufferManagerBase.cpp
 			GeometryShaderGen.cpp
+			GeometryShaderManager.cpp
 			HiresTextures.cpp
 			ImageWrite.cpp
 			IndexGenerator.cpp

Source/Core/VideoCommon/ConstantManager.h

@@ -43,6 +43,9 @@ struct VertexShaderConstants
 	float4 normalmatrices[32];
 	float4 posttransformmatrices[64];
 	float4 pixelcentercorrection;
-	float4 stereoparams;
 };
 
+struct GeometryShaderConstants
+{
+	float4 stereoparams;
+};

Source/Core/VideoCommon/GeometryShaderGen.cpp

@@ -43,11 +43,12 @@ static inline void GenerateGeometryShader(T& out, u32 components, API_TYPE ApiTy
 
 	// uniforms
 	if (ApiType == API_OPENGL)
-		out.Write("layout(std140%s) uniform VSBlock {\n", g_ActiveConfig.backend_info.bSupportsBindingLayout ? ", binding = 2" : "");
+		out.Write("layout(std140%s) uniform GSBlock {\n", g_ActiveConfig.backend_info.bSupportsBindingLayout ? ", binding = 3" : "");
 	else
-		out.Write("cbuffer VSBlock {\n");
+		out.Write("cbuffer GSBlock {\n");
-	out.Write(s_shader_uniforms);
+	out.Write(
-	out.Write("};\n");
+		"\tfloat4 " I_STEREOPARAMS";\n"
+		"};\n");
 
 	uid_data->numTexGens = xfmem.numTexGen.numTexGens;
 	uid_data->pixel_lighting = g_ActiveConfig.bEnablePixelLighting;

Source/Core/VideoCommon/GeometryShaderManager.cpp

@@ -0,0 +1,82 @@
+// Copyright 2013 Dolphin Emulator Project
+// Licensed under GPLv2
+// Refer to the license.txt file included.
+
+#include <cfloat>
+#include <cmath>
+
+#include "VideoCommon/GeometryShaderGen.h"
+#include "VideoCommon/GeometryShaderManager.h"
+#include "VideoCommon/VideoCommon.h"
+#include "VideoCommon/VideoConfig.h"
+#include "VideoCommon/XFMemory.h"
+
+// track changes
+static bool s_projection_changed, s_viewport_changed;
+
+GeometryShaderConstants GeometryShaderManager::constants;
+bool GeometryShaderManager::dirty;
+
+void GeometryShaderManager::Init()
+{
+	memset(&constants, 0, sizeof(constants));
+
+	Dirty();
+}
+
+void GeometryShaderManager::Shutdown()
+{
+}
+
+void GeometryShaderManager::Dirty()
+{
+	s_projection_changed = true;
+	s_viewport_changed = true;
+
+	dirty = true;
+}
+
+// Syncs the shader constant buffers with xfmem
+void GeometryShaderManager::SetConstants()
+{
+	if (s_projection_changed)
+	{
+		s_projection_changed = false;
+
+		if (g_ActiveConfig.iStereoMode > 0 && xfmem.projection.type == GX_PERSPECTIVE)
+		{
+			float offset = (g_ActiveConfig.iStereoSeparation / 1000.0f) * (g_ActiveConfig.iStereoSeparationPercent / 100.0f);
+			constants.stereoparams[0] = (g_ActiveConfig.bStereoSwapEyes) ? offset : -offset;
+			constants.stereoparams[1] = (g_ActiveConfig.bStereoSwapEyes) ? -offset : offset;
+			constants.stereoparams[2] = (g_ActiveConfig.iStereoConvergence / 10.0f) * (g_ActiveConfig.iStereoConvergencePercent / 100.0f);
+		}
+		else
+		{
+			constants.stereoparams[0] = constants.stereoparams[1] = 0;
+		}
+	}
+
+	dirty = true;
+}
+
+void GeometryShaderManager::SetViewportChanged()
+{
+	s_viewport_changed = true;
+}
+
+void GeometryShaderManager::SetProjectionChanged()
+{
+	s_projection_changed = true;
+}
+
+void GeometryShaderManager::DoState(PointerWrap &p)
+{
+	p.Do(dirty);
+
+	if (p.GetMode() == PointerWrap::MODE_READ)
+	{
+		// Reload current state from global GPU state
+		// NOTE: This requires that all GPU memory has been loaded already.
+		Dirty();
+	}
+}

Source/Core/VideoCommon/GeometryShaderManager.h

@@ -0,0 +1,29 @@
+// Copyright 2013 Dolphin Emulator Project
+// Licensed under GPLv2
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include "VideoCommon/ConstantManager.h"
+#include "VideoCommon/GeometryShaderGen.h"
+
+class PointerWrap;
+
+// The non-API dependent parts.
+class GeometryShaderManager
+{
+public:
+	static void Init();
+	static void Dirty();
+	static void Shutdown();
+	static void DoState(PointerWrap &p);
+
+	// constant management
+	static void SetConstants();
+
+	static void SetViewportChanged();
+	static void SetProjectionChanged();
+
+	static GeometryShaderConstants constants;
+	static bool dirty;
+};

Source/Core/VideoCommon/ShaderGenCommon.h

@@ -279,7 +279,8 @@ static inline void GenerateVSOutputStruct(T& object, API_TYPE api_type)
 #define I_NORMALMATRICES        "cnmtx"
 #define I_POSTTRANSFORMMATRICES "cpostmtx"
 #define I_PIXELCENTERCORRECTION "cpixelcenter"
-#define I_STEREOPARAMS          "cstereo"
+
+#define I_STEREOPARAMS      "cstereo"
 
 static const char s_shader_uniforms[] =
 	"\tfloat4 " I_POSNORMALMATRIX"[6];\n"
@@ -290,5 +291,4 @@ static const char s_shader_uniforms[] =
 	"\tfloat4 " I_TRANSFORMMATRICES"[64];\n"
 	"\tfloat4 " I_NORMALMATRICES"[32];\n"
 	"\tfloat4 " I_POSTTRANSFORMMATRICES"[64];\n"
-	"\tfloat4 " I_PIXELCENTERCORRECTION";\n"
+	"\tfloat4 " I_PIXELCENTERCORRECTION";\n";
-	"\tfloat4 " I_STEREOPARAMS";\n";

Source/Core/VideoCommon/VertexManagerBase.cpp

@@ -2,6 +2,7 @@
 
 #include "VideoCommon/BPStructs.h"
 #include "VideoCommon/Debugger.h"
+#include "VideoCommon/GeometryShaderManager.h"
 #include "VideoCommon/IndexGenerator.h"
 #include "VideoCommon/MainBase.h"
 #include "VideoCommon/NativeVertexFormat.h"
@@ -223,6 +224,7 @@ void VertexManager::Flush()
 	// set global constants
 	VertexShaderManager::SetConstants();
 	PixelShaderManager::SetConstants();
+	GeometryShaderManager::SetConstants();
 
 	bool useDstAlpha = !g_ActiveConfig.bDstAlphaPass &&
 	                   bpmem.dstalpha.enable &&

Source/Core/VideoCommon/VertexShaderManager.cpp

@@ -512,18 +512,6 @@ void VertexShaderManager::SetConstants()
 			memcpy(constants.projection, correctedMtx.data, 4*16);
 		}
 
-		if (g_ActiveConfig.iStereoMode > 0 && xfmem.projection.type == GX_PERSPECTIVE)
-		{
-			float offset = (g_ActiveConfig.iStereoSeparation / 1000.0f) * (g_ActiveConfig.iStereoSeparationPercent / 100.0f);
-			constants.stereoparams[0] = (g_ActiveConfig.bStereoSwapEyes) ? offset : -offset;
-			constants.stereoparams[1] = (g_ActiveConfig.bStereoSwapEyes) ? -offset : offset;
-			constants.stereoparams[2] = (g_ActiveConfig.iStereoConvergence / 10.0f) * (g_ActiveConfig.iStereoConvergencePercent / 100.0f);
-		}
-		else
-		{
-			constants.stereoparams[0] = constants.stereoparams[1] = 0;
-		}
-
 		dirty = true;
 	}
 }

Source/Core/VideoCommon/VideoCommon.vcxproj

@@ -61,6 +61,7 @@
     <ClCompile Include="RenderBase.cpp" />
     <ClCompile Include="Statistics.cpp" />
     <ClCompile Include="GeometryShaderGen.cpp" />
+    <ClCompile Include="GeometryShaderManager.cpp" />
     <ClCompile Include="TextureCacheBase.cpp" />
     <ClCompile Include="TextureConversionShader.cpp" />
     <ClCompile Include="VertexLoader.cpp" />
@@ -112,6 +113,7 @@
     <ClInclude Include="ShaderGenCommon.h" />
     <ClInclude Include="Statistics.h" />
     <ClInclude Include="GeometryShaderGen.h" />
+    <ClInclude Include="GeometryShaderManager.h" />
     <ClInclude Include="TextureCacheBase.h" />
     <ClInclude Include="TextureConversionShader.h" />
     <ClInclude Include="TextureDecoder.h" />

Source/Core/VideoCommon/VideoCommon.vcxproj.filters

@@ -146,6 +146,9 @@
     <ClCompile Include="GeometryShaderGen.cpp">
       <Filter>Shader Generators</Filter>
     </ClCompile>
+    <ClCompile Include="GeometryShaderManager.cpp">
+      <Filter>Shader Managers</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="CommandProcessor.h" />
@@ -284,6 +287,9 @@
     <ClInclude Include="GeometryShaderGen.h">
       <Filter>Shader Generators</Filter>
     </ClInclude>
+    <ClInclude Include="GeometryShaderManager.h">
+      <Filter>Shader Managers</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <Text Include="CMakeLists.txt" />

Source/Core/VideoCommon/VideoState.cpp

@@ -8,6 +8,7 @@
 #include "VideoCommon/CommandProcessor.h"
 #include "VideoCommon/CPMemory.h"
 #include "VideoCommon/Fifo.h"
+#include "VideoCommon/GeometryShaderManager.h"
 #include "VideoCommon/PixelEngine.h"
 #include "VideoCommon/PixelShaderManager.h"
 #include "VideoCommon/TextureDecoder.h"
@@ -50,6 +51,9 @@ static void DoState(PointerWrap &p)
 	VertexShaderManager::DoState(p);
 	p.DoMarker("VertexShaderManager");
 
+	GeometryShaderManager::DoState(p);
+	p.DoMarker("GeometryShaderManager");
+
 	VertexManager::DoState(p);
 	p.DoMarker("VertexManager");
 

Source/Core/VideoCommon/XFStructs.cpp

@@ -7,6 +7,7 @@
 #include "VideoCommon/CPMemory.h"
 #include "VideoCommon/DataReader.h"
 #include "VideoCommon/Fifo.h"
+#include "VideoCommon/GeometryShaderManager.h"
 #include "VideoCommon/PixelShaderManager.h"
 #include "VideoCommon/VertexManagerBase.h"
 #include "VideoCommon/VertexShaderManager.h"
@@ -110,6 +111,7 @@ static void XFRegWritten(int transferSize, u32 baseAddress, DataReader src)
 			VertexManager::Flush();
 			VertexShaderManager::SetViewportChanged();
 			PixelShaderManager::SetViewportChanged();
+			GeometryShaderManager::SetViewportChanged();
 
 			nextAddress = XFMEM_SETVIEWPORT + 6;
 			break;
@@ -123,6 +125,7 @@ static void XFRegWritten(int transferSize, u32 baseAddress, DataReader src)
 		case XFMEM_SETPROJECTION+6:
 			VertexManager::Flush();
 			VertexShaderManager::SetProjectionChanged();
+			GeometryShaderManager::SetProjectionChanged();
 
 			nextAddress = XFMEM_SETPROJECTION + 7;
 			break;