Added VertexFixup setting

See issues #164 and #327.

DDrawCompat/Config/Config.cpp

@@ -48,6 +48,7 @@
 #include <Config/Settings/TextureFilter.h>
 #include <Config/Settings/ThreadPriorityBoost.h>
 #include <Config/Settings/VertexBufferMemoryType.h>
+#include <Config/Settings/VertexFixup.h>
 #include <Config/Settings/VSync.h>
 #include <Config/Settings/WinVersionLie.h>
 
@@ -103,6 +104,7 @@ namespace Config
 	Settings::TextureFilter textureFilter;
 	Settings::ThreadPriorityBoost threadPriorityBoost;
 	Settings::VertexBufferMemoryType vertexBufferMemoryType;
+	Settings::VertexFixup vertexFixup;
 	Settings::VSync vSync;
 	Settings::WinVersionLie winVersionLie;
 }

DDrawCompat/Config/Settings/VertexFixup.h

@@ -0,0 +1,22 @@
+#pragma once
+
+#include <Config/EnumSetting.h>
+
+namespace Config
+{
+	namespace Settings
+	{
+		class VertexFixup : public EnumSetting
+		{
+		public:
+			enum Values { CPU, GPU };
+
+			VertexFixup()
+				: EnumSetting("VertexFixup", "gpu", { "cpu", "gpu" })
+			{
+			}
+		};
+	}
+
+	extern Settings::VertexFixup vertexFixup;
+}

DDrawCompat/D3dDdi/DeviceState.cpp

@@ -5,6 +5,7 @@
 #include <Config/Settings/SpriteFilter.h>
 #include <Config/Settings/SpriteTexCoord.h>
 #include <Config/Settings/TextureFilter.h>
+#include <Config/Settings/VertexFixup.h>
 #include <Common/Comparison.h>
 #include <Common/Log.h>
 #include <D3dDdi/Device.h>
@@ -88,6 +89,8 @@ namespace D3dDdi
 		, m_vertexShaderConstB{}
 		, m_vertexShaderConstI{}
 		, m_vertexDecl(nullptr)
+		, m_vertexFixupConfig(Config::vertexFixup.get())
+		, m_vertexFixupData{}
 		, m_changedStates(0)
 		, m_maxChangedTextureStage(0)
 		, m_texCoordIndexes(0)
@@ -100,6 +103,8 @@ namespace D3dDdi
 		const UINT D3DBLENDOP_ADD = 1;
 		const UINT UNINITIALIZED_STATE = 0xBAADBAAD;
 
+		m_vertexFixupData.multiplier[3] = 1.0f;
+
 		m_device.getOrigVtable().pfnSetDepthStencil(m_device, &m_current.depthStencil);
 		m_device.getOrigVtable().pfnSetRenderTarget(m_device, &m_current.renderTarget);
 		m_device.getOrigVtable().pfnSetViewport(m_device, &m_current.viewport);
@@ -298,6 +303,10 @@ namespace D3dDdi
 		{
 			updateTextureStages();
 		}
+		if (m_changedStates & CS_VERTEX_FIXUP)
+		{
+			updateVertexFixupShaderConst();
+		}
 
 		m_changedStates = 0;
 		m_maxChangedTextureStage = 0;
@@ -1012,18 +1021,10 @@ namespace D3dDdi
 			auto resource = (texture == m_app.textures[stage]) ? getTextureResource(stage) : m_device.getResource(texture);
 			if (resource)
 			{
-				D3DDDIARG_SETVERTEXSHADERCONST data = {};
-				data.Register = 253;
-				data.Count = 1;
-
 				auto& si = resource->getFixedDesc().pSurfList[0];
-				ShaderConstF reg = { static_cast<float>(si.Width), static_cast<float>(si.Height),
+				m_vertexFixupData.texCoordAdj[0] = static_cast<float>(si.Width);
-					m_vertexShaderConst[253][2], m_vertexShaderConst[253][3] };
+				m_vertexFixupData.texCoordAdj[1] = static_cast<float>(si.Height);
-
+				m_changedStates |= CS_VERTEX_FIXUP;
-				if (0 != memcmp(&reg, &m_vertexShaderConst[data.Register], sizeof(reg)))
-				{
-					pfnSetVertexShaderConst(&data, &reg);
-				}
 			}
 		}
 
@@ -1107,6 +1108,7 @@ namespace D3dDdi
 		m_changedStates |= CS_RENDER_STATE | CS_RENDER_TARGET | CS_SHADER | CS_TEXTURE_STAGE;
 		m_changedRenderStates.set(D3DDDIRS_COLORKEYENABLE);
 		m_changedRenderStates.set(D3DDDIRS_MULTISAMPLEANTIALIAS);
+		m_vertexFixupConfig = Config::vertexFixup.get();
 
 		for (auto& ps : m_pixelShaders)
 		{
@@ -1211,7 +1213,7 @@ namespace D3dDdi
 		{
 			const float sx = static_cast<float>(scaledVp.Width) / vp.Width;
 			const float sy = static_cast<float>(scaledVp.Height) / vp.Height;
-			updateVertexFixupConstants(vp.Width, vp.Height, sx, sy);
+			updateVertexFixupData(vp.Width, vp.Height, sx, sy);
 		}
 
 		const bool isScissorRectNeeded = isTransformed && 0 != memcmp(&vp, &m_app.viewport, sizeof(vp));
@@ -1230,7 +1232,16 @@ namespace D3dDdi
 	{
 		setPixelShader(mapPixelShader(m_app.pixelShader));
 		setVertexShaderDecl(m_app.vertexShaderDecl);
-		setVertexShaderFunc(getVertexDecl().isTransformed ? getVsVertexFixup() : m_app.vertexShaderFunc);
+
+		if (Config::Settings::VertexFixup::GPU == m_vertexFixupConfig &&
+			getVertexDecl().isTransformed)
+		{
+			setVertexShaderFunc(getVsVertexFixup());
+		}
+		else
+		{
+			setVertexShaderFunc(m_app.vertexShaderFunc);
+		}
 	}
 
 	void DeviceState::updateStreamSource()
@@ -1321,25 +1332,51 @@ namespace D3dDdi
 		}
 	}
 
-	void DeviceState::updateVertexFixupConstants(UINT width, UINT height, float sx, float sy)
+	void DeviceState::updateVertexFixupData(UINT width, UINT height, float sx, float sy)
 	{
-		D3DDDIARG_SETVERTEXSHADERCONST data = {};
-		data.Register = 253;
-		data.Count = 3;
-
 		const float stc = static_cast<float>(Config::spriteTexCoord.getParam()) / 100;
 		const float apc = Config::alternatePixelCenter.get();
 		const auto& zr = m_current.zRange;
 
-		ShaderConstF registers[3] = {
+		m_vertexFixupData.texCoordAdj[2] = stc;
-			{ m_vertexShaderConst[253][0], m_vertexShaderConst[253][1], stc, stc },
+		m_vertexFixupData.texCoordAdj[3] = stc;
-			{ 0.5f + apc - 0.5f / sx - width / 2, 0.5f + apc - 0.5f / sy - height / 2, -zr.MinZ, 0.0f },
-			{ 2.0f / width, -2.0f / height, 1.0f / (zr.MaxZ - zr.MinZ), 1.0f }
-		};
 
-		if (0 != memcmp(registers, &m_vertexShaderConst[data.Register], sizeof(registers)))
+		if (Config::Settings::VertexFixup::GPU == m_vertexFixupConfig)
 		{
-			pfnSetVertexShaderConst(&data, registers);
+			m_vertexFixupData.offset[0] = 0.5f + apc - 0.5f / sx - width / 2;
+			m_vertexFixupData.offset[1] = 0.5f + apc - 0.5f / sy - height / 2;
+			m_vertexFixupData.offset[2] = -zr.MinZ;
+
+			m_vertexFixupData.multiplier[0] = 2.0f / width;
+			m_vertexFixupData.multiplier[1] = -2.0f / height;
+			m_vertexFixupData.multiplier[2] = 1.0f / (zr.MaxZ - zr.MinZ);
+		}
+		else
+		{
+			m_vertexFixupData.offset[0] = 0.5f + apc - 0.5f / sx;
+			m_vertexFixupData.offset[1] = 0.5f + apc - 0.5f / sy;
+
+			m_vertexFixupData.multiplier[0] = sx;
+			m_vertexFixupData.multiplier[1] = sy;
+		}
+
+		m_changedStates |= CS_VERTEX_FIXUP;
+	}
+
+	void DeviceState::updateVertexFixupShaderConst()
+	{
+		if (m_current.vertexShaderFunc == m_app.vertexShaderFunc)
+		{
+			return;
+		}
+
+		D3DDDIARG_SETVERTEXSHADERCONST data = {};
+		data.Register = 253;
+		data.Count = 3;
+
+		if (0 != memcmp(&m_vertexFixupData, &m_vertexShaderConst[data.Register], sizeof(m_vertexFixupData)))
+		{
+			pfnSetVertexShaderConst(&data, &m_vertexFixupData.texCoordAdj);
 		}
 	}
 }

DDrawCompat/D3dDdi/DeviceState.h

@@ -109,6 +109,13 @@ namespace D3dDdi
 			bool isTransformed;
 		};
 
+		struct VertexFixupData
+		{
+			ShaderConstF texCoordAdj;
+			ShaderConstF offset;
+			ShaderConstF multiplier;
+		};
+
 		DeviceState(Device& device);
 		
 		HRESULT pfnCreatePixelShader(D3DDDIARG_CREATEPIXELSHADER* data, const UINT* code);
@@ -153,9 +160,11 @@ namespace D3dDdi
 		void flush();
 		const State& getAppState() const { return m_app; }
 		const State& getCurrentState() const { return m_current; }
+		bool getSpriteMode() const { return m_spriteMode; }
 		Resource* getTextureResource(UINT stage);
 		UINT getTextureStageCount() const;
 		const VertexDecl& getVertexDecl() const;
+		const VertexFixupData& getVertexFixupData() const { return m_vertexFixupData; }
 		bool isLocked() const { return m_isLocked; }
 		void onDestroyResource(Resource* resource, HANDLE resourceHandle);
 		void updateConfig();
@@ -168,7 +177,8 @@ namespace D3dDdi
 			CS_RENDER_TARGET = 1 << 1,
 			CS_SHADER        = 1 << 2,
 			CS_STREAM_SOURCE = 1 << 3,
-			CS_TEXTURE_STAGE = 1 << 4
+			CS_TEXTURE_STAGE = 1 << 4,
+			CS_VERTEX_FIXUP  = 1 << 5
 		};
 
 		struct PixelShader
@@ -230,7 +240,8 @@ namespace D3dDdi
 		void updateShaders();
 		void updateTextureColorKey(UINT stage);
 		void updateTextureStages();
-		void updateVertexFixupConstants(UINT width, UINT height, float sx, float sy);
+		void updateVertexFixupData(UINT width, UINT height, float sx, float sy);
+		void updateVertexFixupShaderConst();
 
 		Device& m_device;
 		State m_app;
@@ -243,6 +254,8 @@ namespace D3dDdi
 		std::array<ShaderConstI, 16> m_vertexShaderConstI;
 		std::map<HANDLE, VertexDecl> m_vertexShaderDecls;
 		VertexDecl* m_vertexDecl;
+		UINT m_vertexFixupConfig;
+		VertexFixupData m_vertexFixupData;
 		UINT m_changedStates;
 		UINT m_maxChangedTextureStage;
 		UINT m_texCoordIndexes;

DDrawCompat/D3dDdi/DrawPrimitive.cpp

@@ -12,6 +12,14 @@ namespace
 	const UINT INDEX_BUFFER_SIZE = D3DMAXNUMPRIMITIVES * 3 * sizeof(UINT16);
 	const UINT VERTEX_BUFFER_SIZE = 1024 * 1024;
 
+	enum VertexFixupFlags
+	{
+		VF_XY       = 1 << 0,
+		VF_Z        = 1 << 1,
+		VF_RHW      = 1 << 2,
+		VF_TEXCOORD = 1 << 3
+	};
+
 	UINT getVertexCount(D3DPRIMITIVETYPE primitiveType, UINT primitiveCount)
 	{
 		switch (primitiveType)
@@ -57,6 +65,7 @@ namespace D3dDdi
 		, m_indexBuffer(device, m_vertexBuffer ? INDEX_BUFFER_SIZE : 0)
 		, m_streamSource{}
 		, m_batched{}
+		, m_vertexFixupFlags(0)
 	{
 		LOG_ONCE("Dynamic vertex buffers are " << (m_vertexBuffer ? "" : "not ") << "available");
 		LOG_ONCE("Dynamic index buffers are " << (m_indexBuffer ? "" : "not ") << "available");
@@ -325,6 +334,74 @@ namespace D3dDdi
 	{
 		auto vertices = m_streamSource.vertices + base * m_streamSource.stride;
 		m_batched.vertices.insert(m_batched.vertices.end(), vertices, vertices + count * m_streamSource.stride);
+
+		if (0 == m_vertexFixupFlags)
+		{
+			return;
+		}
+
+		auto& vertexFixupData = m_device.getState().getVertexFixupData();
+		auto firstVertex = &m_batched.vertices[m_batched.vertices.size() - count * m_streamSource.stride];
+
+		if (m_vertexFixupFlags & VF_XY)
+		{
+			auto vPos = firstVertex;
+			for (unsigned i = 0; i < count; ++i)
+			{
+				auto v = reinterpret_cast<D3DTLVERTEX*>(vPos);
+				v->sx += vertexFixupData.offset[0];
+				v->sy += vertexFixupData.offset[1];
+				v->sx *= vertexFixupData.multiplier[0];
+				v->sy *= vertexFixupData.multiplier[1];
+				vPos += m_streamSource.stride;
+			}
+		}
+
+		if (m_vertexFixupFlags & VF_Z)
+		{
+			auto zPos = reinterpret_cast<BYTE*>(&reinterpret_cast<D3DTLVERTEX*>(firstVertex)->sz);
+			for (unsigned i = 0; i < count; ++i)
+			{
+				auto& z = *reinterpret_cast<float*>(zPos);
+				if (z > 1)
+				{
+					z = 1;
+				}
+				zPos += m_streamSource.stride;
+			}
+		}
+
+		if (m_vertexFixupFlags & VF_RHW)
+		{
+			auto rhwPos = reinterpret_cast<BYTE*>(&reinterpret_cast<D3DTLVERTEX*>(firstVertex)->rhw);
+			for (unsigned i = 0; i < count; ++i)
+			{
+				auto& rhw = *reinterpret_cast<float*>(rhwPos);
+				if (0 == rhw || INFINITY == rhw)
+				{
+					rhw = 1;
+				}
+				rhwPos += m_streamSource.stride;
+			}
+		}
+
+		if (m_vertexFixupFlags & VF_TEXCOORD)
+		{
+			auto tcPos = firstVertex + m_device.getState().getVertexDecl().texCoordOffset[0];
+			for (unsigned i = 0; i < count; ++i)
+			{
+				float* tc = reinterpret_cast<float*>(tcPos);
+				tc[0] *= vertexFixupData.texCoordAdj[0];
+				tc[1] *= vertexFixupData.texCoordAdj[1];
+				tc[0] += vertexFixupData.texCoordAdj[2];
+				tc[1] += vertexFixupData.texCoordAdj[3];
+				tc[0] = roundf(tc[0]);
+				tc[1] = roundf(tc[1]);
+				tc[0] /= vertexFixupData.texCoordAdj[0];
+				tc[1] /= vertexFixupData.texCoordAdj[1];
+				tcPos += m_streamSource.stride;
+			}
+		}
 	}
 
 	void DrawPrimitive::clearBatchedPrimitives()
@@ -456,6 +533,7 @@ namespace D3dDdi
 	{
 		auto& state = m_device.getState();
 		auto vertexCount = getVertexCount(data.PrimitiveType, data.PrimitiveCount);
+		m_vertexFixupFlags = 0;
 		if (!state.isLocked())
 		{
 			if (0 == m_batched.primitiveCount)
@@ -477,6 +555,7 @@ namespace D3dDdi
 				state.setSpriteMode(false);
 			}
 			state.flush();
+			setVertexFixupFlags(data.VStart, 0);
 		}
 
 		if (0 == m_batched.primitiveCount || flagBuffer ||
@@ -521,6 +600,7 @@ namespace D3dDdi
 
 		auto indexCount = getVertexCount(data.PrimitiveType, data.PrimitiveCount);
 		auto vStart = data.BaseVertexOffset / static_cast<INT>(m_streamSource.stride);
+		m_vertexFixupFlags = 0;
 		if (!state.isLocked())
 		{
 			if (m_streamSource.vertices && data.PrimitiveType >= D3DPT_TRIANGLELIST)
@@ -537,6 +617,7 @@ namespace D3dDdi
 				state.setSpriteMode(false);
 			}
 			state.flush();
+			setVertexFixupFlags(vStart, indices[0]);
 		}
 
 		auto [min, max] = std::minmax_element(indices, indices + indexCount);
@@ -870,4 +951,44 @@ namespace D3dDdi
 			}
 		}
 	}
+
+	void DrawPrimitive::setVertexFixupFlags(INT baseVertexIndex, UINT16 index)
+	{
+		auto& state = m_device.getState();
+		if (!state.getVertexDecl().isTransformed ||
+			state.getCurrentState().vertexShaderFunc != state.getAppState().vertexShaderFunc)
+		{
+			return;
+		}
+
+		auto& vertexFixupData = state.getVertexFixupData();
+		if (0 != vertexFixupData.offset[0] ||
+			0 != vertexFixupData.offset[1] ||
+			1 != vertexFixupData.multiplier[0] ||
+			1 != vertexFixupData.multiplier[1])
+		{
+			m_vertexFixupFlags |= VF_XY;
+		}
+
+		auto vertex = reinterpret_cast<const D3DTLVERTEX*>(
+			m_streamSource.vertices + (baseVertexIndex + index) * m_streamSource.stride);
+		if (vertex->sz > 0)
+		{
+			m_vertexFixupFlags |= VF_Z;
+		}
+
+		if (0 == vertex->rhw || INFINITY == vertex->rhw)
+		{
+			m_vertexFixupFlags |= VF_RHW;
+		}
+
+		const UINT D3DDECLTYPE_FLOAT2 = 1;
+		if (state.getSpriteMode() &&
+			Config::Settings::SpriteTexCoord::ROUND == Config::spriteTexCoord.get() &&
+			0 != Config::spriteTexCoord.getParam() &&
+			D3DDECLTYPE_FLOAT2 == state.getVertexDecl().texCoordType[0])
+		{
+			m_vertexFixupFlags |= VF_TEXCOORD;
+		}
+	}
 }

DDrawCompat/D3dDdi/DrawPrimitive.h

@@ -71,6 +71,7 @@ namespace D3dDdi
 		HRESULT flush(const UINT* flagBuffer);
 		HRESULT flushIndexed(const UINT* flagBuffer);
 		bool isSprite(INT baseVertexIndex, UINT16 index0, UINT16 index1, UINT16 index2);
+		void setVertexFixupFlags(INT baseVertexIndex, UINT16 index);
 		INT loadIndices(const void* indices, UINT count);
 		INT loadVertices(UINT count);
 		UINT getBatchedVertexCount() const;
@@ -87,5 +88,6 @@ namespace D3dDdi
 		StreamSource m_streamSource;
 		std::map<HANDLE, BYTE*> m_sysMemVertexBuffers;
 		BatchedPrimitives m_batched;
+		UINT m_vertexFixupFlags;
 	};
 }

DDrawCompat/DDrawCompat.vcxproj

@@ -214,6 +214,7 @@
     <ClInclude Include="Config\Settings\TextureFilter.h" />
     <ClInclude Include="Config\Settings\ThreadPriorityBoost.h" />
     <ClInclude Include="Config\Settings\VertexBufferMemoryType.h" />
+    <ClInclude Include="Config\Settings\VertexFixup.h" />
     <ClInclude Include="Config\Settings\VSync.h" />
     <ClInclude Include="Config\Settings\WinVersionLie.h" />
     <ClInclude Include="D3dDdi\Adapter.h" />

DDrawCompat/DDrawCompat.vcxproj.filters

@@ -717,6 +717,9 @@
     <ClInclude Include="Common\Disasm.h">
       <Filter>Header Files\Common</Filter>
     </ClInclude>
+    <ClInclude Include="Config\Settings\VertexFixup.h">
+      <Filter>Header Files\Config\Settings</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="Gdi\Gdi.cpp">

DDrawCompat/Overlay/ConfigWindow.cpp

@@ -25,6 +25,7 @@
 #include <Config/Settings/StatsPosY.h>
 #include <Config/Settings/StatsTransparency.h>
 #include <Config/Settings/TextureFilter.h>
+#include <Config/Settings/VertexFixup.h>
 #include <Config/Settings/VSync.h>
 #include <D3dDdi/Device.h>
 #include <Gdi/GuiThread.h>
@@ -66,6 +67,7 @@ namespace
 		{ &Config::statsPosY, []() { Gdi::GuiThread::getStatsWindow()->updatePos(); } },
 		{ &Config::statsTransparency, [&]() { Gdi::GuiThread::getStatsWindow()->setAlpha(Config::statsTransparency.get()); }},
 		{ &Config::textureFilter, &D3dDdi::Device::updateAllConfig },
+		{ &Config::vertexFixup, &D3dDdi::Device::updateAllConfig },
 		{ &Config::vSync }
 	};
 }