Make the number of vertex outputs configurable.

src/D3D9/Capabilities.hpp

@@ -15,6 +15,9 @@
 #ifndef D3D9_Capabilities_hpp
 #define D3D9_Capabilities_hpp
 
+#include "Config.hpp"
+#include "MetaMacro.hpp"
+
 #include <d3d9.h>
 
 namespace D3D9
@@ -465,6 +468,23 @@ namespace D3D9
 
 	extern unsigned int textureMemory;
 	extern unsigned int maxAnisotropy;
+
+	enum
+	{
+		MAX_VERTEX_SHADER_CONST = 256,
+		MAX_PIXEL_SHADER_CONST = 224,
+		MAX_VERTEX_OUTPUTS = 12,
+	};
+
+	// Shader Model 3.0 requirements
+	META_ASSERT(MAX_VERTEX_SHADER_CONST >= 256);
+	META_ASSERT(MAX_PIXEL_SHADER_CONST == 224);
+	META_ASSERT(MAX_VERTEX_OUTPUTS == 12);
+
+	// Back-end minimum requirements
+	META_ASSERT(sw::VERTEX_UNIFORM_VECTORS >= MAX_VERTEX_SHADER_CONST);
+	META_ASSERT(sw::FRAGMENT_UNIFORM_VECTORS >= MAX_PIXEL_SHADER_CONST);
+	META_ASSERT(sw::MAX_VERTEX_OUTPUTS >= MAX_VERTEX_OUTPUTS);
 }
 
 #endif   // D3D9_Capabilities_hpp

src/D3D9/Direct3DDevice9.cpp

@@ -5800,7 +5800,7 @@ namespace D3D9
 				}
 				else   // Bind directly to the output
 				{
-					for(int i = 0; i < 12; i++)
+					for(int i = 0; i < MAX_VERTEX_OUTPUTS; i++)
 					{
 						if((usage == shader->output[i][0].usage || (usage == D3DDECLUSAGE_POSITIONT && shader->output[i][0].usage == D3DDECLUSAGE_POSITION)) &&
 						    index == shader->output[i][0].index)

src/D3D9/Direct3DDevice9.hpp

@@ -19,6 +19,7 @@
 
 #include "Direct3D9.hpp"
 #include "Direct3DSwapChain9.hpp"
+#include "Capabilities.hpp"
 
 #include "Stream.hpp"
 
@@ -45,12 +46,6 @@ namespace D3D9
 	class Direct3DIndexBuffer9;
 	class CriticalSection;
 
-	enum
-	{
-		MAX_VERTEX_SHADER_CONST = MIN(256, sw::VERTEX_UNIFORM_VECTORS),
-		MAX_PIXEL_SHADER_CONST = MIN(224, sw::FRAGMENT_UNIFORM_VECTORS),
-	};
-
 	class Direct3DDevice9 : public IDirect3DDevice9, public Unknown
 	{
 		friend CriticalSection;

src/Main/Config.hpp

@@ -87,6 +87,7 @@ namespace sw
 		TOTAL_IMAGE_UNITS = TEXTURE_IMAGE_UNITS + VERTEX_TEXTURE_IMAGE_UNITS,
 		FRAGMENT_UNIFORM_VECTORS = 224,
 		VERTEX_UNIFORM_VECTORS = 256,
+		MAX_VERTEX_OUTPUTS = 12,
 		MAX_FRAGMENT_UNIFORM_BLOCKS = 12,
 		MAX_VERTEX_UNIFORM_BLOCKS = 12,
 		MAX_UNIFORM_BUFFER_BINDINGS = MAX_FRAGMENT_UNIFORM_BLOCKS + MAX_VERTEX_UNIFORM_BLOCKS,   // Limited to 127 by SourceParameter.bufferIndex in Shader.hpp

src/OpenGL/compiler/OutputASM.cpp

@@ -2787,7 +2787,7 @@ namespace glsl
 			}
 			else if(vertexShader)
 			{
-				if((var + registerCount) > sw::VertexShader::MAX_OUTPUT_VARYINGS)
+				if((var + registerCount) > sw::MAX_VERTEX_OUTPUTS)
 				{
 					mContext.error(varying->getLine(), "Varyings packing failed: Too many varyings", "vertex shader");
 					return 0;

src/Renderer/Renderer.cpp

@@ -1878,7 +1878,7 @@ namespace sw
 
 		int pts = state.pointSizeRegister;
 
-		if(state.pointSizeRegister != 0xF)
+		if(state.pointSizeRegister != Unused)
 		{
 			pSize = v.v[pts].y;
 		}

src/Renderer/SetupProcessor.cpp

@@ -89,7 +89,7 @@ namespace sw
 		state.vFace = context->pixelShader && context->pixelShader->vFaceDeclared;
 
 		state.positionRegister = Pos;
-		state.pointSizeRegister = 0xF;   // No vertex point size
+		state.pointSizeRegister = Unused;
 
 		state.multiSample = context->getMultiSampleCount();
 		state.rasterizerDiscard = context->rasterizerDiscard;
@@ -133,7 +133,7 @@ namespace sw
 					if(context->pixelShader->semantic[interpolant][component - project].active())
 					{
 						int input = interpolant;
-						for(int i = 0; i < 12; i++)
+						for(int i = 0; i < MAX_VERTEX_OUTPUTS; i++)
 						{
 							if(context->pixelShader->semantic[interpolant][component - project] == context->vertexShader->output[i][component - project])
 							{

src/Renderer/SetupProcessor.hpp

@@ -45,8 +45,8 @@ namespace sw
 			bool interpolateW              : 1;
 			bool perspective               : 1;
 			bool pointSprite               : 1;
-			unsigned int positionRegister  : 4;
-			unsigned int pointSizeRegister : 4;
+			unsigned int positionRegister  : BITS(VERTEX_OUTPUT_LAST);
+			unsigned int pointSizeRegister : BITS(VERTEX_OUTPUT_LAST);
 			CullMode cullMode              : BITS(CULL_LAST);
 			bool twoSidedStencil           : 1;
 			bool slopeDepthBias            : 1;
@@ -56,7 +56,7 @@ namespace sw
 
 			struct Gradient
 			{
-				unsigned char attribute : BITS(Unused);
+				unsigned char attribute : BITS(VERTEX_OUTPUT_LAST);
 				bool flat               : 1;
 				bool wrap               : 1;
 			};

src/Renderer/Vertex.hpp

@@ -18,11 +18,13 @@
 #include "Color.hpp"
 #include "Common/MetaMacro.hpp"
 #include "Common/Types.hpp"
+#include "Main/Config.hpp"
 
 namespace sw
 {
-	enum Out   // Default vertex output semantic
+	enum Out
 	{
+		// Default vertex output semantics
 		Pos = 0,
 		C0 = 1,   // Diffuse
 		C1 = 2,   // Specular
@@ -36,7 +38,13 @@ namespace sw
 		T7 = 10,
 		Fog = 11,    // x component
 		Pts = Fog,   // y component
-		Unused
+
+		// Variable semantics
+		V0 = 0,
+		Vn_1 = MAX_VERTEX_OUTPUTS - 1,
+
+		Unused,
+		VERTEX_OUTPUT_LAST = Unused,
 	};
 
 	struct UVWQ
@@ -72,7 +80,7 @@ namespace sw
 				float pSize;         // Point size
 			};
 
-			float4 v[12];   // Generic components using semantic declaration
+			float4 v[MAX_VERTEX_OUTPUTS];   // Generic components using semantic declaration
 		};
 
 		// Projected coordinates

src/Renderer/VertexProcessor.cpp

@@ -942,7 +942,7 @@ namespace sw
 
 		if(context->vertexShader)   // FIXME: Also when pre-transformed?
 		{
-			for(int i = 0; i < 12; i++)
+			for(int i = 0; i < MAX_VERTEX_OUTPUTS; i++)
 			{
 				state.output[i].xWrite = context->vertexShader->output[i][0].active();
 				state.output[i].yWrite = context->vertexShader->output[i][1].active();

src/Renderer/VertexProcessor.hpp

@@ -52,8 +52,8 @@ namespace sw
 
 			bool fixedFunction             : 1;
 			bool textureSampling           : 1;
-			unsigned int positionRegister  : 4;
-			unsigned int pointSizeRegister : 4;   // 0xF signifies no vertex point size
+			unsigned int positionRegister  : BITS(MAX_VERTEX_OUTPUTS);
+			unsigned int pointSizeRegister : BITS(MAX_VERTEX_OUTPUTS);
 
 			unsigned int vertexBlendMatrixCount               : 3;
 			bool indexedVertexBlendEnable                     : 1;
@@ -134,7 +134,7 @@ namespace sw
 			};
 
 			Input input[VERTEX_ATTRIBUTES];
-			Output output[12];
+			Output output[MAX_VERTEX_OUTPUTS];
 		};
 
 		struct State : States

src/Shader/VertexProgram.cpp

@@ -594,7 +594,7 @@ namespace sw
 	{
 		if(shader)
 		{
-			for(int i = 0; i < 12; i++)
+			for(int i = 0; i < MAX_VERTEX_OUTPUTS; i++)
 			{
 				unsigned char usage = shader->output[i][0].usage;
 

src/Shader/VertexRoutine.cpp

@@ -580,7 +580,7 @@ namespace sw
 	{
 		Vector4f v;
 
-		for(int i = 0; i < 12; i++)
+		for(int i = 0; i < MAX_VERTEX_OUTPUTS; i++)
 		{
 			if(state.output[i].write)
 			{
@@ -669,7 +669,7 @@ namespace sw
 
 	void VertexRoutine::writeVertex(const Pointer<Byte> &vertex, Pointer<Byte> &cache)
 	{
-		for(int i = 0; i < 12; i++)
+		for(int i = 0; i < MAX_VERTEX_OUTPUTS; i++)
 		{
 			if(state.output[i].write)
 			{

src/Shader/VertexRoutine.hpp

@@ -49,7 +49,7 @@ namespace sw
 		Int clipFlags;
 
 		RegisterArray<16> v;   // Varying registers
-		RegisterArray<12> o;   // Output registers
+		RegisterArray<MAX_VERTEX_OUTPUTS> o;   // Output registers
 
 		const VertexProcessor::State &state;
 

src/Shader/VertexShader.cpp

@@ -25,7 +25,7 @@ namespace sw
 	{
 		version = 0x0300;
 		positionRegister = Pos;
-		pointSizeRegister = -1;   // No vertex point size
+		pointSizeRegister = Unused;
 		instanceIdDeclared = false;
 
 		for(int i = 0; i < MAX_INPUT_ATTRIBUTES; i++)
@@ -57,7 +57,7 @@ namespace sw
 		parse(token);
 
 		positionRegister = Pos;
-		pointSizeRegister = -1;   // No vertex point size
+		pointSizeRegister = Unused;
 		instanceIdDeclared = false;
 
 		for(int i = 0; i < MAX_INPUT_ATTRIBUTES; i++)

src/Shader/VertexShader.hpp

@@ -16,6 +16,7 @@
 #define sw_VertexShader_hpp
 
 #include "Shader.hpp"
+#include "Main/Config.hpp"
 
 namespace sw
 {
@@ -40,8 +41,7 @@ namespace sw
 		enum {MAX_INPUT_ATTRIBUTES = 16};
 		Semantic input[MAX_INPUT_ATTRIBUTES];       // FIXME: Private
 
-		enum {MAX_OUTPUT_VARYINGS = 12};
-		Semantic output[MAX_OUTPUT_VARYINGS][4];   // FIXME: Private
+		Semantic output[MAX_VERTEX_OUTPUTS][4];   // FIXME: Private
 
 	private:
 		void analyzeInput();