Fix alignment of projected vertex coordinates

src/Device/Config.hpp

@@ -66,7 +66,7 @@ namespace sw
 		MAX_TEXTURE_LOD = MIPMAP_LEVELS - 2,   // Trilinear accesses lod+1
 		RENDERTARGETS = 8,
 		NUM_TEMPORARY_REGISTERS = 4096,
-		MAX_INTERFACE_COMPONENTS = 32 * 4,
+		MAX_INTERFACE_COMPONENTS = 32 * 4,  // Must be multiple of 4 for 16-byte alignment.
 	};
 }
 

src/Device/Renderer.cpp

@@ -888,7 +888,7 @@ namespace sw
 
 			if((v0.clipFlags & v1.clipFlags & v2.clipFlags) == Clipper::CLIP_FINITE)
 			{
-				Polygon polygon(&v0.builtins.position, &v1.builtins.position, &v2.builtins.position);
+				Polygon polygon(&v0.position, &v1.position, &v2.position);
 
 				int clipFlagsOr = v0.clipFlags | v1.clipFlags | v2.clipFlags;
 
@@ -972,8 +972,8 @@ namespace sw
 		Vertex &v0 = triangle.v0;
 		Vertex &v1 = triangle.v1;
 
-		const float4 &P0 = v0.builtins.position;
-		const float4 &P1 = v1.builtins.position;
+		const float4 &P0 = v0.position;
+		const float4 &P1 = v1.position;
 
 		if(P0.w <= 0 && P1.w <= 0)
 		{
@@ -1162,17 +1162,17 @@ namespace sw
 
 		Vertex &v = triangle.v0;
 
-		float pSize = v.builtins.pointSize;
+		float pSize = v.pointSize;
 
 		pSize = clamp(pSize, 1.0f, static_cast<float>(vk::MAX_POINT_SIZE));
 
 		float4 P[4];
 		int C[4];
 
-		P[0] = v.builtins.position;
-		P[1] = v.builtins.position;
-		P[2] = v.builtins.position;
-		P[3] = v.builtins.position;
+		P[0] = v.position;
+		P[1] = v.position;
+		P[2] = v.position;
+		P[3] = v.position;
 
 		const float X = pSize * P[0].w * data.halfPixelX[0];
 		const float Y = pSize * P[0].w * data.halfPixelY[0];

src/Device/Vertex.hpp

@@ -23,14 +23,12 @@ namespace sw
 {
 	ALIGN(16, struct Vertex
 	{
-		float v[MAX_INTERFACE_COMPONENTS];
+		float4 position;
+		float pointSize;
 
-		struct
-		{
-			float4 position;
-			float pointSize;
-		} builtins;
-		struct
+		int clipFlags;
+
+		alignas(16) struct
 		{
 			int x;
 			int y;
@@ -38,8 +36,7 @@ namespace sw
 			float w;
 		} projected;
 
-		int clipFlags;
-		int padding[2];
+		alignas(16) float v[MAX_INTERFACE_COMPONENTS];
 	});
 
 	static_assert((sizeof(Vertex) & 0x0000000F) == 0, "Vertex size not a multiple of 16 bytes (alignment requirement)");

src/Pipeline/SetupRoutine.cpp

@@ -86,9 +86,9 @@ namespace sw
 					Return(0);
 				}
 
-				Int w0w1w2 = *Pointer<Int>(v0 + OFFSET(Vertex, builtins.position.w)) ^
-							 *Pointer<Int>(v1 + OFFSET(Vertex, builtins.position.w)) ^
-							 *Pointer<Int>(v2 + OFFSET(Vertex, builtins.position.w));
+				Int w0w1w2 = *Pointer<Int>(v0 + OFFSET(Vertex, position.w)) ^
+							 *Pointer<Int>(v1 + OFFSET(Vertex, position.w)) ^
+							 *Pointer<Int>(v2 + OFFSET(Vertex, position.w));
 
 				A = IfThenElse(w0w1w2 < 0, -A, A);
 
@@ -268,9 +268,9 @@ namespace sw
 			// Sort by minimum y
 			if(triangle)
 			{
-				Float y0 = *Pointer<Float>(v0 + OFFSET(Vertex, builtins.position.y));
-				Float y1 = *Pointer<Float>(v1 + OFFSET(Vertex, builtins.position.y));
-				Float y2 = *Pointer<Float>(v2 + OFFSET(Vertex, builtins.position.y));
+				Float y0 = *Pointer<Float>(v0 + OFFSET(Vertex, position.y));
+				Float y1 = *Pointer<Float>(v1 + OFFSET(Vertex, position.y));
+				Float y2 = *Pointer<Float>(v2 + OFFSET(Vertex, position.y));
 
 				Float yMin = Min(Min(y0, y1), y2);
 
@@ -281,9 +281,9 @@ namespace sw
 			// Sort by maximum w
 			if(triangle)
 			{
-				Float w0 = *Pointer<Float>(v0 + OFFSET(Vertex, builtins.position.w));
-				Float w1 = *Pointer<Float>(v1 + OFFSET(Vertex, builtins.position.w));
-				Float w2 = *Pointer<Float>(v2 + OFFSET(Vertex, builtins.position.w));
+				Float w0 = *Pointer<Float>(v0 + OFFSET(Vertex, position.w));
+				Float w1 = *Pointer<Float>(v1 + OFFSET(Vertex, position.w));
+				Float w2 = *Pointer<Float>(v2 + OFFSET(Vertex, position.w));
 
 				Float wMax = Max(Max(w0, w1), w2);
 
@@ -292,13 +292,13 @@ namespace sw
 			}
 
 			*Pointer<Float>(primitive + OFFSET(Primitive, pointCoordX)) =
-				*Pointer<Float>(v0 + OFFSET(Vertex, builtins.position.x));
+				*Pointer<Float>(v0 + OFFSET(Vertex, position.x));
 			*Pointer<Float>(primitive + OFFSET(Primitive, pointCoordY)) =
-				*Pointer<Float>(v0 + OFFSET(Vertex, builtins.position.y));
+				*Pointer<Float>(v0 + OFFSET(Vertex, position.y));
 
-			Float w0 = *Pointer<Float>(v0 + OFFSET(Vertex, builtins.position.w));
-			Float w1 = *Pointer<Float>(v1 + OFFSET(Vertex, builtins.position.w));
-			Float w2 = *Pointer<Float>(v2 + OFFSET(Vertex, builtins.position.w));
+			Float w0 = *Pointer<Float>(v0 + OFFSET(Vertex, position.w));
+			Float w1 = *Pointer<Float>(v1 + OFFSET(Vertex, position.w));
+			Float w2 = *Pointer<Float>(v2 + OFFSET(Vertex, position.w));
 
 			Float4 w012;
 

src/Pipeline/VertexRoutine.hpp

@@ -15,10 +15,10 @@
 #ifndef sw_VertexRoutine_hpp
 #define sw_VertexRoutine_hpp
 
-#include "Device/Color.hpp"
-#include "Device/VertexProcessor.hpp"
 #include "ShaderCore.hpp"
 #include "SpirvShader.hpp"
+#include "Device/Color.hpp"
+#include "Device/VertexProcessor.hpp"
 
 namespace vk
 {
@@ -27,7 +27,7 @@ namespace vk
 
 namespace sw
 {
-	class VertexRoutinePrototype : public Function<Void(Pointer<Byte>, Pointer<Byte>, Pointer<Byte>, Pointer<Byte>)>
+	class VertexRoutinePrototype : public Function<Void(Pointer<Byte>, Pointer<UInt>, Pointer<Byte>, Pointer<Byte>)>
 	{
 	public:
 		VertexRoutinePrototype() : vertex(Arg<0>()), batch(Arg<1>()), task(Arg<2>()), data(Arg<3>()) {}
@@ -35,7 +35,7 @@ namespace sw
 
 	protected:
 		Pointer<Byte> vertex;
-		Pointer<Byte> batch;
+		Pointer<UInt> batch;
 		Pointer<Byte> task;
 		Pointer<Byte> data;
 	};
@@ -70,7 +70,7 @@ namespace sw
 		void readInput(UInt &index);
 		void computeClipFlags();
 		void writeCache(Pointer<Byte> &cacheLine);
-		void writeVertex(const Pointer<Byte> &vertex, Pointer<Byte> &cacheLine);
+		void writeVertex(const Pointer<Byte> &vertex, Pointer<Byte> &cacheEntry);
 	};
 }