Add epilog pass to copy outputs out to pipeline

src/Pipeline/PixelProgram.cpp

@@ -42,6 +42,7 @@ namespace sw
 		}
 
 		spirvShader->emit(&routine);
+		spirvShader->emitEpilog(&routine);
 
 		for(int i = 0; i < RENDERTARGETS; i++)
 		{

src/Pipeline/PixelRoutine.cpp

@@ -32,7 +32,7 @@ namespace sw
 	PixelRoutine::PixelRoutine(const PixelProcessor::State &state, SpirvShader const *spirvShader)
 		: QuadRasterizer(state, spirvShader)	/* addressing */
 	{
-		spirvShader->emitEarly(&routine);
+		spirvShader->emitProlog(&routine);
 
 		if (forceClearRegisters)
 		{

src/Pipeline/SpirvShader.cpp

@@ -643,7 +643,7 @@ namespace sw
 
 	// emit-time
 
-	void SpirvShader::emitEarly(SpirvRoutine *routine) const
+	void SpirvShader::emitProlog(SpirvRoutine *routine) const
 	{
 		for (auto insn : *this)
 		{
@@ -795,4 +795,32 @@ namespace sw
 			}
 		}
 	}
+
+	void SpirvShader::emitEpilog(SpirvRoutine *routine) const
+	{
+		for (auto insn : *this)
+		{
+			switch (insn.opcode())
+			{
+			case spv::OpVariable:
+			{
+				auto resultId = insn.word(2);
+				auto &object = getObject(resultId);
+				if (object.kind == Object::Kind::InterfaceVariable && object.storageClass == spv::StorageClassOutput)
+				{
+					auto &dst = routine->getValue(resultId);
+					int offset = 0;
+					VisitInterface(resultId,
+								   [&](Decorations const &d, AttribType type) {
+									   auto scalarSlot = d.Location << 2 | d.Component;
+									   routine->outputs[scalarSlot] = dst[offset++];
+								   });
+				}
+				break;
+			}
+			default:
+				break;
+			}
+		}
+	}
 }

src/Pipeline/SpirvShader.hpp

@@ -313,9 +313,9 @@ namespace sw
 		std::vector<InterfaceComponent> inputs;
 		std::vector<InterfaceComponent> outputs;
 
-		void emitEarly(SpirvRoutine *routine) const;
-
+		void emitProlog(SpirvRoutine *routine) const;
 		void emit(SpirvRoutine *routine) const;
+		void emitEpilog(SpirvRoutine *routine) const;
 
 		using BuiltInHash = std::hash<std::underlying_type<spv::BuiltIn>::type>;
 		std::unordered_map<spv::BuiltIn, BuiltinMapping, BuiltInHash> inputBuiltins;

src/Pipeline/VertexProgram.cpp

@@ -66,6 +66,8 @@ namespace sw
 		{
 			Nucleus::setInsertBlock(returnBlock);
 		}
+
+		spirvShader->emitEpilog(&routine);
 	}
 
 	RValue<Pointer<Byte>> VertexProgram::uniformAddress(int bufferIndex, unsigned int index)

src/Pipeline/VertexRoutine.cpp

@@ -28,7 +28,7 @@ namespace sw
 		: state(state),
 		  spirvShader(spirvShader)
 	{
-	  	spirvShader->emitEarly(&routine);
+	  	spirvShader->emitProlog(&routine);
 	}
 
 	VertexRoutine::~VertexRoutine()