[dxbc] Implement passthrough geometry shader

This is needed when vertex or domain shader code is
passed to CreateGeometryShaderWithStreamOutput.

- Fix compilation with new DxbcProgramInfo.

src/dxbc/dxbc_compiler.cpp

@@ -179,6 +179,36 @@ namespace dxvk {
   }
 
 
+  void DxbcCompiler::processXfbPassthrough() {
+    m_module.setExecutionMode (m_entryPointId, spv::ExecutionModeInputPoints);
+    m_module.setExecutionMode (m_entryPointId, spv::ExecutionModeOutputPoints);
+    m_module.setOutputVertices(m_entryPointId, 1);
+    m_module.setInvocations   (m_entryPointId, 1);
+
+    for (auto e = m_isgn->begin(); e != m_isgn->end(); e++) {
+      emitDclInput(e->registerId, 1,
+        e->componentMask, e->systemValue,
+        DxbcInterpolationMode::Undefined);
+    }
+
+    // Figure out which streams to enable
+    uint32_t streamMask = 0;
+
+    for (size_t i = 0; i < m_xfbVars.size(); i++)
+      streamMask |= 1u << m_xfbVars[i].streamId;
+    
+    for (uint32_t mask = streamMask; mask != 0; mask &= mask - 1) {
+      const uint32_t streamId = bit::tzcnt(mask);
+
+      emitXfbOutputSetup(streamId, true);
+      m_module.opEmitVertex(m_module.constu32(streamId));
+    }
+
+    // End the main function
+    emitFunctionEnd();
+  }
+  
+  
   Rc<DxvkShader> DxbcCompiler::finalize() {
     // Depending on the shader type, this will prepare
     // input registers, call various shader functions
@@ -2078,7 +2108,7 @@ namespace dxvk {
       emitOutputSetup();
       emitClipCullStore(DxbcSystemValue::ClipDistance, m_clipDistances);
       emitClipCullStore(DxbcSystemValue::CullDistance, m_cullDistances);
-      emitXfbOutputSetup(streamId);
+      emitXfbOutputSetup(streamId, false);
       m_module.opEmitVertex(streamVar);
     }
 
@@ -4254,6 +4284,21 @@ namespace dxvk {
   }
   
   
+  DxbcRegisterPointer DxbcCompiler::emitArrayAccess(
+          DxbcRegisterPointer     pointer,
+          spv::StorageClass       sclass,
+          uint32_t                index) {
+    uint32_t ptrTypeId = m_module.defPointerType(
+      getVectorTypeId(pointer.type), sclass);
+    
+    DxbcRegisterPointer result;
+    result.type = pointer.type;
+    result.id = m_module.opAccessChain(
+      ptrTypeId, pointer.id, 1, &index);
+    return result;
+  }
+
+
   uint32_t DxbcCompiler::emitLoadSampledImage(
     const DxbcShaderResource&     textureResource,
     const DxbcSampler&            samplerResource,
@@ -6292,10 +6337,21 @@ namespace dxvk {
   }
 
 
-  void DxbcCompiler::emitXfbOutputSetup(uint32_t streamId) {
+  void DxbcCompiler::emitXfbOutputSetup(
+          uint32_t                          streamId,
+          bool                              passthrough) {
     for (size_t i = 0; i < m_xfbVars.size(); i++) {
       if (m_xfbVars[i].streamId == streamId) {
-        DxbcRegisterPointer srcPtr = m_oRegs[m_xfbVars[i].outputId];
+        DxbcRegisterPointer srcPtr = passthrough
+          ? m_vRegs[m_xfbVars[i].outputId]
+          : m_oRegs[m_xfbVars[i].outputId];
+
+        if (passthrough) {
+          srcPtr = emitArrayAccess(srcPtr,
+            spv::StorageClassInput,
+            m_module.constu32(0));
+        }
+        
         DxbcRegisterPointer dstPtr;
         dstPtr.type.ctype  = DxbcScalarType::Float32;
         dstPtr.type.ccount = m_xfbVars[i].dstMask.popCount();

src/dxbc/dxbc_compiler.h

@@ -382,6 +382,15 @@ namespace dxvk {
     void processInstruction(
       const DxbcShaderInstruction&  ins);
     
+    /**
+     * \brief Emits transform feedback passthrough
+     * 
+     * Writes all captured input variables to the
+     * corresponding xfb outputs, and sets up the
+     * geometry shader for point-to-point mode.
+     */
+    void processXfbPassthrough();
+    
     /**
      * \brief Finalizes the shader
      * \returns The final shader object
@@ -851,6 +860,13 @@ namespace dxvk {
             DxbcRegisterValue       value,
             DxbcOpModifiers         modifiers);
     
+    ////////////////////////////////
+    // Pointer manipulation methods
+    DxbcRegisterPointer emitArrayAccess(
+            DxbcRegisterPointer     pointer,
+            spv::StorageClass       sclass,
+            uint32_t                index);
+
     ///////////////////////////////////////
     // Image register manipulation methods
     uint32_t emitLoadSampledImage(
@@ -1065,7 +1081,8 @@ namespace dxvk {
     void emitXfbOutputDeclarations();
 
     void emitXfbOutputSetup(
-            uint32_t                          streamId);
+            uint32_t                          streamId,
+            bool                              passthrough);
     
     ///////////////////////////////
     // Hull shader phase methods

src/dxbc/dxbc_module.cpp

@@ -64,6 +64,25 @@ namespace dxvk {
   }
   
   
+  Rc<DxvkShader> DxbcModule::compilePassthroughShader(
+    const DxbcModuleInfo& moduleInfo,
+    const std::string&    fileName) const {
+    if (m_shexChunk == nullptr)
+      throw DxvkError("DxbcModule::compile: No SHDR/SHEX chunk");
+    
+    DxbcAnalysisInfo analysisInfo;
+
+    DxbcCompiler compiler(
+      fileName, moduleInfo,
+      DxbcProgramType::GeometryShader,
+      m_osgnChunk, m_osgnChunk,
+      analysisInfo);
+    
+    compiler.processXfbPassthrough();
+    return compiler.finalize();
+  }
+
+
   void DxbcModule::runAnalyzer(
           DxbcAnalyzer&       analyzer,
           DxbcCodeSlice       slice) const {

src/dxbc/dxbc_module.h

@@ -60,6 +60,20 @@ namespace dxvk {
       const DxbcModuleInfo& moduleInfo,
       const std::string&    fileName) const;
     
+    /**
+     * \brief Compiles a pass-through geometry shader
+     *
+     * Applications can pass a vertex shader to create
+     * a geometry shader with stream output. In this
+     * case, we have to create a passthrough geometry
+     * shader, which operates in point to point mode.
+     * \param [in] moduleInfo DXBC module info
+     * \param [in] fileName SPIR-V shader name
+     */
+    Rc<DxvkShader> compilePassthroughShader(
+      const DxbcModuleInfo& moduleInfo,
+      const std::string&    fileName) const;
+    
   private:
     
     DxbcHeader   m_header;