[dxbc] Implemented DclHsMaxTessFactor

src/dxbc/dxbc_compiler.cpp

@@ -247,6 +247,9 @@ namespace dxvk {
       case DxbcOpcode::DclOutputControlPointCount:
         return this->emitDclOutputControlPointCount(ins);
       
+      case DxbcOpcode::DclHsMaxTessFactor:
+        return this->emitDclHsMaxTessFactor(ins);
+        
       case DxbcOpcode::DclTessDomain:
         return this->emitDclTessDomain(ins);
       
@@ -1109,6 +1112,11 @@ namespace dxvk {
   }
   
   
+  void DxbcCompiler::emitDclHsMaxTessFactor(const DxbcShaderInstruction& ins) {
+    m_hs.maxTessFactor = ins.imm[0].f32;
+  }
+  
+  
   void DxbcCompiler::emitDclTessDomain(const DxbcShaderInstruction& ins) {
     const spv::ExecutionMode executionMode = [&] {
       switch (ins.controls.tessDomain) {
@@ -5033,6 +5041,12 @@ namespace dxvk {
       const uint32_t tessFactorArrayIndex
         = m_module.constu32(tessFactor.index);
       
+      // Apply global tess factor limit
+      DxbcRegisterValue tessValue = emitRegisterExtract(value, mask);
+      tessValue.id = m_module.opFClamp(getVectorTypeId(tessValue.type),
+        tessValue.id, m_module.constf32(0.0f),
+        m_module.constf32(m_hs.maxTessFactor));
+      
       DxbcRegisterPointer ptr;
       ptr.type.ctype  = DxbcScalarType::Float32;
       ptr.type.ccount = 1;
@@ -5043,7 +5057,7 @@ namespace dxvk {
         tessFactor.array, 1,
         &tessFactorArrayIndex);
       
-      emitValueStore(ptr, emitRegisterExtract(value, mask),
+      emitValueStore(ptr, tessValue,
         DxbcRegMask(true, false, false, false));
     } else {
       Logger::warn(str::format(

src/dxbc/dxbc_compiler.h

@@ -200,6 +200,8 @@ namespace dxvk {
     DxbcCompilerHsPhase currPhaseType   = DxbcCompilerHsPhase::None;
     size_t              currPhaseId     = 0;
     
+    float maxTessFactor = 64.0f;
+    
     uint32_t vertexCountIn   = 0;
     uint32_t vertexCountOut  = 0;
     
@@ -484,6 +486,9 @@ namespace dxvk {
     void emitDclOutputControlPointCount(
       const DxbcShaderInstruction&  ins);
     
+    void emitDclHsMaxTessFactor(
+      const DxbcShaderInstruction&  ins);
+    
     void emitDclTessDomain(
       const DxbcShaderInstruction&  ins);
     

src/dxbc/dxbc_decoder.h

@@ -273,9 +273,12 @@ namespace dxvk {
    * \brief Immediate value
    * 
    * Immediate argument represented either
-   * as a 32-bit or 64-bit unsigned integer.
+   * as a 32-bit or 64-bit unsigned integer,
+   * or a 32-bit or 32-bit floating point number.
    */
   union DxbcImmediate {
+    float    f32;
+    double   f64;
     uint32_t u32;
     uint64_t u64;
   };

src/dxbc/dxbc_defs.cpp

@@ -726,14 +726,16 @@ namespace dxvk {
     /* DclTessOutputPrimitive               */
     { 0, DxbcInstClass::Declaration },
     /* DclHsMaxTessFactor                   */
-    { },
+    { 1, DxbcInstClass::Declaration, {
+      { DxbcOperandKind::Imm32, DxbcScalarType::Float32 },
+    } },
     /* DclHsForkPhaseInstanceCount          */
     { 1, DxbcInstClass::HullShaderInstCnt, {
-      { DxbcOperandKind::Imm32, DxbcScalarType::Uint32  },
+      { DxbcOperandKind::Imm32, DxbcScalarType::Uint32 },
     } },
     /* DclHsJoinPhaseInstanceCount          */
     { 1, DxbcInstClass::HullShaderInstCnt, {
-      { DxbcOperandKind::Imm32, DxbcScalarType::Uint32  },
+      { DxbcOperandKind::Imm32, DxbcScalarType::Uint32 },
     } },
     /* DclThreadGroup                       */
     { 3, DxbcInstClass::Declaration, {

src/dxbc/dxbc_util.cpp

@@ -23,8 +23,9 @@ namespace dxvk {
       }
     } else {
       // Global resource slots
-      //   0 -   7: Uniform access views
-      //   8 -  11: Stream output buffers
+      //   0 -   7: Unordered access views
+      //   8 -  15: UAV counter buffers
+      //  16 -  19: Stream output buffers
       // Per-stage resource slots:
       //   0 -  13: Constant buffers
       //  14 -  29: Samplers